Anthropometric Thermal Evaluation and Recommendation Method of Physiotherapy for Athletes

FACULDADE DE ENGENHARIA DA UNIVERSIDADE DO PORTO Anthropometric Thermal Evaluation and Recommendation Method of Physiotherapy for Athletes Ana Sofia Roque Domingues MASTER DISSERTATION Mestrado Integrado em Bioengenharia Supervisor: Ricardo Vardasca, Ph.D., AMBCS Co-Supervisor: Eduardo Marques, MSc. June 2014 Abstract Nowadays, injuries are identified as persistent and serious problems in the sports activity. Ankle sprains compose a significant part of these injuries and present high frequency and associated costs. Among other issues, this condition is related to the common persistence of sequel and re-occurrences. The disregard for its severity and a reduced follow-up of the treatment may be pointed as associated causes. The aim of this dissertation consisted in the assessment and monitoring of physiotherapy procedures for cases of ankle sprain. Infrared thermography, a non-contact, non-invasive, non-ionizing, precise and safe medical imaging modality, has been used for analyzing the temperature distribution of the affected area since this type of pathology has a marked effect on it. In order to establish an association between the collected data and the respective subject, the development of a methodology for extracting anthropometric measures with the Microsoft Kinect was also an objective of this work . Although not being a primary goal, it was also intended to use these two techniques for establishing a set of anatomical control points in whole body thermograms. This contribution was seen has a pilot study for the automation of the thermal images analysis that is still manually performed. The thermal evaluation was conducted on rugby players from CDUP (Centro de Desporto da Universidade do Porto) team, who were followed for a 30 days period. This monitoring process was focused on athletes that were recovering from ankle sprains and receiving physiotherapeutic treatment. A parallel collection was performed using the Microsoft Kinect sensor to capture RGB-D images from subjects external to the team. An algorithm was developed in order to extract the referred measures from these images, being created a complete methodology of image acquisition and processing. The descriptive evaluation of the collected thermograms and the statistical analysis of the absolute thermal symmetry values obtained has allowed to confirm the presence of a pathology in the injured subjects. It has been also possible to associate the recovery stage of each injury with the temperature distribution observed and, in every case, confirm a positive evolution of the ankle sprain. Regarding the anthropometric evaluation, 12 measures have been automatically extracted from RGB-D images of each subject, whose values were compared with the obtained from the Qualisys system. The relative errors obtained allowed to identify the most critical steps of this estimation. Using these measures and others additionally determined, a gender discriminant algorithm was developed and tested, presenting a percentage of correct classifications of 83%. The results obtained by applying the developed algorithm in whole body thermograms were very promising, having been marked relevant anatomical points. In the future, with the proper optimization, the integration of both techniques may answer the need of having a fully automated method for thermal images analysis. The main goals of this work have been achieved, proving that the infrared thermography is a useful tool for assessing and monitoring ankle sprains and being possible to establish a suitable methodology for the acquisition and analysis of thermograms. The estimation of anthropometric landmarks and measures using a low-cost depth sensor confirmed to be an important auxiliary procedure both for subjects profiling and thermograms analysis. i ii Acknowledgments This work would not have been possible without the contribution and support of several people. First and foremost I would like to thank my supervisors: Dr. Ricardo Vardasca for the motivation, continuous advising and constructive discussion, and Eduardo Marques for the everyday guidance, helpful ideas, inspiring advices and endless patience throughout the entire project. I also acknowledge the VCMI (Visual Computing and Machine Intelligence) group and the INESC TEC for the great work environment provided and for all the support. Moreover, I would like to acknowledge Márcio Borgonovo and Professor Joaquim Gabriel for all the advices and assistance, and Professor Luísa Estriga for the help in establishing contacts and for all the ideas and pertinent questions. I would also like to thank everyone from CDUP for always being present to help, specially, Manuel Ruivo for the assistance and medical insights and Miguel Moreira for the availability and interest. I also take this opportunity to thank all those people who have never hesitated to help and take an active, and sometimes repetitive, role in this project, the participants: a great part of this thesis is literally you and I could not be more grateful. During this path, I have always counted with the unconditional support of those who are closest to me and to whom I should thank every day. To my parents, who have given me the affection, help and trust that made me go throughout one more phase of my life and who are the greatest examples of persistence and good-willing. To my brother, my all-time role-model, and to my sister-in-law, my little nephew and my grandparents that are always encouraging me. A very special thanks to João Amaral for the everyday support, for all his patience, for giving me the right advices in the right time and for never stop believing in me. I could not pass the opportunity of thanking my “fellow travelers”: Filipa, Inês and Rita, for all the help they gave me and for making every day far better. In a moment of closure, I would also like to thank all my friends that pass these unforgettable 5 years with me and to all those that, by one reason or another, I could not always be with: you are always showing me that friendship is what makes every experience matter. Sofia Domingues iii iv ” It always seems impossible until it’s done. ” Nelson Mandela v vi Contents 1 Introduction 1 1.1 Motivation........................................... 1 1.2 Objectives........................................... 3 1.3 Contributions ......................................... 3 1.4 Structure of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2 Literature Review 5 2.1 AnkleJoint .......................................... 5 2.1.1 AnatomyandPhysiology............................... 5 2.1.2 Injuries ........................................ 9 2.1.3 ThermalPhysiology ................................. 10 2.2 MedicalThermalImaging .................................. 12 2.2.1 HistoricalOverview ................................. 13 2.2.2 PrincipleofIRT.................................... 13 2.2.3 MedicalThermography................................ 14 2.2.4 Thermography in Sports Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3 Anthropometry in Biometric Identification . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.3.1 Biometrics ...................................... 17 2.3.2 Anthropometry.................................... 18 2.4 Conclusions.......................................... 23 3 Methodology 25 3.1 EthicalIssues ......................................... 25 3.2 Thermography......................................... 25 3.2.1 Sample Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 3.2.2 ImageAcquisition .................................. 26 3.2.3 ImageAnalysis.................................... 27 3.3 Anthropometry ........................................ 29 3.3.1 DataCollection.................................... 29 3.3.2 Anthropometric Measures Estimation . . . . . . . . . . . . . . . . . . . . . . . . 31 3.3.3 GenderEstimation .................................. 38 3.3.4 Evaluation of the Anthropometric Assessment . . . . . . . . . . . . . . . . . . . . 41 3.4 Identification of Anatomical Control Points in Thermograms . . . . . . . . . . . . . . . . 41 3.4.1 Regions of Interest in Thermograms . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.5 Conclusions.......................................... 43 4 Results 45 4.1 Thermographic Monitoring of Ankle Sprains . . . . . . . . . . . . . . . . . . . . . . . . 45 4.1.1 Sample Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.1.2 StatisticalAnalysis.................................. 46 4.1.3 Follow-up....................................... 47 4.2 Anthropometric Assessment with Microsoft Kinect . . . . . . . . . . . . . . . . . . . . . 53 4.2.1 SilhouetteExtraction................................. 53 vii xiv ABBREVIATIONS Chapter 1 Introduction The ankle is the most commonly injured joint in the human body [1], accounting for 25% of all the locomotor system trauma [2]. It is estimated that, on average, each person experiences 2 to 3 ankle injuries during a life time [3]. Ankle sprains represent 85% of all the injuries suffered in this region and present the higher frequency. They occur when a sudden twisting movement of the foot makes the supporting ligaments stretch or tear [3]. Studies [4, 5] have shown that this condition has an incidence of 1 per 10000 patients per day and an annual peak of 7.2 per 1000 patients. The most affected age group comprises ages ranging from 21 to 30 years old. As most common consequences of this injury, it is possible to identify: pain, loss of mobility and functional instability [6]. In sports, particularly in those involving sharp cutting movements, such as basketball, soccer, rugby and volleyball, ankle sprains are a persistent problem [3, 7]. For being a contact sport with hard collisions, rugby is associated with high injury risks, mostly musculoskeletal, concussions or related to joints [8]. The major part of the injuries correspond to ligaments tear and occur in the tackle phase, showing the effect of collision in causing these injuries. In professional sports teams, when a key player suffers a severe injury that forces him to leave competition, enormous financial losses are usually inevitable. For that reason considerable investments are being made in the prevention and efficient treatment, in order to avoid first and recurrent ankle sprains [6]. In this section, the presented problem will be addressed with more detail, giving relevance to the treatment and rehabilitation procedures. The financial impact associated, some issues related to the conducted treatment and a potential improvement of this process will also be approached. Finally, the objectives of this work will be established and the dissertation structure outlined. 1.1 Motivation Nowadays, ankle sprains are considered an actual public health problem, mainly because of its high frequency and associated costs [2]. Although no estimation is found for several countries, considering the values known from Netherlands, an average of e360 is spend in every sprained ankle case [9]. The financial impact of this injury in sports is not only related to the treatment costs but also with the absence of 1 2Introduction important athletes that have a key role in the success of the team and, therefore, in its economical revenue [10]. There is often a disregard for the importance of this injury, by patients and physicians. Since diagnosis tests are not always conclusive and the pain experienced by the patient is subjective, ankle sprains are sometimes seen as trivial injuries and their treatment is underrated. For this reason, there are sequels and reoccurrences cases both in athletes and in the general population. In fact, in a recent work it has been reported that, 3 years after the injury, only 36 to 85% of the patients achieve a full recovery [2]. Several studies [11, 12, 13] indicate that, in ankle sprain cases, conventional treatment is more appropriate and presents better results than surgery. It consists in the application of the RICE (Rest Ice Compression Elevation) protocol followed by early immobilization, using preferably a semi-rigid brace, and rehabilitation [2]. The monitoring of the rehabilitation process is crucial for a consistent and effective treatment. Only with a careful supervision by the physiotherapist and with patient cooperation, a complete recovery is accomplished and the risk of re-sprain is minimized [2].The decision of returning to play is based on subjective information that sometimes leads the athletes to take serious risks. Since there is no efficient way to attest the fully recovery of an ankle sprains, physiotherapists and doctors rely on their experience and in the athletes feedback. However, different athletes have different healing capacities, response to medications and pain tolerance biasing this process. Therefore, a rehabilitation monitoring tool would help in these decisions, by giving quantitative information about the injury regression [14]. Infrared Thermography (IRT) may assume an important role in the monitoring of physiotherapeutical procedures. This technique has already a long history in medical applications. It was first used to evaluate pain in 1964 [15] and since then it has proven to be a reliable and appropriate tool to address many conditions. The human body emits an amount of infrared (IR) radiation that is detectable by thermal cameras and processed in order to produce corresponding images, thermograms, where each pixel corresponds to a temperature value [16, 17]. In cases of pathologies associated with infection or inflammation, some increases in skin temperature may be registered in the affected areas. Ankle sprains are typically hyperthermic and appear as non-homogeneous regions in the thermograms [18]. At the present, IRT already is used as an injury assessment tool for athletic animals and a lot of research is being conducted in order to apply this technique in humans [19]. IRT is a non-invasive, non-contact, nonionizing, objective and, therefore, harmless and safe medical imaging method, which has a high potential to be used in the previously mentioned problem. In order to perform a complete monitoring of this type of therapeutic process, it becomes relevant to maintain a complete profile of the subject. Soft biometrics are attributes that characterize people, however they lack in the robustness required to differentiate them. With a selection of soft biometrics, it is possible to create a patient’s profile, considering characteristics, such as: gender, ethnicity, height, weight, BMI (Body Mass Index) and measures of specific parts of the body, anthropometric measures [20]. The Microsoft Kinect is a sensor that combines a standard RGB camera with a depth camera and has been initially developed for game purposes [21]. However, an increasingly interest is being showed in order to use this equipment to perform biometric and anthropometric studies [22]. Therefore, the use of this technology appears has an interesting option to improve the purposed work. Since the needed technologies for further investigate this proposed solution, as well as valuable human resources, are available in the University of Porto, exploring this field is a relevant and pertinent opportunity. 1.2 Objectives 3 1.2 Objectives The aim of this study is to evaluate the possibility of applying infrared thermography in the assessment and monitoring of physiotherapy procedures for cases of ankle sprain in rugby players. Since body conditions are different from subject to subject, it is also intended to establish a methodology for estimating biometric features that can be used to profile the patient, such as: gender, height and other anthropometric measures. In order to accomplish the referred aim, the following objectives were established: •Establish a standardized protocol of thermal imaging capture and analysis for the follow-up of ankle sprains; •Perform a statistical analysis to evaluate the thermal patterns analyzed and characterize this injury; •Define and implement a methodology for anthropometric measurement of a subject, by using RGB-D images from Microsoft Kinect; •Evaluate the accuracy and robustness of Microsoft when used as a tool in the extraction of anthropometric measures by comparing the results with the obtained using a proper validation system; •Complement the subject’s characterization with the gender, determining it automatically using the anthropometric data. Some difficulties and disparities are still experienced in the analysis of thermograms for medical applications since there is no automatic analysis method available. Therefore, it is also intended to work towards this goal and develop an algorithm to detect body contours and identify specific anatomical points. 1.3 Contributions Three main contributions have result from the proposed work: 1. It has identified a new and suitable application of thermography in the monitoring of physiotherapy in sprained ankles which can be applied to other musculoskeletal disorders. 2. An algorithm for determining anthropometric measures and the gender of a subject by using RGB-D images obtained from the Microsoft Kinect was presented and its performance evaluated, consisting in a different methodology from those already existent in literature and having results comparable and similar to the state-of-art. 3. It has suggested the use of depth sensors, such as the Microsoft Kinect, to extract control points in thermograms, making a first contribution towards this goal. 1.4 Structure of the Dissertation This dissertation is composed by 6 chapters (Figure 1.1). After this introductory section, Chapter 2 addresses the anatomical and physiological characterization of the ankle joint and its most common injuries, namely ankle sprains. It also presents the IRT technique, with more focus in its application in the medicine. Anthropometry will also be object of detail in the second chapter, with respect to its application in biometrics. In Chapter 3, the experimental methodology applied for recording and analyzing thermal images will 4Introduction be presented. The acquisition process for Microsoft Kinect will also be detailed, as well as the algorithm developed for the detection of anthropometric measures and its validation. Chapter 4 presents the main results obtained, which are discussed and related to the literature in Chapter 5. Finally, Chapter 6 gives a global perspective of the developed work, presenting relevant conclusions and final remarks. It also provides guidelines and questions to be explored in future work. Figure 1.1: Schematic representation of dissertation structure. Chapter 2 Literature Review In this chapter, the anatomical and physiological concepts related to this work will be presented, including: the injuries that most affect the ankle joint, the anatomy of this region and the physiological basis of thermal imaging. The following sections present the techniques in study and the state-of-art of the application of thermography in medicine and in sports injuries and also of the use of anthropometric measurements as biometric features. 2.1 Ankle Joint The ankle joint forms the articulation between the leg and the foot. Due to its anatomical structure, it is more stable when the foot is dorsiflexed. When assuming other positions, the ankle becomes more fragile and vulnerable to different injuries [1]. In this section, the anatomy of the ankle will be addressed, as well as the injuries that may occur in this region, in particularly those affecting its ligaments – the sprains. The physiological phenomena in which thermography is based is also presented. 2.1.1 Anatomy and Physiology The lower limb has several important functions, such as: support of the body weight, gravity control and locomotion [23]. It presents three segments: gluteal region, legs and feet. The joint between the last two is named ankle or talocrural joint and involves three bones: the talus of the foot, the tibia and fibula of the leg. The ankle joint is synovial and passes the body weight from the tibia to the talus, allowing foot movements of dorsiflexion, plantar flexion and limited inversion and eversion (Figure 2.1). The articular surfaces of this joint are covered in hyaline cartilage [1, 23]. Since the anterior and posterior margins of the tibia and fibula are practically nonexistent and the medial and lateral margins are very extensive, together they create a “deep bracket-shaped socket” [23] for the talus (Figure 2.2.B). The articular surface of talus has the shape of short half-cylinder (Figure 2.2.C) with one end turned to the lateral side and the other to the medial. Its upper side has two ends covered by hyaline cartilage and fits into the bracket-shaped socket structure. Two membranes enclose the articular cavity: a synovial and a fibrous. The synovial is attached to the articular surfaces margins and is covered by the fibrous one, which is also attached to the adjacent bones. 5 6Literature Review Figure 2.1: Movements of the foot ensured by the ankle joint: plantar flexion, dorsiflexion and limited eversion and inversion (From [1]). . Figure 2.2: Bone structure of the ankle joint: A. Foot in plantar flexion (anterior view). B. Schematic representation of the joint. C. Articular face of the talus (superior view) (From [23]). The ankle joint is most stable when it is in dorsiflexion. In this position, the talus’ wider surface (anterior) moves into the joint and fits tighter into the “socket” formed by tibia and fibula. On the contrary, when plantar flexed, it is the narrower part (posterior) that fits the joint, resulting in a less stable position [23]. 2.1.1.1 Ligaments Medial and lateral ligaments stabilize the ankle joint. The medial ligament, also named deltoid, is large and triangular shaped. In the apex, it is attached to the medial malleolus and, in the base, to a line that connects the tuberosity of the navicular bone to the tubercle of the talus. Considering the points of attachment, four parts of the medial ligament can be identified (Figure 2.3): •The tibionavicular: attached in front to the tuberosity of the navicular bone; •The tibiocalcaneal: inserted in the sustentaculum tali of the calcaneus bone; •The posterior tibiotalar: inserted in the medial side and with the medial tubercle of the talus; 2.1 Ankle Joint 7 •The anterior tibiotalar: attached to the medial surface of the talus. Figure 2.3: Medial ligament of the ankle joint (From [23]). The lateral ligament is also composed by three different ligaments (Figure 2.4): •The anterior talofibular: connects the anterior margin of the lateral malleolus to the adjacent region of the talus; •The posterior talofibular ligament: passes from the malleolar fossa, through the medial side of the lateral malleolus, to the posterior process of the talus; •The calcaneofibular ligament: attached to the malleolar fossa on the posteromedial side of the lateral malleolus and to a tubercle on the lateral surface of the calcaneus . Figure 2.4: Lateral ligament of the ankle joint: A. Lateral View. B. Posterior View (From [23]). 8Literature Review 2.1.1.2 Arterial and Venous Circulation The arterial blood supply of the foot (Figure 2.5.A) is ensured by branches of the posterior tibial and dorsalis pedis arteries. The posterior tibial artery passes through the medial side of the ankle and enters the foot through the tarsal tunnel. Between the medial malleolus and the heel, the posterior tibial artery originates a small medial plantar artery and a larger lateral plantar artery. The dorsalis pedis begins when the anterior tibial passes the ankle joint and ends up joining the deep plantar arch in the sole of the foot. Tarsal branches of the dorsalis pedis pass over the tarsal bones and supply adjacent structures, anastomosing with a network of vessels formed around the ankle [23]. Interconnected networks of superficial (Figure 2.5.B)and deep veins rise from the foot. The deep veins follow the same path as arteries and superficial veins originated from different sides of the dorsal venous arch pass onto the medial side or onto the back side of the leg (the great saphenous vein and the small saphenous vein, respectively) [23]. Figure 2.5: Blood circulation of the foot: A. Arterial circulation. B. Superficial veins. (From [23]). 2.1.1.3 Muscles Leg muscles inserted in this joint and in foot bones ensure ankle movement. The tibialis anterior muscle has origin in the lateral surface of the tibia and is inserted into the medial cuneiform of the ankle and in the first metatarsal bones. It is responsible for the dorsiflexion and inversion movements of the foot. During locomotion, this muscle stabilizes the ankle, supporting the medial arch of the foot. A superficial muscle attached to the head of the fibula, the fibularis longus, allows the eversion and plantar flexion of the foot. It passes through the lateral side of the leg, becomes a tendon posteriorly to the lateral malleolus of the ankle and finally attaches to the medial cuneiform and first metatarsal. The tibialis posterior also participates in support and plantar flexion and inversion movements and the fibularis tertius in dorsiflexion and eversion movements of the foot [23]. 2.1 Ankle Joint 9 2.1.2 Injuries In order to understand the wide variety of injuries that may occur in the ankle joint, it is very important to have a specific anatomical perspective of this region. Usually the ankle joint is observed as a fibro-osseous ring with a coronal orientation [23]: •The distal ends of tibia and fibula form the upper part; •The ligaments connecting the medial and lateral malleolus to the tarsal bones form the sides of the ring; •The subtalar joint (not part of the ankle joint) delineates the bottom. Considering this structure, it is easier for a physician to analyze the damage resulting from each type of injury. For example, a disruption of the ring may occur not only by a damage to the bones – fracture –, but also by a damage to the ligaments – sprain – that is likely to be missed by plain radiographs [23]. 2.1.2.1 Ankle Sprains When ligaments of the ankle are torn partially or completely, it is considered that a sprain has occurred (Figure 2.6). A strong and forced inversion of the foot appears as the most common cause of this injury, resulting in lateral ligaments damage. More remotely, medial ligaments may also be affected by an eversion of the foot, being frequently associated with a malleoli fracture [1]. A sprained ankle usually occurs when the foot is plantar flexed, since it is less stable in this position and, consequently, needing more ligament support. The most common sprain involves the anterior talofibular ligament that is also affected by inversion sprains. The calcaneofibular ligament is sometimes also involved in this type of injury (Figure 2.6) [24]. Figure 2.6: Ankle sprain - Illustrative representation of the trauma and damaged ligaments: calcaneofibular and anterior talofibular (From [1]). 16 Literature Review [17, 18, 42] have been conducted in matter, covering knee pathologies: Arfaoui et al. [17], for example, concluded that IRT was a reliable tool to detect specific temperature patterns in athletes with knee osteoarthritis. The studies already conducted sustain the possibility of using IRT for addressing other injury cases. Due to their high occurrence and epidemiology (see Chapter 1), ankle sprains are largely studied in different areas. However, there are not many works regarding the application of IRT to this injury. Nevertheless, some important observations have been made: Asagai et al. [43] reported some cases of high temperature around the trauma site and reduced temperature in the periphery, on the first day after the injury; this cold skin pattern was interpreted as indicative of poor prognosis and long recovery time [19]; Schmitt and Guillot [44], reported, several years ago, that the thermal distribution of temperature in ankle sprains was related to the recovery time recommended: a bilateral isothermia reflected a minor injury treated in 1-2 weeks; a thermal asymmetry between sprained and healthy ankle of 1.5oC to 2.0oC required usually a recovery of 4 weeks. Healthy people present a symmetric thermal distribution. Thermal symmetry of the human body has been defined [45] as the ‘degree of similarity’ of two ROIs mirrored across the main longitudinal axis of the human body and which are identical in shape and size, being taken at the same angle. When an injury induces a localized increase in temperature, it will disturb this normal symmetric pattern [18]. In fact, it has been established that a different superior to 0.5±0.3 oC is abnormal and may point to a pathological condition [46]. As any other physiological manifestation, a temperature change may be a result of different causes and cannot be assuredly related to a specific condition without further analysis. Therefore, observational data of the injury collected for a significant period of time and a deeper understanding of the biological phenomena have to be taken into account by the physicians and physical therapists. In order to use IRT to address sports injuries, it is crucial that a comparative analysis is performed, considering injured and correspondent non-injured areas. It is also ideal that a continuous follow-up is conducted, for a better understanding of the progression of these modifications. By considering all these factors, it would become possible to create a database of normal and pathological data. [18]. Thermography is now seen as a helpful technology that may help athletes, coaches, physicians and physical therapists in cases of injuries for prevention, early detection and therapy assessment. It is the least expensive non-invasive method that presents absolutely no harm to the patients. Among other issues and challenges that are currently presented to IRT, the construction of a biometric system applied in thermal images to complement this technique and the automatic discovery of anatomic control points to delineate the geometrical and standardized regions of interest that are currently used, are pointed as relevant breakthroughs [28]. These two challenges will be subject of analysis in the present work. 2.3 Anthropometry in Biometric Identification Face, voice and gait are parameters that have always been used to recognize people. Regardless the process, a person’s identification is performed by analyzing physiological and behavioral characteristics, from which some are classified as biometrics for being markedly distinguishing [47, 48]. Soft biometrics refers to the traits that also characterize people, although lacking in the robustness required to differentiate them. Some examples of these features are: eyes color, gait, stature and ethnicity [49]. Anthropometry is the part of anthropology that studies human body measurements. This field comprises the extraction and detailed analysis of different parameters, such as: weight, height, body lengths and 2.3 Anthropometry in Biometric Identification 17 circumferences, skinfold thickness, etc. [50]. Since these dimensions describe the human body, they can serve as a reference to construct a subject’s profile and actually be used as biometric features. In this section, the biometrics and anthropometry fields will be presented in more detail, with special reference to the use of anthropometric measures as soft biometrics. In order to develop an evaluation system with a biometric signature based on anthropometric measures, it is crucial to have a complete understanding of these subjects. The use of Microsoft Kinect for this purpose will also be analyzed, considering some studies developed in the area. 2.3.1 Biometrics The biometrics field provides reliable methodologies for person’s recognition. Biometric identification has already many applications in a broad range of areas including, for example, surveillance and multi-user platforms [48, 51]. A system based in these features is basically a pattern recognition system capable of collecting data from an individual, extract relevant features from it and compare them to a specific dataset. This dataset is composed by the information of different attributes needed to match the individual to one of them [47, 51]. Biometric traits may be grouped on hard and soft biometrics. In order to be classified as hard biometrics, a human characteristic has to satisfy four basic requirements [47]: •Universality: every person must present it; •Collectability: it has to be possible to quantitatively measure it; •Distinctiveness: it must be sufficiently different in any two subjects; •Permanence: the characteristic has to be sufficiently invariant over a period of time. Some examples of hard biometrics are: fingerprints, iris, retina, face appearance, etc., which had already been explored in several studies, proving to be useful in different applications. More recently, researchers have started to give an increasing relevance to soft biometrics [52]. 2.3.1.1 Soft Biometrics During the 19th century, Alphonse Bertillon was the first to present the concept of human identification systems based on morphological characteristics. This was the beginning of biometry as a field of study [20]. However, only several years later, soft biometrics were first introduced by Jain et al. [53], referring to them as traits that provide information about the subjects, without being capable of establishing an absolute identification. This is due to the fact that in soft biometrics some of the referred requirements, such as distinctiveness and permanence, may not be completely satisfied [20]. A selection of the characteristics already classified as soft biometrics is presented in Table 2.1. With the technological and scientific development occurring in this area, more are being identified. These biometric features are qualified in terms of distinctiveness and permanence. In this group, distinctiveness is equivalent to the strength with each trait allows subject identification and is, in most cases, reduced. Height, for example, presents medium distinctiveness since considering a random sample it may be a distinguishing factor, despite being the same for a large number of people. Likewise, soft biometrics may also not be invariant for a long period of time. 18 Literature Review Table 2.1: Soft biometric traits (Adapted from [20]). Trait Permanence Distinctiveness Facial measurements High Medium Ethnicity High Medium Gender High Low Height Medium/High Medium Weight Low/Medium Medium Gait Medium Medium Body measurements Medium/High Medium/High Clothes color Low Medium Currently, these traits are being largely used in different applications: session-based systems, database search for preliminary exclusion and in systems based in hard biometrics in order to improve their reliability [20, 52]. When compared to the hard, soft biometric features may seem to present only disadvantages. However, several works in this area [53, 54] have shown the opposite. Some main advantages can be referred: the increase in speed and accuracy of already existing systems, computational efficiency, deducibility from classical traits and user-friendly behavior in data acquisition. As mentioned before, limitations in permanence and distinctiveness may be pointed. However, by considering a well-established group of soft biometric traits, this drawback may be overpassed [20], being possible to create a complete profile of a subject [49]. The different soft biometrics may be grouped in different categories: facial, body and accessory (eye glasses, for example). Body soft biometrics comprise traits as: gait, weight, height and body measurements; and will be the main focus of this work, namely, the two lastly referred. The height estimation algorithms comprise some basic steps: they have to first recognize the foreground and the background, then, automatically extract real coordinates of the environment and finally, considering this information, they have to estimate the distance from the top of the head to the feet [20]. This topic is already vastly presented in the literature and studies are now more focused on optimizing these algorithms in order to improve their precision [20]. Work in body measurements extraction is developed regarding different purposes. A relevant example is the study conducted by BenAbdelkader et al. [55], which involved body height and shoulder breadth extraction in order to create a multimodal identification system. 2.3.2 Anthropometry Anthropometry studies are applied in different areas, such as: in biomechanics for posture evaluation, in medicine for prevention of musculoskeletal lesions and in nutrition, addressing, for example, the relation between eating habits and changes in body dimensions [56, 57]. There are two types of anthropometric data: static and dynamic. The first refer to body dimensions and include: height, arms and legs length, waist and tight circumference, etc. These measures are usually collected considering fixed anatomical points or landmarks and with the subject assuming standardized postures (Figure 2.9). Dynamic measurements are those collected in order to evaluate the movement amplitude associated with the performance of defined tasks [56]. Over the years, an increasing need for the standardization of static measurement practices has arisen, resulting in the establishment of a set of practices that have to be applied in order to obtain a comprehensive 2.3 Anthropometry in Biometric Identification 19 Figure 2.9: Standardized position for anthropometric measurements: A. Upper leg length measurement. B. Upper arm length measurement (From [58]). anthropometric profile. Specific equipment is required in these studies, including: stadiometers (for stature and sitting height), weighting scales, anthropometric tape (for girths), anthropometers or segmometers (for heights and lengths), anthropometric boxes (for iliospinale height) and several types of calipers. Before any measurement, specific anatomical landmarks need to be marked. They are identifiable body points that establish where the measurement sites are located. The landmarks usually considered, and which are referred in the International Society for the Advancement of Kinanthropometry (ISAK) standards document [57], are presented in Figure 2.10. Figure 2.10: Anatomical landmarks (From [57]). 20 Literature Review 2.3.2.1 Measurement Items Anthropometric measurements are classified in different categories: basic, skinfolds, lengths and breadths. The basic measurement items are: body mass, stature and sitting height. Skinfolds refer to the thickness of some body areas (triceps, biceps, subscapular, abdominal, etc.) where the skin is intentionally fold. Body circumferences, or girths, are also collected from different regions (head, neck, forearm, chest, waist, thigh, etc.) and are measured using a flexible tape. Lengths may be direct, if measured from landmark to landmark, or derived, if based in projected heights taken from the floor (Figure 2.11). Finally, breadths are horizontal lengths of specific regions that are measured with calipers. Even though an unequivocal identification may not be possible, using only these anthropometric features, a complete profile of the subjects is achieved [57]. Figure 2.11: Measures considered for obtaining lengths: A. Direct lengths. B. Projected heights (From [57]). Over the years, anthropometric studies have been carried out using traditional methods, which had precision-related issues and were time-consuming practices. In fact, conducting a reliable and reproducible anthropometric study requires large samples, because of the high variability of human dimensions, and extensive work with each individual. Therefore, the development of digital devices for body scanning introduced the possibility of overcoming these aspects, by automatically acquiring anthropometric measurements [59]. 2.3.2.2 Microsoft Kinect in Anthropometry The detection of the human body in images using computer vision techniques is a wide studied subject comprising complex questions. The major difficulties are associated with the high variability of the scenario in terms of illumination conditions, background configuration and acquisition point of view. The use of range images in this area became a practical solution to these issues. However, the first data acquisition devices of this type of images were expensive and had considerable dimensions, limiting their use [22]. With the appearance of low cost depth cameras, this subject gained more relevance. In fact, these devices 2.3 Anthropometry in Biometric Identification 21 have been successfully employed in different fields, including robotics, forensics, object recognition and scene segmentation [60]. Anthropometry has also been an explored field of application for low cost cameras. As already referred, anthropometric measurements were traditionally performed recurring to time-consuming manual methods with uncertain precision. The first automatic methods for body dimension extraction were based in 3D point clouds [59, 61]. Even though this methodology resulted in high-quality data, it was complex and required a 360 degrees collection, involving expensive equipment and strict environmental conditions [21]. Considering this, low cost range cameras are naturally a useful tool in this area [59]. Several studies [60, 62, 63] identified Microsoft Kinect has a reliable and accurate device for clinical purposes. This product is a low price multisensory device, capable of collecting both depth and RGB images [22] at a frame rate of 30 fps [60]. It was originally created in 2010, for game purposes, as an interface for the Microsoft Xbox 360 [21]. However, over the years its use has spread to technological and scientific applications. The Microsoft Kinect uses near-IR light reflected from the scenario to measure the distance of any point to the sensor, modeling the surfaces curvature. With this information, a real-time depth map with the range values is created. When used to detect the human figures, the KinectSDK (Software Development Kit)1or OpenCV 2modules must be used to process this information and manage to estimate 20 and 15 joints of the body, respectively (Figure 2.12) [21, 22, 64]. Figure 2.12: Skeletal joints extracted from the Microsoft Kinect depth images: A. Joints from KinectSDK (From [21]). B. Joints from OpenCV (From [64]). Considering these joints, it is possible to describe the subject’s posture and extract anthropometric features. Since these measurements cannot be directly obtained from Kinect, there are some studies [21, 61] exploring different methods to estimate them from the extracted information. Samejima et al. [21] developed a full body dimensions estimation method, considering some of the points extracted from Microsoft Kinect. In this study, using the National Institute of Advanced Industrial Science and Technology (AIST) anthropometric database, four body dimensions (height, iliocristal height, biacromial breadth and bitrochanteric breadth) were chosen by a linear multiple regression analysis in order to estimate other 52, with a regression formula based in a principal component analysis. Only height was directly obtained from the KinectSDK. The errors obtained in the measurements estimation were relatively low, all inferior to 16%, however, the sample used in this analysis was reduced (10 subjects). Therefore, this method would have to be further 1http://www.microsoft.com/en-us/kinectforwindows/ 2http://opencv.org/ 22 Literature Review tested to define its accuracy. Also in this context, Azouz et al. [61] presented an algorithm for automatic recognition of anthropometric landmarks on 3D human scans. This method presented a good accuracy for most of the landmarks, according to a validation process involving 30 human models with different shapes. The images were obtained from the CAESAR (Civilian American and European Surface Anthropometry Resource) database [65], which may also be used in future work for data training. Although requiring a larger dataset (200 images for training and 30 for validation) and not addressing anthropometric measurements estimation, when compared to the previously referred, this study was more representative and also presented low error rates: less than 2 cm for most of the landmarks. Studies on biometric applications of Microsoft Kinect have also been conducted. Regarding body dimensions, the study performed by Araujo et al. [51] may be used as reference. It has explored the viability of using Kinect as a tool to create a biometric profile, considering anthropometric measurements. These parameters were extracted from static images of different subjects, without requiring the subject to place himself in standard position in front of the sensor. The results showed the existence of a pose dependency on the information obtained from the Microsoft Kinect, reporting some inconsistency in the identification of skeleton points due to the different poses assumed. A standardized and cooperating position had to be required or different poses had to be considered in the training data. When using soft biometrics, such as body measures, for subject’s characterization, it is important to consider that these traits are not individually distinctive or permanent. Therefore, the use of a specific set of anthropometric measures combined may minimize this weakness. Moreover, the fusion of different types of soft biometrics can be crucial to this end, being object of study in several recent works [66, 67, 20]. Comprising a set of distinct traits, such as gender, age, ethnicity, height and body sizes, more robustness and accuracy are achieved. In this work, gender will be considered to complement the information gathered from the anthropometric analysis, for being a discrete trait capable of providing more accuracy and minimize the complexity of the user’s profile process. 2.3.2.3 Gender Estimation Although being less distinctive, soft biometric features such as gender and ethnicity, are also important descriptors of a person and may be automatically extracted from anthropometric traits. Some studies [68, 69, 67] have already reported methodologies for estimating a subject’s gender from recorded images, presenting promising results. However, facial features are commonly analyzed for this purpose, not being considered other traits that may assume a relevant and complementary role in this type of classification. In several levels of the human body’s anatomy, it is possible to identify sexual dimorphic features [70]. Most anthropometric studies on this field are focused in only one feature [71, 72, 73] instead of in a more distinguishing combination of them. Fullenkamp et al. [70] have studied this matter, purposing three different optimized combinations of anhropometric measurements for gender discrimination, considering different populations. In that study, the required measurements were performed by non-automatized traditional methods, having presented high accuracy values for the three discriminant functions: 99.5% for the North-America (2391 participants) and Italy (800 participants) functions and of 98.5% for the Netherlands’ (1265 participants). 2.4 Conclusions 23 2.4 Conclusions Considering the presented studies, it is possible to verify that there is an absence of a methodology for assessing, with IRT, the evolution of a pathology recovery during physiotherapy treatments. It can also be concluded that there is the need for a new solution for automatically estimate anthropometric measures. In this context, the use of Microsoft Kinect for the determination of body dimensions may provide positive results and enable the use of these parameters and of other related biometric traits, in order to create a subject’s profile. Regarding thermograms analysis, there is also an absence of a method for conducting an automatic segmentation of these images into regions of interest, which would allow a more standardized thermographic evaluation. 24 Literature Review Chapter 3 Methodology In this section, the methodology used to achieve the goals that have been addressed in chapter 1 will be presented. First, the materials and protocols used in the acquisition of the required data will be detailed. The following steps of image processing and analysis will also be addressed, including the methods applied to extract and validate results from both thermograms and RGB-D images. 3.1 Ethical Issues This work was conducted accordingly to the ethical principles for medical research involving human subjects present in the Declaration of Helsinki [74], as well as with the Law no. 21/2014 of April 16th for clinical investigation of the Portuguese Assembly of the Republic [75]. 3.2 Thermography In order to perform a thermographic evaluation of a specific pathological condition, it was important to properly define the views to capture and the regions of interest to analyze. Standardization was also a crucial aspect of thermographic evaluations (see Section 2.2). The acquisition protocol must be carefully prepared and followed in order to collect reliable data [19]. In this section, the sample used in this study will be characterized, describing its participants, and then the procedures used to acquire and analyze thermograms will be detailed. 3.2.1 Sample Characterization This study was conducted in 30 subjects, from whom 13 were male rugby players: 9 with ankle sprains and receiving physiotherapy and 4 healthy that were used as controls. The other 17 subjects (8 males and 9 females) were external to the team and healthy, being also considered as controls. In order to verify the health condition of the analyzed sample, they were asked to complete a form (Appendix A) based on the EURO-QOL questionnaire (Euro-QoL score must be zero for controls). All controls were characterized as healthy, i.e., absent of pathology in the ankles. 25 32 Methodology to the actual 3D shape of the body. However, the presented results are affected by considerable errors that may be minimized. In order to improve the existing methods and results, more attention was paid to this aspect and obtaining an improved body silhouette has been a key step for making a positive determination of the selected landmarks and measures. Therefore, considering the information provided by the Microsoft Kinect, an algorithm was applied in each frame, consisting in (Figure 3.4): Figure 3.4: Schematic representation of the method outline. 1. Depth image pre-processing: The depth image was first segmented using the mask obtained directly from Kinect and this image was then smoothed using a median filter. 2. 2D Silhouette improvement: An initial silhouette was extracted from the RGB image by applying the Canny edge detector to a border of the body created using the mask from Kinect. The resultant open contour was composed by several segments that are refined considering the expected body shape. Finally, these segments were analyzed and those selected as being part of the optimal silhouette were connected using a spline fitting process, closing the contour. An improved silhouette was obtained. 3. 3D Silhouette improvement: The depth image was segmented using the improved silhouette. Then, the existing outliers were identified, corresponding to points that were seen has foreground in this new silhouette but were considered to be background in the mask from Kinect. These pixels’ values were assigned by using a median filter. 3.3 Anthropometry 33 4. Anthropometric landmarks estimation: Considering the data obtained from the previous steps and additional information regarding body proportions, the 2D coordinates of anatomical landmarks were estimated and marked in the RGB image. 5. Anthropometric measures extraction: Using the real-world coordinates of the estimated set of anthropometric landmarks, 3D Euclidean distances were calculated between different pairs of points to determine anthropometric lengths and heights of the subject. 3.3.2.1 Depth Image Pre-Processing The depth image obtained from the Kinect acquisition setup was first processed to smooth the surface of the detected body. In order to achieve this, the depth image was segmented with the mask obtained directly from the Kinect and only this 3D data was used in further analysis. The smoothing process was performed, accordingly to the literature [80], by applying a median filter with a 4×4 window to the resultant image. 3.3.2.2 2D Silhouette Improvement The mask obtained directly from the acquisition software was first used to define a border of the body, by subtracting its dilation from its erosion. The Canny edge detector was then applied to the grayscale image existent only inside this border, which decreases the amount of edges that are detected and correspond to noise. Since the feet region was extremely affected by noise, the mask obtained from the acquisition software was used for these points without further optimization. For the rest of the body, the primary silhouette obtained from the edge detector consists in an open contour composed by individual objects, unconnected lines. Each object is itself composed by different segments that correspond to pixel sequences with a single direction (Figure 3.5). The silhouette was then improved by applying a sequence of morphological operations to remove most of the detected noise. Figure 3.5: Silhouette obtained from the Canny edge detector: open contour composed by objects with different segments. Each object was refined by removing all the end segments with a slope that differs from the object overall tendency in more than 40o(Figure 3.6.A). This parameter was chosen by empirically testing a complete set of angles, differing in 5ofrom each other. 34 Methodology From a visual inspection of the result, it was possible to empirically verify that the Canny edge detector creates noisy segments mainly inside the optimal silhouette, while in the exterior almost no extra edges were detected. Considering this, the algorithm was set to keep only the more external points existing inside the silhouette (Figure 3.6.B). This kind of procedure, assumes that the background is homogeneous, which has to be taken into account when capturing the RGB-D images in order to reduce the noise existing outside the silhouette. Figure 3.6: Outputs from the different refinement processes: A. Refinement based on the slope of the objects’ end segments: Pixels in green correspond to the ones removed. B. Removal of the more internal pixels: the middle axis of each region is presented in red and the pixels removed marked in yellow. Finally, the contour was closed by beginning at the longest object, which was assumed to belong to the improved silhouette. A 35×35 window centered in the end point of this object (query point) was defined and all the start points inside this window are considered as candidates to be linked to the query point. Candidate points have to: •Belong to the same region of the body as the query point or to one that is considered to be adjacent. •Belong to the same side of a region as the query point (lateral or medial), except on the feet, hands and on the top of the head. Regions of the body are defined considering the skeletal joints provided by Kinect. The distance from each point of the silhouette to the 15 existent joints was firstly determined. Each point was associated with the closest joint, creating the referred regions (Figure 3.7). Although belonging to the shoulders region, the points inferior to this joint, correspondent to the armpits area, have a completely different tendency and shape. Therefore, they are considered separately, in some steps of the algorithm. In order to choose the most appropriate object to continue the silhouette, for each candidate four measures are determined: 1. The Euclidean distance between the query point and the candidates; 2. The difference between the slope of the object border and the slope of a linear linking between the query point and the candidate. The object border slope is determined by the weighted mean of the three last segments slopes. This weight is higher for the last segment and decreases with the distance to the discontinuity. If the object has 3 or less segments, the whole object is considered. 3.3 Anthropometry 35 Figure 3.7: Regions of the body defined in the silhouette by considering the skeletal joins obtained from Microsoft Kinect. 3. The length of the candidate object; 4. The distance from the candidate to the closest joint. For each object, these measures were normalized from 0 to 1. Considering the body region in analysis, a different weight was attributed to each measure (see Table 3.4 ) and the global sum was calculated. The candidate selection was performed by choosing the one with the minimum sum result. The weights assigned to each measure have been selected taking into account the geometrical shape of each region. Parts of the body with a curved contour (head, neck and shoulders superior to the joint) have been assigned with lower weights for tendency (less relevant in these cases) and higher values for the euclidean distance. Distance to the joint has been considered for the hands, head and neck, more circular parts where the joint was placed in its center. The length of the object was assigned with a value different from zero for those regions where smaller objects mostly corresponded to noise: head, neck, elbows, torso, hips and knees. Table 3.4: Weights assigned to each measure considered in the candidate selection. Region Euclidean Tendency Object Distance Distance Length to Joint Head, Neck 0.5 0 0.2 0.3 Shoulders (superior to the joint) 1 0 0 0 Shoulders (inferior to the joint) 0.8 0.2 0 0 Elbows, Torso, Hips, Knees 0.8 0.1 0.1 0 Hands 0.7 0 0 0.3 Finally, a spline fitting was applied and used to make the link between the two points, closing the silhouette. In order to have a visual confirmation of the efficiency of the silhouette extraction method, after a smoothing process, this contour was used to segment the grayscale image. 36 Methodology 3.3.2.3 3D Silhouette Improvement The pre-processed depth image was segmented using the silhouette obtained. Pixels that are located inside this silhouette and were considered to be background in the Kinect mask (Figure 3.8) are marked as outliers, since their depth value is likely wrong, and are, finally, reconstructed. Figure 3.8: Visual identification of outliers - pixels considered background in the Kinect mask and as foreground in the improved silhouette: A. Outliers marked in red in the 2D depth image. B. 3D point cloud before the refinement - outliers are easily identified in the point cloud for having a completely different depth value from the remaining pixels. For this reconstruction an 11×11 window, centered at each one of the outliers, was firstly considered and it was iteratively increased until it reaches 25 pixels with an intensity different from 0. A median filter was finally applied considering this window and the resulting value was assigned to the center pixel. By considering this refined image, a point cloud representation of the whole body was finally achieved. 3.3.2.4 Estimating Anthropometric Landmarks and Measures In order to estimate the desired anthropometric heights, breadths and lengths, the silhouette and depth information was first used to determine the 3D coordinates of a set of landmarks (Figure 3.9) : top of the head (vertex) and left and right tip of the middle finger (dactylion), wrist (stylion), elbow (radiale), shoulder (acromicale), iliac crest (iliocristale), femur head (tronchanterion), knee (tibiale laterale) and ankle (sphyrion fibulare). Without further processing, it was possible to extract 3 anthropometric landmarks directly from the information resultant from the first steps: •Vertex: is considered to be the 3D point with the highest Y coordinate from the point cloud. •Dactylion: is assigned to the point that belongs to the hand and that is the lower end of the arm-hand middle axis. The location of the shoulders was determined by considering the skeletal joints of the shoulders and torso obtained from Kinect. The torso joint was linked to each one of the shoulder joints and the coordinates of the shoulder anthropometric landmarks were determined to be in the intersection of these lines with the body silhouette. 3.3 Anthropometry 37 Figure 3.9: Anthropometric landmarks used to extract the corresponding measures (Adapted from [58]). The wrists were determined by looking only at the silhouette points belonging to the hands’ regions. The silhouette was simplified (Figure 3.10) in order to enable the detection of the corner existent laterally in the intersection between the arm and the hand. The angles created by the different simplified lines are analyzed to determine this landmark location (Figure 3.10.A). If the position of the hand does not allow this detection, the same procedure is applied medially. If found, the obtained location was marked laterally and considered to be the wrist (Figure 3.10.B). When this information cannot be extracted from neither one of the sides, the landmark was located in the middle point of segment closer to the arm and still belonging to the hand. Figure 3.10: Identification of the wrist landmark: A. Method applied when it is possible to identified this landmark laterally B. Alternative method consisting in a medial level identification. Having no distinguishable contour characteristic, the elbows could not be identified using a similar approach. Therefore, there has been the need to consider body proportions. 38 Methodology The human body proportions have been object of study for centuries. The usually accepted measurement unit for drawing the human body is the "head size" and, considering this value, relations between different points of the body are inferred [81]. This measure was taken as the vertical distance from the top of the head (determined previously) to the point with the lowest Y coordinate and still belonging to the skeletal joint of the head. Although it was not possible to guarantee that this second point was correctly marked, this estimation relied on the accuracy of the Kinect estimation of skeletal joints, having been successfully determined in all the analyzed cases. The distance from the wrist to the elbow is accepted by some authors [82] to correspond to the size of the person’s head. Therefore, having determined the lateral location of the wrist, it was possible to locate the elbow by simply estimating the size of the subject’s head. The femurs heads and the iliac crests could not either be identified by simply looking at the silhouette contour. Therefore, the joints from Kinect corresponding to the hips were considered to be approximately at the level of the femurs heads, only being marked laterally. The iliac crests were also determined by considering anatomical proportions, being assumed that the vertical distance from the femur head to the iliac crest corresponds to half a head [82]. The knee landmarks (tibiales laterale) were searched in a 10×10 window centered in each knee skeletal joint. The Y coordinate was identified in the point where the distance from the lateral side to the medial side is smaller. The corresponding X coordinate was laterally marked. Since the feet region has a lot of noise, the points identified as corresponding to the ankle joints by the Kinect were used as landmarks and marked laterally. A correspondence between this 2D landmarks and the 3D points from the depth point cloud was performed and, finally, the distance between some pairs of landmarks equivalent to the selected anthropometric lengths and breadths was computed. To estimate the selected heights, the floor level had to be determined. This has been considered to be the lowest value observed for the Y coordinate. 3.3.3 Gender Estimation The anthropometric measures obtained by the previously described method were used to estimate the gender of the subjects. This process was based on the conclusions taken in a previous study [70] that has established which measures were better to distinguish genders and established three different gender discriminant functions based on those measures for Italy, Netherlands and United States of America populations. In order to consider a discriminant function that best fitted the collected dataset, a similar analysis had to be performed based on the Portuguese population. However, this database was not available and, therefore, an approximation had to be considered, having been used the USA discriminant function (Equation 3.1, where wb is waist back length, bb is bustpoint breadth, hb is hip breadth sitting, ac is ankle circumference, bf is bilateral femural epicondyle breadth sitting, cc is chest circumference and cg is chest girth). Three aspects were considered in this selection: •The Italy discriminant function was dependent of external information (such as the weight) and an automatic classifier was desired, implying no inputs from the user or analyzed subject; •The Netherlands’ population largely differs from the Portuguese in terms of anthropometry; •There is a great variability in terms of anthropometric characteristics in the USA population, being less restrictive. 3.3 Anthropometry 39 score =wb ×0.06 +bb ×0.16 −hb ×0.13 +ac ×0.14 +b f ×0.03 −cc ×0.08 +cg ×0.08 (3.1) Applying this function to each subject, the gender was considered to be "Male" if the scored was superior to zero and "Female" if inferior. The estimation of this anthropometric characteristic has, therefore, required the determination of different measures, namely (Figure 3.11): •Waist back length: it was considered to be the vertical distance between the medial point at the estimated iliac crests and the neck-head separation - point at the body’s middle axis belonging to the head and with the lowest y value. •Bustpoint breadth: considering the front view segmentation, it was possible to determine the corners existing in the armpits region. Since the subject’s did not assumed a "T pose", the detected points do not correspond to the real armpits. Instead, they are located in the chest level, being considered as reference to this measure. The 3D distance between these corner points was taken as the bustpoint breadth. •Hip breadth sitting: although the hip breadth of a subject when seated is naturally expect to be larger than standing, this values had to be taken as equivalent, since the algorithm was only adequate to subjects assuming a standing position. Therefore, the 3D distance between the femur heads was considered. However, to improve and validate the obtained results, the error introduced by this approximation had to be estimated in a study comparing the measures of a subject sitting and standing. •Ankle circumference: The ankle landmarks previously determined were used in this step. At the same vertical level, two points were marked (laterally and medially) for each ankle, establishing a diameter. The ankle was considered to be circular and its perimeter was calculated (2×π×(d 2), with d being the ankle’s diameter). •Bilateral femural epicondyle breadth sitting: Taking into account the landmarks that were possible to estimate with the proposed methods, this measure was considered to correspond to the distance between the tibiale lateral points. As referred in the hip breadth sitting case, the errors introduced by this approximation had also to be object of a study designed for this purpose. •Chest circumference: the chest circumference is calculated in the bustpoint line and was, in this study, approximated to an ellipse. In order to determine its perimeter, two axis had to be estimated: the major (from the front view) that was equivalent to the bustpoint breadth and the minor (from the lateral view). The minor axis was calculated in the lateral view by detecting the first maximum point in the horizontal axis to be detected in the silhouette below the neck. The correspondent point in the back was marked and the 3D distance between this points calculated. The circumference measure was calculated by applying an approximation for ellipse perimeter (Equation 3.2, where Mis the length of the major axis and mthe length of the minor). p=2×rM2+m2 2(3.2) •Chest girth: the method applied for estimating of the chest girth was similar to the described for the respective circumference. Chest girth corresponds to the circumference measured in the armpit line. 40 Methodology In the lateral side, there was no exact way to determine the location of this point considering only the RGB information from the arm. Therefore, the scapula protuberance was used to estimate the vertical location of this point, since it was anatomically coincident. Two points were marked in this level (anterior and posterior) and the distance between them taken as the minor axis of the chest girth. The vertical distance between this point (anterior) and the bustpoint level (determined in the for the chest circumference) was used to mark the location of the armpits in the front view, directly above the bustpoint breadth landmarks. The ellipse perimeter was calculated considering the equation 3.2. Figure 3.11: Anthropometric measurements used in gender discrimination: a) waist back length; b) bustpoint breadth; c) hip breadth sitting; d) ankle circumference; e) bilateral femural epicondyle breadth sitting; f) chest circumference; g) chest girth (Adapted from [70]). For estimating the minor axis correspondent to the chest circumference and girth ellipses, the lateral images had to be segmented and processed following a similar method to the applied in the front view: 1. The contour of the subject’s body was obtained by using the RGB image and first apply the Canny edge detector to a defined border. This border was created by considering the subtraction of the result of the dilation of the Kinect mask from the result of the respective erosion. Head, lower legs and feet were removed in this step, since this regions are extremely affected by noise (see Section 3.2). 2. If the resulting contour was already closed, it was used as the ideal silhouette and smoothed, using the same method as in Section 3.2. If the contour had discontinuities, the original mask of the Kinect was considered. Since the measurements extracted from this step were used in an approximated calculation, this approximation was not considered to be relevant. 3. If an improved silhouette was obtained in 2, it is used to reconstruct the pixels considered as background in the Kinect mask and that are seen as foreground in this, using a median filter and following the same procedure applied for the front view. In order to determine this set of anthropometric measures, some approximations had to be considered, which were expected, in advance, to introduce errors in the estimated scores. However, since the final result that was intended to obtain corresponded to a categorical variable (male or female) and not to an accurate 3.4 Identification of Anatomical Control Points in Thermograms 41 continuous value, these approximations were allowed, but as minimized as possible. The contribution of the gender to the extraction of an automated profile of a subject is relevant to any patient evaluation and also justified the use of this type of analysis, even with the needed approximations. 3.3.4 Evaluation of the Anthropometric Assessment For each subject, 90 frames were considered. The algorithm was applied to all the frames separately and a mean value was calculated for each measure. Three different analysis were performed, being two quantitative (1 and 3) and one qualitative (2): 1. In order to first evaluate the efficiency of the applied methodology, the obtained silhouette was used to generate a binary mask (1 for foreground and 0 for background) and compared to a ground truth mask. The ground truth also consisted in a binary mask, manually segmented from the RGB image, using a image editing software 2. Since the algorithm behaves similarly to the different frames from the same subject, only one frame was considered for each case. The performed segmentations were then compared by considering metrics usually applied in binary classification problems - the foreground pixels were seen as positives and the background as negatives. True Positives (TP), False Positives (FP) and False Negatives (FN) are counted and combined to form three metrics: percentage of bad classifications (Equation 3.3), precision (Equation 3.4) and recall(Equation 3.5). The mask more similar to the ground truth presents an inferior value of PBC and superior values of precision and recall. PBC(%) = 100 ×FN +FP T P +FN +FP +TN (3.3) precision =TP T P +FP (3.4) recall =TP T P +FN (3.5) 2. The detected landmarks were marked in 2D RGB images and 3D point clouds in order to make a qualitative evaluation of the obtained results. 3. In order to evaluate the efficiency of the anthropometric measures estimation, using the developed algorithm based on the RGB-D images from Kinect, the mean values obtained for each measure were compared to a ground truth set obtained with Qualisys. 3.4 Identification of Anatomical Control Points in Thermograms The segmentation of thermograms in appropriate regions of interest for further analysis is still a manual procedure that is, not only, time-consuming, but also increases the inter-user variability, biasing the results. In the last years, there has been some research on this topic in order to fully automatize the thermograms analysis methodology [83] . However, there is still no software widely accepted and suitable for all the different thermography applications regarding the human body [28]. 2http://www.gimp.org/ 48 Results Subject 3 When this athlete was first examined, he had suffered an sprain in the left ankle, one month before, and was already recovered and practicing. The thermograms first collected (Day 1) are presented in Figure 4.2. It is possible to observe that there is no visual evidence of injury, since both ankles have a similar thermal distribution. Twenty days after the beginning of the monitoring procedure, Subject 3 suffered a new sprain, this time in the right ankle. The evolution on the thermal symmetry of the ankle region for both front and lateral views is presented in Figure 4.3. He started receiving therapy in the last day of thermal acquisitions. It is possible to observe that, for the first 3 samples, there was a natural variation on the thermal symmetry distribution, but all values were between -0.5 and 0.5. This corroborates the conclusion extracted from thermograms examination: the athlete could be classified as healthy. The samples collected after the injury occur, reveal an increase in the right ankle temperature when compared to the left (the thermal symmetry value is positive), which indicates the presence of inflammation in the surrounding tissues. The injury is more clearly observed in the lateral view, which can indicate that the affected ligaments are located in this side. A different exam had to be conducted to confirm this assumption, for example, an MRI or an ultrasonography. The visual inspection of the thermograms corresponding to the day where the thermal asymmetry was supposed to be higher in both views (Figure 4.2 - Day 25), also shows an asymmetry in the thermal distribution. A superior temperature is observed in the injured ankle, which is evidence that an inflammatory process is occurring in this region. Although not being possible to infer about the physiotherapy effect, it can be observed that a natural recovery process had occurred during the 9 days of follow-up. An extended monitoring period was necessary to extract more accurate conclusions. Figure 4.2: Thermograms of Subject 3 in the different days of examination (1 and 25): A. Front view (day 1). B. Left lateral view (day 1). C. Right lateral view (day 1). D. Front view (day 25). E. Left lateral view (day 25). F. Right lateral view (day 25). 4.1 Thermographic Monitoring of Ankle Sprains 49 Figure 4.3: Graphical representation of the Subject 3 evolution during the thermal examination period. Subject 5 Subject 5 suffered a severe sprain in the left ankle with ligaments tear. The physiotherapist recommended the athlete to rest for 15 days before start receiving therapy. Thermograms of the injured region were captured before the first physiotherapy session (Figure 4.4 - Day 1) and showed evidence of an intense inflammatory process in the affected area. The extreme difference between the temperature recorded in the two ankles was clear in both frontal and lateral views. Figure 4.4: Thermograms of Subject 5 in the different days of examination (1 and 29): A. Front view (day 1). B. Left lateral view (day 1). C. Right lateral view (day 1). D. Front view (day 29). E. Left lateral view (day 29). F. Right lateral view (day 29). 50 Results Thermography was used to monitor the rehabilitation of this subject, being collected a set of images of the injured region before each treatment session. The athlete did not return to practice during the follow-up period. However, it was possible to report a notorious improvement of the sprain due to the applied therapy and performed exercises (Figure 4.5). Figure 4.5: Graphical representation of the Subject 5 evolution during the thermal examination period. There is a clear approximation of the thermal symmetry values to 0, which is due to the decrease in mean temperature of the injured ankle. This decrease is related to the end of the inflammatory process, showing the regression of the sprain. This evolution may also be confirmed through the examination of the thermograms collected in the last day of observations (Figure 4.4 - Day 29). Subject 9 Subject 9 is a case of re-occurrence of an ankle sprain. This athlete had already an history of reincidences in both ankles: 2 sprains in the left and 2 in the right in a period of 4 years. During the thermographic follow-up period, the subject suffered a mild sprain of his right ankle. After receiving only one session of physiotherapy, the athlete reported an improvement and returned to play. However, 29 days after this treatment, the athlete had a relapse, suffering the same type of injury in the same ankle. Thermograms of front and lateral sides from both injuries were collected and are presented in Figure 4.6. The mean temperature values, show in both cases a higher temperature in the right ankle, evidencing the existence of inflammation in this region. Thermal symmetry values (Table 4.5) of the first injury pointed to a severe injury, requiring more attention. 4.1 Thermographic Monitoring of Ankle Sprains 51 Table 4.5: Thermal symmetry values for Subject 9 in Day 1 and Day 29. Day Front (thermal symmetry) Lateral (thermal symmetry) 1 2.89 0.58 29 2.30 0.27 Figure 4.6: Thermograms of Subject 9 in the different days of examination (1 and 29): A. Front view (day 1). B. Left lateral view (day 1). C. Right lateral view (day 1). D. Front view (day 29). E. Left lateral view (day 29). F. Right lateral view (day 29). A different scale was used in this case to ensure a better visualization. Subject 12 Subject 12 suffered an ankle sprain in the right foot one day before the first examination. This injury was considered, by the physiotherapist, to be less severe than the presented by Subject 5. The state of the injury at this point was recorded in thermograms (Figure 4.7) and confirms this assumption. However, it is possible to observe a clear difference between the temperatures of the right and the left ankle, indicating an intense inflammation in the right foot. The last thermograms collected from the subject show an improvement in the affected region, with a decrease in the thermal asymmetry. This evolution is evidence of an adequate physiotherapeutic treatment, with positive and visible results. This recovery process is possible to observe in the graphical representation of the thermal symmetry values presented in Figure 4.8. The values resultant from both views are tending to zero, as rehabilitation occurs. The applied treatment and the natural healing phenomena results in the decrease of the inflammatory process, which decreases the temperature of the injured ankle. 52 Results Figure 4.7: Thermograms of Subject 12 in the different days of examination (1 and 19): A. Front view (day 1). B. Left lateral view (day 1). C. Right lateral view (day 1). D. Front view (day 19). E. Left lateral view (day 19). F. Right lateral view (day 19). Figure 4.8: Graphical representation of the Subject 12 evolution during the thermal examination period. The evolution of the remaining pathological cases could not be further analyzed since they were less severe and the treatment comprised no more than 2 sessions. For Subject 5 and Subject 12, a more detailed analysis was performed, being conducted a MannWhitney test for the thermal symmetry values estimated using a linear regression for the first 19 days of rehabilitation. The resultant p-values obtained from the Mann-Whitney test (Table 4.6) were inferior to 4.2 Anthropometric Assessment with Microsoft Kinect 53 0.05. This indicates that there is enough statistical evidence to conclude that the thermal symmetry values of both views allow to differentiate these two cases. Table 4.6: Results from the Mann-Whitney test conducted for the two cases followed during rehabilitation: p-value, statistic (U) and Z value. Variable p-value U Z Front view (absolute thermal symmetry) 0.000 49.500 -3.826 Lateral view (absolute thermal symmetry) 0.020 101.000 -2.321 4.2 Anthropometric Assessment with Microsoft Kinect The outputs resulting from the different stages of the RGB-D images processing were analyzed and evaluated separately in order to establish which were the most critical steps. 4.2.1 Silhouette Extraction The silhouette obtained from the body segmentation was compared with a ground truth image (manually segmented) and evaluated considering three metrics: precision, recall and percentage of bad classifications. These values were compared to the ones obtained with the mask from Kinect and are summarized in Table 4.7. Table 4.7: Mean results obtained in the metrics used for silhouette evaluation and the respective ratio in number of pixels. Metric Method Result Ratio (pixels) Precision Estimated 0.95 27155.33 /28489.00 Kinect 0.95 26889.00 /28207.67 Recall Estimated 0.96 27155.33 /28200.00 Kinect 0.95 26889.00 /28200.00 PBC Estimated 0.0078 2378.33 /306541.33 Kinect 0.0079 2436.67 /306541.33 The results show that the extracted silhouette has more points coincident with the ground truth than the mask estimated by the Kinect, presenting a higher recall (0.96 instead of 0.95), a lower percentage of bad classifications (0.79% instead of 0.78%) and equivalent precision (0.95 in both cases). Although it appears minimal, this difference is relevant since it is detected at pixel level corresponding to a considerable improvement. It is, therefore, possible to conclude that, with the developed algorithm, an optimization of the original mask was achieved. 4.2.2 Anthropometric Landmarks: Descriptive Evaluation The landmarks obtained with the developed algorithm were marked in the RGB images and in the depth point clouds, providing a 2D and 3D evaluation of the results. For each subject, an example of these two outputs is presented in Figure 4.9, considering the same frame. 54 Results By observing the presented representations, it is possible to conclude that the landmarks were estimated in the expected region. However, a quantitative analysis of the estimated measures needs to be considered for more conclusive observations. Figure 4.9: 2D and 3D representation of the anthropometric landmarks automatically determined with the developed algorithm: A. Grayscale image. B. Point cloud (The (R) and (L) correspond to right and left sides, respectively). 4.2.3 Anthropometric Measures: Quantitative Assessment The mean values obtained for each subject considering the 50 analyzed frames were compared to the correspondent values obtained with the Qualisys system. Absolute and relative errors were determined for each measure of each subject and the mean values considering all subjects were finally determined. These results are presented in Table 4.8 along with the relative errors presented in [21]. Table 4.8: Results obtained from the quantitative assessment of the anthropometric measures estimation. The relative error (%) and the absolute error (millimeters) are estimated by comparing the mean values obtained for all subjects with the results from Qualisys. State-of-art relative errors (%) [21] are also presented to the measures referenced in literature. To these, the lower relative errors are presented in bold. Measures Absolute Error Relative Error Relative Error (mm) (%) from [21] (%) Lengths Tibiale laterale-Sphyrion fibulare 36.01 ±29.89 8.69 ±6.97 - Tronchanterion-Tibiale laterale 22.36 ±14.79 5.27 ±3.40 - Iliocristale-Tronchanterion 11.39 ±10.74 9.48 ±10.41 - Stylion-Dactylion 30.35 ±9.46 16.04 ±4.68 - Radiale-Stylion 16.64 ±11.41 6.64 ±4.42 6.68 Acromiale-Radiale 63.40 ±29.46 20.10 ±9.62 12.76 Breadths Biacromiale 28.80 ±24.01 7.87 ±6.02 13.99 Bitronchanteric 55.61 ±13.28 14.29 ±3.25 15.64 Heights Vertex 75.20 ±31.23 4.40 ±1.70 3.31 Acromiale 72.14 ±28.89 5.16 ±1.99 3.69 Iliocristale 25.96 ±17.14 2.53 ±1.62 7.72 Tibiale laterale 33.14 ±16.57 7.36 ±3.62 6.58 The absolute error is presented to verify how much the estimated measures have varied from the expected, being the iliocristale-tronchanterion the one presenting the smaller error. However, in order to 4.3 Gender Estimation 55 establish a valid comparative analysis between the different measures, the relative error is more appropriate. Considering all measures, the ones presenting the highest relative errors were: the acromiale-radiale, the stylion-dactylion and the bitronchanteric, by descending order. The wrong estimation of the first two measures may be evidence of the existence of a systematic error in the detection of one or more landmarks of the arm. Regarding the bitronchanteric breadth, it important to stress that these points were determined by considering the Kinect skeletal joints correspondent to the hip, only marking them laterally. Therefore, this approximation may be in the origin of the presented error. In order to accurately verify which landmarks have been wrongly determined and minimize the obtained errors, the establishment of a fixed referential for both Qualisys and Microsoft Kinect could be investigated in future work, establishing the error associated with each landmark and improving the detection of the most critical. When comparing the obtained relative errors with the literature, it is possible to observe that the developed algorithm has presented better relative error for half of the compared measures: radiale-stylion length, biacromiale and bitronchanteric breadths and iliocristale height. Errors obtained for the rest of the measures could be minimized in future work through a better determination of the correspondent landmarks: vertex,acromiale,tibiale laterale and radiale, by exploring the color context and making a less use of body proportions approximations, for example. It is also possible to observe that the relative errors obtained for all the estimated heights were inferior to 10%, which was also observed in the values reported in the reference literature. Most part of the remaining measures present superior errors. Samejima et al. [21] suggested that, having inferior lengths, these measures could be more affected by the accuracy limits of the depth sensor. In this study, the same justification can be assumed. 4.3 Gender Estimation Considering the estimated anthropometric measures, it was possible to automatically determine the gender of the tested subjects by observing the score that each one obtained for the discriminant function (Table 4.9). Table 4.9: Results from the gender estimation algorithm. The classification M corresponds to "Male" and F to "Female". Scores and the respective estimated classification from misclassified subjects are presented in bold. Sample Score Estimated Classification Real Classification A-1.79 F F B-21.84 F F C16.96 M M D23.98 M M E-9.92 F F F-2.83 F F G13.60 M M H 32.78 M F I -18.40 F M J-6.78 F F K7.58 M M L10.92 M M 56 Results The gender discrimination method was successfully applied in the most part of the subjects, presenting a percentage of correct classifications of 83%. The two cases where this algorithm failed were examples of the effect that a slightly different posture (Figure 4.10.A) or the presence of large clothes (Figure 4.10.B) may have in this classification. Figure 4.10: Critical aspects identified in the gender discrimination algorithm: A. Pose dependency in the lateral view of Subject I. B. The effect of clothes in the lateral and front view of Subject H. For Subject H, the used clothes has affected the estimation of landmarks in the front view (ankles) and in the lateral view (chest circumferences). The lateral image of Subject I presents the : the point detected in the front side of the chest belonged to the left side of the body, instead of the right side that was expected. It is also important to refer that the used samples were very similar, not presenting marked differences in terms of height, weight or age. In order to be correctly validated, this method had to be tested in a larger population with more distinctive subjects. Nevertheless, this experiment already showed that it is possible to use automatically obtained anthropometric measures for gender discrimination. 4.4 Identification of Anatomical Control Points in Thermograms The estimation of the primary and secondary control points in whole body thermograms of the anterior view, allowed the identification of the most common ROIs. An example of the output of this algorithm is presented in Figure 4.11.A. Connecting the detected points, it is possible to define regions (Figure 4.11.B) that are are close to the expected (see Figure 3.12). Figure 4.11: Anatomical control points estimated in whole body thermograms: A. Set of points detected. B. Definition of regions by connecting related control points. 4.4 Identification of Anatomical Control Points in Thermograms 57 Since this was just a first stage of development, no quantitative analysis was performed. However, the qualitative evaluation indicates that most of the control points are well estimated. Wrists and elbows, being deeply dependent of pose, are the ones less consistent to what was visually expected and could, therefore, be improved in future work. 64 Conclusion in a relevant starting point for the achievement of a fully automated system in thermographic analysis. However, some steps still require manual inputs and must be object of a more thorough study. To conclude, this work has made a positive contribution to the establishment of useful methodologies for injury evaluation, thermograms automated analysis and anthropometric profiling of an athlete, being a pioneering attempt for helping to reduce players absence caused by injury and their associated costs. References [1] P Tate. Seeley’s Principles of Anatomy and Physiology. McGraw-Hill Education, 2011. [2] Y Guillodo, A Le Goff, and A Saraux. Adherence and effectiveness of rehabilitation in acute ankle sprain. Annals of physical and rehabilitation medicine, 54(4):225–235, 2011. [3] L Garfunkel, J Kaczorowski, and C Christy. Pediatric Clinical Advisor: Instant Diagnosis and Treatment. Clinical Advisor Series. CVMOSBY Company, 2007. [4] B Waterman, B Owens, S Davey, M Zacchilli, and P Belmont. The epidemiology of ankle sprains in the united states. The Journal of Bone & Joint Surgery, 92(13):2279–2284, 2010. [5] Y Guillodo, T Simon, A Le Goff, and A Saraux. Interest of rehabilitation in healing and preventing recurrence of ankle sprains. Annals of physical and rehabilitation medicine, 56(7):503–514, 2013. [6] D Fong, L Hong, Yand Chan, Pg Yung, and K Chan. A systematic review on ankle injury and ankle sprain in sports. Sports Medicine, 37(1):73–94, 2007. [7] S Thacker, D Stroup, C Branche, J Gilchrist, R Goodman, and E Weitman. The prevention of ankle sprains in sports a systematic review of the literature. The American journal of sports medicine, 27(6):753–760, 1999. [8] A McIntosh. Rugby injuries. Journal of Science and Medicine in Sport, 49:120–139, 2005. [9] E Verhagen, M Van Tulder, A Van der Beek, L Bouter, and W Van Mechelen. An economic evaluation of a proprioceptive balance board training programme for the prevention of ankle sprains in volleyball. British Journal of Sports Medicine, 39(2):111–115, 2005. [10] E Kemler, I van de Port, F Backx, and C van Dijk. A systematic review on the treatment of acute ankle sprain. Sports medicine, 41(3):185–197, 2011. [11] P Kannus and P Renstrom. Treatment for acute tears of the lateral ligaments of the ankle. operation, cast, or early controlled mobilization. The Journal of bone and joint surgery. American volume, 73(2):305–312, 1991. [12] G Kerkhoffs, H Handoll, R De Bie, B Rowe, and P Struijs. Surgical versus conservative treatment for acute injuries of the lateral ligament complex of the ankle in adults. Cochrane Database Syst Rev, 2, 2007. [13] R Van Rijn, A Van Os, R Bernsen, P Luijsterburg, B Koes, and S Bierma-Zeinstra. What is the clinical course of acute ankle sprains? a systematic literature review. The American journal of medicine, 121(4):324–331, 2008. [14] W Sands. Thermography and return to play decisions. National Strength and Conditioning Association (NSCA). [15] S Albert, M Glickman, and M Kallish. Thermography in orthopedics. Annals of the New York Academy of Sciences, 121(1):157–170, 1964. [16] F Ring and K Ammer. Infrared thermal imaging in medicine. Physiological measurement, 33(3):R33, 2012. 65 66 REFERENCES [17] A Arfaoui, G Polidori, R Taiar, and C Popa. Infrared thermography in sports activity. Infrared Thermography, pages 141–168, 2012. [18] C Hildebrandt, K Zeilberger, F Ring, and C Raschner. The application of medical infrared thermography in sports medicine. Ultrasound, 10:2, 2012. [19] C Hildebrandt, C Raschner, and K Ammer. An overview of recent application of medical infrared thermography in sports medicine in austria. Sensors, 10(5):4700–4715, 2010. [20] A Dantcheva, C Velardo, A D Angelo, and J Dugelay. Bag of soft biometrics for person identification. Multimedia Tools and Applications, 51(2):739–777, 2011. [21] I Samejima, K Maki, M Kagami, Sand Kouchi, and H Mizoguchi. A body dimensions estimation method of subject from a few measurement items using kinect. In Systems, Man, and Cybernetics (SMC), 2012 IEEE International Conference on, pages 3384–3389. IEEE, 2012. [22] M Reyes, A Clapes, J Ramirez, J Revilla, and S Escalera. Automatic digital biometry analysis based on depth maps. Computers in Industry, 64(9):1316–1325, 2013. [23] R Drake, A Vogl, and A Mitchell. Gray’s Anatomy for Students. Gray’s Anatomy. Elsevier Health Sciences, 2009. [24] G Harris and P Smith. Foot and ankle motion analysis: clinical treatment and technology. CRC Press, 2007. [25] T Hubbard and C Hicks-Little. Ankle ligament healing after an acute ankle sprain: an evidence-based approach. Journal of athletic training, 43(5):523, 2008. [26] E Holme, S Magnusson, K Becher, T Bieler, P Aagaard, and M Kjaer. The effect of supervised rehabilitation on strength, postural sway, position sense and re-injury risk after acute ankle ligament sprain. Scandinavian journal of medicine & science in sports, 9(2):104–109, 1999. [27] R van Rijn, J van Ochten, P Luijsterburg, M van Middelkoop, B Koes, and S Bierma-Zeinstra. Effectiveness of additional supervised exercises compared with conventional treatment alone in patients with acute lateral ankle sprains: systematic review. BMJ: British Medical Journal, 341, 2010. [28] R Vardasca and R Simoes. Current issues in medical thermography. In Topics in Medical Image Processing and Computational Vision, pages 223–237. Springer, 2013. [29] E Arens and H Zhang. The skin’s role in human thermoregulation and comfort. Center for the Built Environment, 2006. [30] A Szentkuti, H Skala Kavanagh, and S Grazio. Infrared thermography and image analysis for biomedical use. Periodicum biologorum, 113(4):385–392, 2011. [31] Y Houdas and E Ring. Human body temperature-its measurement and regulation. 1982. [32] F Ring. The historical development of thermal imaging in medicine. Rheumatology, 43(6):800–802, 2004. [33] F Achmidt, R Henderson, and C Wolgemuth. Introduction to thermal science. John Wiley & Sons, New York, 1993. [34] M Planck. On the law of distribution of energy in the normal spectrum. Annalen der Physik, 4(553):1, 1901. [35] R Vardasca. The effect of work related mechanical stress on the peripheral temperature of the hand. 2010. [36] F Ring. Thermographic terminology. Acta Thermographica, pages 1–30, 1978. REFERENCES 67 [37] J Engel. Thermography in locomotor diseases: recommended procedure. European Journal of Rheumatology and Inflammation, 2(3):299–306, 1979. [38] K Ammer. The glamorgan protocol for recording and evaluation of thermal images of the human body. Thermology international, 18(4):125–144, 2008. [39] F Ring, K Ammer, B Wiecek, P Plassmann, C Jones, Anna Jung, and P Murawski. Quality assurance for thermal imaging systems in medicine. Therm, 156:103, 1997. [40] R Simpson, H McEvoy, G Machin, K Howell, M Naeem, P Plassmann, F Ring, P Campbell, C Song, and J Tavener. In-field-of-view thermal image calibration system for medical thermography applications. International Journal of Thermophysics, 29(3):1123–1130, 2008. [41] W Sands, J McNeal, and M Stone. Thermal imaging and gmnastics injuries: A means of screening and injury identification. Science of Gymnastics Journal, 3(2), 2011. [42] J Selfe, C Sutton, N J Hardaker, S Greenhalgh, A Karki, and P Dey. Anterior knee pain and cold knees: A possible association in women. The knee, 17(5):319–323, 2010. [43] Y Asagai, A Imakiire, and T Oshiro. Thermographic study of low level laser therapy for acute phase injury. In Presentation at World Association of Laser Therapy annual conference in Athens, Greece, volume 1000, 2000. [44] M Schmitt and Y Guillot. Thermography and muscular injuries in sports medicine. In Recent advances in medical thermology, pages 439–445. Springer, 1984. [45] R Vardasca. Thermal symmetry of the limbs in healthy subjects. Thermology international, 2012. [46] R Simoes, R Vardasca, and C Nogueira-Silva. Thermal skin reference values in healthy late pregnancy. Journal of Thermal Biology, 37(8):608–614, 2012. [47] A Jain, Ar Ross, and S Prabhakar. An introduction to biometric recognition. Circuits and Systems for Video Technology, IEEE Transactions on, 14(1):4–20, 2004. [48] B Munsell, A Temlyakov, C Qu, and S Wang. Person identification using full-body motion and anthropometric biometrics from kinect videos. In Computer Vision–ECCV 2012. Workshops and Demonstrations, pages 91–100. Springer, 2012. [49] C Velardo and J Dugelay. Real time extraction of body soft biometric from 3d videos. In Proceedings of the 19th ACM international conference on Multimedia, pages 781–782. ACM, 2011. [50] M McDowell and National Center for Health Statistics (US). Anthropometric reference data for children and adults: United States, 2003-2006. US Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Health Statistics, 2008. [51] R Araujo, G Grana, and V Andersson. Towards skeleton biometric identification using the microsoft kinect sensor. In Proceedings of the 28th Annual ACM Symposium on Applied Computing, pages 21–26. ACM, 2013. [52] C Velardo and J Dugelay. Weight estimation from visual body appearance. In Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE International Conference on, pages 1–6. IEEE, 2010. [53] A Jain, S Dass, and K Nandakumar. Soft biometric traits for personal recognition systems. In Biometric Authentication, pages 731–738. Springer, 2004. [54] S Samangooei, B Guo, and M Nixon. The use of semantic human description as a soft biometric. In Biometrics: Theory, Applications and Systems, 2008. BTAS 2008. 2nd IEEE International Conference on, pages 1–7. IEEE, 2008. 68 REFERENCES [55] C BenAbdelkader and L Davis. Estimation of anthropomeasures from a single calibrated camera. In Automatic Face and Gesture Recognition, 2006. FGR 2006. 7th International Conference on, pages 499–504. IEEE, 2006. [56] R Lima. Adaptação ergonómica e antropométrica de calçado para pessoas com necessidades especiais: um estudo de caso. 2012. [57] M Marfell-Jones and International Society for Advancement of Kinanthropometry. International Standards for Anthropometric Assessment. International Society for the Advancement of Kinanthropometry, 2006. [58] I Westat. National health and nutrition examination survey iii: body measurements (anthropometry). 1988. [59] M Robinson and M Parkinson. Estimating anthropometry with microsoft kinect. 2013. [60] K Khoshelham and S Elberink. Accuracy and resolution of kinect depth data for indoor mapping applications. Sensors, 12(2):1437–1454, 2012. [61] Z Ben Azouz, C Shu, and A Mantel. Automatic locating of anthropometric landmarks on 3d human models. 2006. [62] C Martin, D Burkert, K Choi, Ni Wieczorek, P McGregor, R Herrmann, and P Beling. A realtime ergonomic monitoring system using the microsoft kinect. In Systems and Information Design Symposium (SIEDS), 2012 IEEE, pages 50–55. IEEE, 2012. [63] R Clark, Y Pua, K Fortin, C Ritchie, K Webster, L Denehy, and A Bryant. Validity of the microsoft kinect for assessment of postural control. Gait & posture, 36(3):372–377, 2012. [64] Prime sensor nite 1.3 algorithms notes. Framework Programmer’s Guide - Prime Sense, 2010. [65] K Robinette, H Daanen, and E Paquet. The caesar project: a 3-d surface anthropometry survey. In 3-D Digital Imaging and Modeling, 1999. Proceedings. Second International Conference on, pages 380–386. IEEE, 1999. [66] H Lin, H Lu, and L Zhang. A new automatic recognition system of gender, age and ethnicity. In Intelligent Control and Automation, 2006. WCICA 2006. The Sixth World Congress on, volume 2, pages 9988–9991. IEEE, 2006. [67] A Jain, K Nandakumar, X Lu, and U Park. Integrating faces, fingerprints, and soft biometric traits for user recognition. In Biometric Authentication, pages 259–269. Springer, 2004. [68] G Shakhnarovich, PA Viola, and B Moghaddam. A unified learning framework for real time face detection and classification. In Automatic Face and Gesture Recognition, 2002. Proceedings. Fifth IEEE International Conference on, pages 14–21. IEEE, 2002. [69] S Gutta, J Huang, P Jonathon, and H Wechsler. Mixture of experts for classification of gender, ethnic origin, and pose of human faces. Neural Networks, IEEE Transactions on, 11(4):948–960, 2000. [70] A Fullenkamp, K Robinette, and H Daanen. Gender differences in nato anthropometry and the implication for protective equipment. Technical report, DTIC Document, 2008. [71] R Brown, D Ubelaker, and M Schanfield. Evaluation of purkait’s triangle method for determining sexual dimorphism. Journal of forensic sciences, 52(3):553–556, 2007. [72] Christopher A King, MY I¸scan, and Susan R Loth. Metric and comparative analysis of sexual dimorphism in the thai femur. Journal of forensic sciences, 43(5):954–958, 1998. [73] P Barrio, G Trancho, and J Sánchez. Metacarpal sexual determination in a spanish population. Journal of forensic sciences, 51(5):990–995, 2006. REFERENCES 69 [74] World Medical Association. Declaration of helsinki, ethical principles for medical research involving human subjects. 52 nd WMA General Assembly, Edinburgh, Scotland, 2000. [75] Assembleia da Republica Portuguesa. Lei n. 21/2014 de 16 de abril. Diario da Republica, 2014. [76] A Rovai, J Baker, and M Ponton. Social Science Research Design and Statistics: A Practitioner’s Guide to Research Methods and IBM SPSS. Watertree Press LLC, 2013. [77] L Martinez. Analise de Dados com SPSS. Escolar Editora. [78] Microsoft developer network - msdn library: Kinect sensor. Microsoft Corporation. [79] H Zhou and H Hu. Human motion tracking for rehabilitation - a survey. Biomedical Signal Processing and Control, 3(1):1–18, 2008. [80] Lu Xia, Chia-Chih Chen, and JK Aggarwal. Human detection using depth information by kinect. In Computer Vision and Pattern Recognition Workshops (CVPRW), 2011 IEEE Computer Society Conference on, pages 15–22. IEEE, 2011. [81] J Hamm. Drawing the Head and Figure. A GD/Perigee book. Perigee Books, 1983. [82] P Richer and R Hale. Artistic Anatomy. Watson-Guptill Publications, 1986. [83] A Conci, R Lima, C Fontes, L Motta, and R Resmini. A new method for automatic segmentation of the region of interest of thermographic breast image. Thermology International, 20(4):134–135, 2010. [84] D Carmona, M Quintana, D Salces, D Rodrigo, and P León. La termografia como medio de prevencion, recuperacion y seguimiento de lesiones en futbolistas. [85] J Sampedro, S Piñonosa, I Fernandez, et al. Thermography as a new assessment tool in basketball. pilot study carried out with a professional player in the acb. Cuadernos de Psicología del Deporte, 12(Suppl. 1):51–56, 2012. [86] R Vardasca, F Ring, P Plassmann, and C Jones. Thermal symmetry of the upper and lower extremities in healthy subjects. Thermology International, 2012. [87] S Uematsu. Symmetry of skin temperature comparing one side of the body to the other. Thermology, 1(1):4–7, 1985. [88] P Goodman, M Murphy, G Siltanen, M Kelley, and L Rucker. Normal temperature asymmetry of the back and extremities by computer-assisted infrared imaging. Thermology, 1(4):194–202, 1986. [89] S Uematsu, D Edwin, W Jankel, J Kozikowski, and M Trattner. Quantification of thermal asymmetry: Part 1: Normal values and reproducibility. Journal of neurosurgery, 69(4):552–555, 1988. [90] L Schwarz, A Mkhitaryan, D Mateus, and N Navab. Human skeleton tracking from depth data using geodesic distances and optical flow. Image and Vision Computing, 30(3):217–226, 2012. [91] I Fernandez-Cuevas, J Marins, P Carmona, M Garcia-Concepcion, J Lastras, and M Quintana. Reliability and reproducibility of skin temperature of overweight subjects by an infrared thermography software designed for human beings. Thermology International, 2012. 70 REFERENCES 71 Appendix A. Assessment Sheet Questionário Género Idade Altura Peso Medicamentos Tempo de experiência como atleta Dominância (superior/inferior) Lesões e Cirurgias Já sofreu alguma lesão? Quantas? Indique os tipos de lesão e há quanto tempo ocorreram (aproximadamente). Já foi sujeito a alguma cirurgia? Quantas? Indique as cirurgias realizadas e há quanto tempo ocorreram (aproximadamente) 72 Por favor, escolha a opção adequada. MOBILIDADE Não tenho problemas em percorrer nenhuma distância. Tenho alguma dificuldade em percorrer metros. Estou confinado a uma cama. CUIDADOS PESSOAIS Não tenho problemas com cuidados pessoais. Tenho alguma dificuldade em vestir-me e lavar-me sozinho. Não sou capaz de me vestir ou lavar sozinho. ATIVIDADES QUOTIDIANAS (ex.: trabalho, estudo, trabalhos domésticos, etc.) Não tenho dificuldades em executar as minhas atividades quotidianas. Tenho alguma dificuldade em executar as minhas atividades quotidianas. DOR/DESCONFORTO Não sinto dor nem desconforto. Sinto dor ou desconforto moderado. Sinto dor ou desconforto extremo. ANSIEDADE/DEPRESSÃO Não sou ansioso nem depressivo. Sou moderadamente ansioso ou depressivo. Sou extremamente ansioso ou depressivo. Note que este questionário é anónimo. O seu único propósito é identificar algumas informações do sujeito relevantes para o estudo e garantir que este satisfaz os requisitos estabelecidos. 73 B. Informed Consent INFORMAÇÃO AO PARTICIPANTE Objetivo do estudo O presente estudo tem como objetivo principal, a avaliação da utilização da termografia na monitorização de lesões durante o processo de fisioterapia, nomeadamente de entorses de tornozelo. O que é a termografia de infravermelhos? A termografia de infravermelhos é uma técnica que permite a análise quantitativa e precisa da distribuição da temperatura à superfície da pele. É uma técnica que utiliza apenas a radiação infravermelha que é emitida pelo corpo humano, isto é, é um método não-invasivo, não-ionizante, portanto seguro, que capta a energia natural que é libertada pelo corpo, não apresentando qualquer risco para a saúde do participante. Porquê o estudo com a termografia de Infravermelhos? Sendo a temperatura um parâmetro fisiológico, através da recolha de imagens por termografia, é possível quantificar objetivamente algumas lesões existentes que alteram o padrão normal de distribuição da temperatura nas regiões onde ocorrem, como é o caso das entorses de tornozelo. O que tenho que fazer? (a) Não consumir refeições pesadas, álcool, bebidas quentes ou fumar durante as 2 horas anteriores às recolhas. (b) Não deve ser aplicado qualquer tipo de creme na pele antes da recolha de imagens. (c) É necessário que antes da participação neste projeto responda a um inquérito e assine o consentimento informado. (d) Serão recolhidas imagens a partir de uma câmara Microsoft Kinect, sendo necessário que permaneça imóvel durante 20 segundos (em cada aquisição). (e) Anteriormente à recolha das imagens é necessário que proceda a uma climatização de 10 minutos. (f) Durante a recolha dos termogramas, é necessário que permaneça relativamente imóvel. Quanto tempo irá demorar? Todo o processo demorará cerca de 45 minutos. O que acontecerá à informação do meu exame? As imagens capturadas serão apenas utilizadas para investigação. A identificação do participante será armazenada com segurança e é estritamente confidencial. CONTACTO: Ana Domingues - [email protected]