Automated Tracheal Morphometrics Using Deep Learning: Toward Custom Tracheostomy Tubes

Automated Tracheal Morphometrics Using Deep Learning: Toward Custom Tracheostomy Tubes Bram Tona,∗, Sally Farooqb, Jelmer Veenstrac aAmbient Intelligence, Saxion University of Applied Sciences, M.H. Tromplaan 28, Enschede, 7513 AB, Overijsel, The Netherlands bMedisch Spectrum Twente, Koningsplein 1, Enschede, 7512 KZ, Overijsel, The Netherlands cIndustrial Design, Saxion University of Applied Sciences, M.H. Tromplaan 28, Enschede, 7513 AB, Overijsel, The Netherlands Abstract Tracheostomy is a life-saving procedure with complications often linked to suboptimal tracheostomy tube fit. Customised tubes could reduce these risks, but their development requires precise tracheal morphometric data. This study presents an automated workflow for extracting high-resolution tracheal measurements from CT scans, using 3D Slicer and a deep learning segmentation model (VISTA-3D). One hundred anonymised scans were analysed, with 90 included after exclusions. Automated measurements revealed average tracheal lengths of 66.1±12.2 mm, anteroposterior dimensions of 16.2±2.5 mm, transverse dimensions of 17.9±2.4 mm, and maximum inscribed circle diameters of 14.1±1.9 mm. Sex-specific differences were observed, with males exhibiting larger dimensions. The results align with existing literature, validating the automated approach. This method enables efficient, objective, and high-resolution morphometric analysis, supporting the development of demographic-specific tracheostomy tubes and improving pre-operative planning. Keywords: trachea, morphometrics, computational, tracheostomy ∗Corresponding author Email addresses: [email protected] (Bram Ton), [email protected] (Sally Farooq), [email protected] (Jelmer Veenstra) Preprint submitted to medRxiv.org November 27, 2025 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. 1. Introduction Tracheostomy is a surgical technique in which an opening is created in the anterior wall of the trachea to facilitate airflow, often in the context of upper airway obstruction. It can be performed either via an open surgical technique or percutaneous techniques. Indications can be emergent, such as in cases of acute upper airway obstruction or traumatic injury, or elective, for example in cases of chronic aspiration, neurological conditions, or prolonged ventilator dependence [1]. Reported rates of complications vary greatly [2]. Complications of tracheostomy are generally divided into transoperative, early, and late. Transoperative complications include haemorrhage (most common), pneumothorax, injury to the recurrent laryngeal nerve, oesophageal perforation [3] and intraoperative loss of airway [4]. Loss of airway can result from obstruction of the tracheostomy tube lumen by blood clots or mucus. A number of the late complications of tracheostomy arise from prolonged contact between the cannula and the tracheal wall, resulting in local trauma. Continuous friction is thought to cause granulation tissue formation, which can mature to form a fibrotic tissue layer, potentially resulting in tracheal stenosis. Tracheal stenosis resulting in lumen occlusion from 60-70% to total occlusion is particularly unfortunate [4]. Reported rates of clinically significant tracheal stenosis range from 3 to 12% [5]. The use of an oversized tube is a risk factor [5]. Malposition of the distal cannula tip against the posterior tracheal wall is another risk factor for trauma, a problem especially relevant in obese patients with increased neck thickness [3]. Chronic irritation can also lead to chondritis, weakening the tracheal rings and potentially resulting in tracheomalacia [3]. Among the most severe late complications is tracheoinnominate fistula, a rare but life-threatening event. It develops from local trauma to the anterior tracheal wall [6] and excessive cannula movement [3] leading to erosion into the innominate artery [3, 6]. The risk is particularly high when the tracheostomy is placed low (below the 3rd third ring) [3]. Another uncommon late complication is tracheo-oesophageal fistula, which likely occurs secondary to chronic trauma and ischaemia of the posterior tracheal wall [3]. Choice of tracheostomy tube size is important to mitigate the risk of aforementioned complications. Parameters to consider include tracheal diameter, tracheal length (proximal, i.e. depth of soft tissue in the neck and distal), and the angle at which the tube sits in the trachea [7]. How well tube 2 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint dimensions fit tracheal dimensions is therefore an important consideration, yet standard sizing assumes a relatively uniform anatomy despite the limited data available on normal adult tracheal measurements. Ideally, each patient should have access to tailor-made tracheostomy tubes to ensure a perfect fit, and thus mitigate the risk of complications and increase the quality of life. Though many advances have been made in this direction, especially with the rise of additive manufacturing, mass adoption is still far ahead. A compromise in this regard is to offer a set of products tailored to specific target groups. To identify these target groups and obtain the optimal parameters, a large number of tracheal measurements have to be made. CT scans provide a valuable source of information to obtain these parameters. Obtaining representative measurements for each target group, requires large numbers of measurements. Obtaining these measurements manually would be too resource intensive, hence this article explores how these measurements can automated. 2. Related work Several factors influencing the dimensions of the trachea are sex [8, 9], the phase of pulmonary ventilation [10, 11], and tracheomalacia [11]. The work of Ebrahimian et al. focuses on tracheal dimensions in light of detecting tracheomalacia [11]. Their work compares manual and automated measurements during both phases of pulmonary ventilation. For manual measurements, they use the aortic arch as a reference for taking a singlesection area measurement. For the automated measurement, they define average metrics for the anteroposterior (AP) dimension and surface area along the length of the trachea. They define the length of the trachea as the path from the thoraic inlet to a point slightly above the carina. For the normal patients the AP dimension ranged from ≈16–18 mm, and the surface area for this group ranged from 204–273 mm2. Kamel et al. makes a comparison between morphometrics derived from CT scans and cadaver data [9]. Based on the analysis of 60 CT scans, an average maximum AP dimension of 21.4±3.2 mm is found and an average maximum transverse dimension of 25.7±3.7 mm is found. In their study the patients scanned were required to hold their breath during scanning. Based on cadaver data from 10 subjects, an AP dimension of 20.8±2.9 mm was 3 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint found for men, for women this was 15.5±1.1 mm. The measured transverse dimension for men was 21.4±1.6 mm and for women was 17.8±1.5 mm. Prince and Stark’s main focus was relating the dimensions of the thyroid gland to the dimensions of the trachea [12]. The transverse dimension of the trachea was measured at the point were the thyroid had the largest transverse dimension. Their study examined 118 patients, with the majority being male (115). They found an average transverse tracheal dimension of 19.97 ±2.37 mm. The work of El-Anwar et al. measure the AP and transverse dimensions at the level of the thyroid gland [8]. They analysed 100 scans and found an average AP dimension of 16.56 ±2.77 mm and an average transverse dimension of 15.5±2.1 mm. Considering that this study was conducted in Egypt, genetic and geographic factors might impact the dimensions. Green et al. present a small pilot study in which they used CT scans to preoperatively select the appropriate tracheostomy tube based on several key measurements [13]. In this study, 8 out of 11 patients had the correct tube size at first trial. Most works still rely on manual workflows for morphometry of the trachea. This work fills this gap by presenting an automated workflow. Automatisation enables analysing large numbers of scans with little costs, and can provide fine grained morphometrics per demographic group. Furthermore, it enables a higher spatial resolution, a large number of measurements can be made which would otherwise be too time consuming. Higher spatial resolution means that the tracheal shape deformation along the trachea can be modelled. 3. Materials and methods 3.1. Data The initial data set consists of 100 CT scans provided by the Medisch Spectrum Twente (MST), The Netherlands. An internal ethics committee of the MST granted permission to use the data. A selection of recent adult neck scans was provided by the radiology department. The data was anonymised before being exported, only the year of birth and sex remained in the metadata. A demographic breakdown of the subjects included in this research is given in Figure 1. The figure shows that the data is skewed towards males and towards a slightly elderly part of the population. 4 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint Figure 1: Demographic breakdown 3.2. Methods The open source program Slicer was used [14] to automate the morphometric measurements of the trachea. This software was used because the processing steps available in the graphical user interface (GUI) can also be automated using Python code. The major advantage of this was that parameter settings and procedures could easily be evaluated visually in short iteration cycles. Thereafter these same steps can be added to the Python script to automate the workflow. 3.2.1. Segmentation The first step was the conversion of the data from the Digital Imaging and Communications in Medicine (DICOM) format to the Neuroimaging Informatics Technology Initiative (NIfTI) format. This was done using the dcm2niix tool. This conversion was required as the segmentation model works with this format. For the segmentation itself, the VISTA-3D segmentation model [15] was used. This segmentation model is a deep learning based foundation model capable of segmenting 127 classes from CT scan data, including the trachea and thyroid. The segmented trachea starts roughly at the thyroid cartilage 5 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint and includes part of both primary bronchi. The segmentation results are represented by an integer label per voxel. Cleaning. The segmentation process is not perfect, hence some minor cleaning operations are required next. The cleaning operations make use of the segmentation toolbox [16] of Slicer. The regions at the extents of the trachea can contain small noisy blobs which are part of the segmentation. To remove these, we retain the largest blob only. In the case of the thyroid, the segmentation sometimes misses the region around the isthmus. This implies that the left and right lobe are not always connected, hence the strategy of retaining only the largest blob will not work. In this case the strategy is to remove all blobs below a certain volume threshold. The threshold parameter is defined in voxels. Based on the image spacing and a target threshold volume of 1 cm3, the target number of voxels for the threshold is determined. The segmented trachea sometimes contains small holes. These holes interfere with the determination of the cross sectional area. A morphological closing operation is performed using a 4 mm kernel to ‘fill’ these holes. 3.2.2. Centreline After segmentation and cleaning, the next step is to extract the centreline of the trachea. This centreline is used to define at which positions the cross sectional areas of the trachea will be determined. The Vascular Modeling Toolkit (VMTK) extension of Slicer is used for determining the centreline and calculating the cross sectional areas. The VMTK was designed for analysing blood vessels [17], but also works well for other tubular structures such as the trachea. This toolkit is used to extract the centrelines of the trachea, and usually results in at least three lines. One for the main part of the trachea and two for the left and right parts of the primary bronchi. The longest line, representing the main part of the trachea, is selected for further processing. To be able to derive some statistics of the cross sectional areas, it is important that they are all measured at the same location along the centreline. The measurements start at a point just above the carina and move upward towards the larynx. The extracted centreline is resampled so that the new control points have a fixed distance of 5 mm. Centreline cut-off. The centreline extraction near the extents of the segmented trachea are not well defined. This is caused by an irregular segmentation at these areas. To compare individual trachea to each other, it is 6 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint (a) Oriented bounding box (b) Plane Figure 2: End-point landmark determination of tracheal centreline important to define a consistent end-point along the centreline. A landmark used to determine this end-point is the plane which is perpendicular to the centreline and just touches the lowest part of the segmented thyroid. The normal of this plane is defined by the vector starting at the carina (index 0) and pointing to the point which is 20 mm away from the carina (index 4). The point used in describing the plane is the coordinate of the lowest part of the segmented thyroid (Figure 2b). Using this plane, the end-point is defined as the last point of the centreline which still resides below the defined plane.An initial approach to determine this end-point was to use the lowest plane of the oriented bounding box (Figure 2a) of the segmented thyroid. This approach was not successful as the segmentation of the thyroid was not always complete. 3.2.3. Metrics This section describes several of the metrics which are derived from the CT scans in an automated way. First is the cross-sectional area of the trachea at fixed intervals, thereafter the diameter of the maximum inscribed circle is described. At last the measurement of the anteroposterior and transverse dimensions are elaborated. Figure 3 provides an overview of how some of the metrics are determined. A cross section at each node of the centreline is defined, but in the figure only three are displayed as an example. 7 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint Figure 3: Centreline with examples of cross sections Length. In our work we define the length of the trachea as the summation of all distances between control points of the centreline up to the previously defined end-point of the trachea (See Section 3.2.2). Area. To obtain the actual cross sectional area, the internal workings of the toolkit define a cross cut perpendicular to the centreline at the position of the specified control point. The internal workings of the Slicer program can be called from within a Python script, hence it is possible to extract the cross sectional areas of the trachea in this way. Maximum inscribed circle. The maximum inscribed circle (MIC) of is an important parameter as it directly relates to the diameter of the tracheostomy tube. The MIC is determined by first storing a binary image of the tracheal cross-section. Cross-sectional images are captured at each control point along the centreline of the trachea. After this a distance transform is applied to the binary images, the maximum value of this transform determines the radius of the maximum inscribed circle in pixels. The final step is a unit conversion from pixels to mm. This is done using the spacing information from the CT scan file. Anteroposterior and Transverse dimensions. The same binary image can be used to determine the anteroposterior (AP) and the transverse (T) dimension 8 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint of the trachea. To determine the AP dimension, the binary image is summed in the first dimension and the maximum value is used. To determine the transverse dimension, the binary image is summed in the second dimension and the maximum value is determine once more. Conversion from pixel units to mm is once again done using the spacing information from the CT file. 4. Results 4.1. Exclusion Out of the 100 subjects analysed, it was not possible to define a conclusive centreline cut-off point for four of the subjects. Out of these, two subjects has an insufficient complete segmentation of the upper part of the trachea which made it impossible to find the last point of the centreline which still resided below the thyroid plane. In one case the subject was intubated during the CT scan, which led to a corrupted segmentation of the trachea. In the other case there was a paratracheal air cyst [18] present which confused the centreline finding algorithm. These four subjects are therefore not used for further analysis. Each of the scans has also been manually inspected to verify if a significant part of the main bronchi was present in the scan. Presence of these bronchi is necessary to determine the point just above the carina. In four cases the bronchi was not present, and in two cases the bronchi part was not sufficient enough for the centreline algorithm to work. These six cases have also been excluded from further analysis. The total number of subjects included for statistical analysis is therefore now 90. 4.2. Metrics This section dives into the details of the following metrics of the trachea: length, area, minimum inscribed circle (MIC), the anteroposterior dimension and the transverse dimension. 4.2.1. Length A histogram describing the trachea length is given in Figure 4. We see that the majority of the length measurements fall within the range 60–70 mm. 9 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint [16] C. Pinter, A. Lasso, G. Fichtinger, Polymorph segmentation representation for medical image computing, Computer Methods and Programs in Biomedicine 171 (2019) 19–26. doi:10.1016/j.cmpb.2019.02.011. [17] R. Izzo, D. Steinman, S. Manini, L. Antiga, The Vascular Modeling Toolkit: A Python Library for the Analysis of Tubular Structures in Medical Images, Journal of Open Source Software 3 (25) (2018) 745. doi:10.21105/joss.00745. [18] H.-J. Bae, E.-Y. Kang, H. S. Yong, Y. K. Kim, O. H. Woo, Y.-W. Oh, K. W. Doo, Paratracheal air cysts on thoracic multidetector CT: Incidence, morphological characteristics and relevance to pulmonary emphysema, The British Journal of Radiology 86 (1021) (2013) 20120218– 20120218. doi:10.1259/bjr.20120218. 16 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint The copyright holder for thisthis version posted November 30, 2025. ; https://doi.org/10.1101/2025.11.27.25341143doi: medRxiv preprint