PROCESSING AND ANALYSIS OF PULMONARY DISEASE DIAGNOSTICS BASED ON CONVOLUTIONAL NEURAL NETWORKS

SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 10 OCTOBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 39 PROCESSING AND ANALYSIS OF PULMONARY DISEASE DIAGNOSTICS BASED ON CONVOLUTIONAL NEURAL NETWORKS T.M. Magrupov1, R.R. Akhmadjonov2, M.U. Nosirov3 Tashkent State Technical University, Tashkent, Uzbekistan1,2,3 https://doi.org/10.5281/zenodo.17389860 Abstract. The article investigates the methodological and algorithmic foundations of processing and analyzing biomedical images of pulmonary diseases. Methods, training algorithms, and architecture of convolutional neural networks for classifying lung images are developed. A database of biomedical images of pulmonary diseases is formed with their subsequent processing. A method for digitizing images and clinical data is proposed to obtain data for models in a form convenient and ready for analysis. The proposed algorithms provide automatic extraction and analysis of key features in images. Keywords: neural network model, image processing and analysis, lung disease classification, model training and evaluation. Introduction. In recent years, lung diseases have ranked among the leading causes of mortality and disability worldwide. According to the World Health Organization (WHO), pulmonary and respiratory diseases remain one of the most common causes of death, affecting millions of people annually [1–3]. WHO data indicate that more than 3 million people die each year from various pulmonary disorders, including pneumonia, fibrosis, tuberculosis, and chronic obstructive pulmonary disease (COPD). These diseases not only lead to serious complications but also reduce the quality of life of patients and increase healthcare costs. The rapid growth of biomedical imaging technologies such as multislice computed tomography (MSCT) has enabled the collection of high-resolution lung images, allowing clinicians to detect and analyze structural changes in lung tissue. However, the manual analysis of such images requires significant time and expert experience, and subjective interpretation may cause diagnostic errors. Therefore, there is an urgent need to develop automated systems that can process, analyze, and classify biomedical images with high accuracy. Recent advances in artificial intelligence, particularly in the field of deep learning, have opened new opportunities in medical diagnostics. Convolutional neural networks (CNNs) are especially effective in image recognition and classification tasks, as they automatically extract and learn hierarchical features from raw data without the need for manual feature engineering [4–7]. CNN-based approaches have demonstrated promising results in various domains, including radiology, histopathology, and pulmonary disease detection. The purpose of this study is to apply CNN-based methods to the analysis of biomedical lung images, enabling the automatic identification and classification of pulmonary diseases based on medical imaging data. Research Objective. The objective of this study is to develop a convolutional neural network model for the analysis, processing, and classification of biomedical lung images to determine the type of diseases. The research focuses on combining image preprocessing techniques, deep learning model SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 10 OCTOBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 40 design, and model evaluation within an integrated framework. The scientific novelty of the work lies in the creation of a database of lung images obtained from multislice computed tomography scans, the introduction of an image digitization approach that ensures compatibility with CNN input requirements, and the development of a customized CNN architecture adapted for lung disease classification. The proposed solution provides automated feature extraction and supports the accurate differentiation of various pulmonary conditions. To achieve the research goal, the following main tasks were completed: formation and preprocessing of biomedical lung image datasets, conversion of the data into a digital numerical format suitable for computational analysis, design of the CNN architecture, training the model on representative datasets, and evaluation of the model’s performance through experimental validation. Materials and Methods. The study focuses on the structure and functioning of convolutional neural networks (CNNs) for the classification of pulmonary diseases based on medical imaging data. In medical diagnostics, CNNs have proven to be among the most efficient tools for analyzing complex visual data due to their capability to automatically extract meaningful hierarchical features from raw images. This approach significantly reduces the dependence on manual feature engineering and expert-defined parameters, which are often time-consuming and prone to human error. In the context of lung disease classification, the CNN architecture designed in this research is composed of multiple interconnected layers, each responsible for performing a specific transformation on the input data. The process begins with low-level feature extraction, where the network identifies fundamental image elements such as edges, contours, and texture variations that correspond to anatomical structures in the lungs. Subsequent layers perform mid-level and highlevel pattern recognition, capturing more abstract representations such as tissue density, lesion boundaries, and structural anomalies indicative of disease. The overall workflow of the proposed methodology includes several key stages: formation of a biomedical image database, image preprocessing, data digitization and normalization, CNN architecture design, model training, evaluation, and validation (Figure 1). Each stage is crucial to ensuring the model’s stability, reproducibility, and diagnostic accuracy. The database creation stage ensures the collection and organization of high-quality data, while preprocessing enhances image clarity and removes noise. Data digitization converts medical images into numerical tensors suitable for machine computation. The CNN architecture design stage focuses on determining the optimal layer configuration and hyperparameters that balance accuracy and computational efficiency. Model training involves feeding the prepared data into the neural network to enable learning through iterative optimization. Finally, model validation ensures that the trained model performs reliably when applied to previously unseen data. Database Formation. The data collection process is crucial for the successful classification of lung images using artificial intelligence. In this study, the neural network model is trained on images obtained through multislice computed tomography (MSCT) in DICOM format for the detection of lung diseases. At the same time, the main focus was placed on: data anonymity and confidentiality; conversion of the DICOM format into other formats; integration with metadata; elimination of data distortions; relabeling and balancing; data size and structure; and methods of database augmentation. Image Preprocessing ensures maximum efficiency in the model training process. The collected MSCT images were processed before they could be directly used for the models. Data Digitization. SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 10 OCTOBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 41 After the preprocessing phase, both medical images and associated clinical parameters were systematically transformed into numerical formats suitable for computational analysis. Each radiographic image was converted into a multidimensional numerical array (tensor), where every pixel intensity was represented as a numerical value reflecting grayscale or color channel information. This tensor-based representation enabled the model to interpret visual data in a structured mathematical form, facilitating feature extraction and pattern recognition during the training process. Fig. 1 – Algorithm of the Biomedical Lung Image Classification Process To complement imaging data, clinical parameters such as patient age, gender, smoking history, respiratory rate, oxygen saturation, and diagnostic test results were carefully encoded into numerical vectors using standard data encoding methods. Categorical variables were transformed through one-hot encoding or label encoding, while continuous parameters were standardized or normalized to a common scale. This ensured that both image and non-image data could be processed jointly by the neural network without bias toward particular ranges of values. Data normalization played a critical role in maintaining numerical stability and convergence efficiency during training. Each tensor and numerical feature was scaled to have zero mean and unit variance, which optimized the performance of gradient-based optimization algorithms such as Adam or SGD. The transformed tensors were further converted into GPUcompatible formats, enabling high-speed matrix operations and parallelized learning through deep learning frameworks. Designing the Convolutional Neural Network Architecture. A convolutional neural network (CNN) architecture was carefully designed to perform automated classification of pulmonary diseases based on medical imaging data. The architecture follows a hierarchical design, where each layer progressively extracts and abstracts features from the input images, enabling the model to learn both local and global patterns related to lung pathology. The proposed CNN begins with an input layer, which receives preprocessed image tensors of a fixed spatial dimension. Following this, multiple convolutional layers apply learnable filters (kernels) to detect low-level visual features such as edges, corners, and textures. As the network deepens, subsequent convolutional layers capture more complex and semantically rich features, including nodule SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 10 OCTOBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 42 shapes, opacity regions, and tissue density variations patterns that are clinically indicative of pulmonary disorders. To reduce spatial dimensionality while preserving essential information, max-pooling layers were inserted after selected convolutional blocks. This not only decreased computational cost but also provided translation invariance, allowing the network to recognize patterns regardless of their position in the image. Additionally, batch normalization layers were used to stabilize the learning process, improving convergence speed and generalization. To address potential overfitting — a common issue in deep networks trained on medical datasets — dropout layers were integrated at strategic points within the architecture. By randomly deactivating a fraction of neurons during training, dropout encouraged the network to learn more robust and generalizable representations of disease features. The fully connected (dense) layers served as the final stage of reasoning, combining the extracted features into a compact decision vector. The output layer employed a softmax activation function, producing a probability distribution across all disease categories (e.g., normal, pneumonia, tuberculosis, COVID-19, etc.), which enabled interpretable classification outcomes. The CNN’s performance was optimized through systematic hyperparameter tuning, including adjustments to the learning rate, batch size, kernel size, and number of epochs. The Adam optimizer was utilized for efficient gradient descent with adaptive learning rates, and categorical cross-entropy was selected as the loss function to handle multi-class classification effectively. Experimental analysis confirmed that the designed CNN successfully learned to detect key radiological features associated with pathological conditions such as nodules, consolidations, fibrosis, and abnormal lung textures, demonstrating strong potential for supporting clinical diagnostic processes. Model Training. To identify the types of lung diseases, the model was trained on images obtained through MSCT. The data were divided into training and testing sets using train_test_split, and the model was trained on the training dataset. Model Evaluation. After training, the model was evaluated for its effectiveness on the test data. At this stage, the model demonstrated its ability to correctly classify lung diseases based on the test dataset. Furthermore, the model was validated on real-world data in clinical and practical trials. Conclusion. In this study, lung images obtained through multislice computed tomography (MSCT) were analyzed to investigate cases of fibrosis and pneumonia, with the aim of developing an automated diagnostic framework based on deep learning. A convolutional neural network (CNN) architecture was designed and implemented for the analysis, processing, and classification of pulmonary diseases. The system incorporated a comprehensive workflow that included dataset construction, image preprocessing, data digitization, model training, and performance evaluation. The experimental findings demonstrated that the proposed CNN-based approach achieved high efficiency, robustness, and accuracy in distinguishing between normal and pathological lung conditions. The architecture effectively learned discriminative features from medical images including structural irregularities, tissue density variations, and the presence of nodular or fibrotic patterns which are key diagnostic indicators in radiological assessments. The results confirm that convolutional neural networks represent a reliable and scalable tool for the automatic detection and classification of lung diseases, significantly reducing manual workload and diagnostic SCIENCE AND INNOVATION INTERNATIONAL SCIENTIFIC JOURNAL VOLUME 4 ISSUE 10 OCTOBER 2025 ISSN: 2181-3337 | SCIENTISTS.UZ 43 subjectivity. Furthermore, the integration of the developed methodology into computer-aided diagnostic (CAD) systems can enhance the diagnostic process by providing decision support to radiologists, improving consistency and reducing the likelihood of human error. REFERENCES 1. Levitan, R., Stephen, J. The Importance of Lung Sound Analysis in Medical Diagnosis. Journal of Pulmonary Medicine, vol. 23, no. 5, pp. 345–357, 2020. 2. Wang, Y., Chen, X. Deep Learning Applications in Medical Diagnosis: A Comprehensive Review. International Journal of Medical Informatics, vol. 112, pp. 66–77, 2019. 3. Lee, K.S. Neural Networks for Automated Respiratory Sound Analysis. Biomedical Signal Processing and Control, vol. 51, pp. 79–85, 2019. 4. Singh, P., Kumar, A. Challenges in Respiratory Sound Data Collection for Disease Classification. Respiratory Medicine and Research, vol. 42, no. 2, pp. 103–110, 2021. 5. Ghorbani, M., Amir, H.M. Assessment of Acoustic Features in Respiratory Sound Analysis. IEEE Transactions on Biomedical Engineering, vol. 67, no. 9, pp. 2895–2903, 2020. 6. He, Q.-H., Yu, B., Hong, X., Lv, B., Liu, T., Ran, J., Bi, Y.-T. An Improved Lung Sound Denoising Method by Wavelet Packet Transform with PSO-based Threshold Selection. Intelligent Automation & Soft Computing, vol. 24, no. 2, pp. 223–230, 2018. 7. Magrupov, T.M., Nazirov, R.M., Abdullaev, I.N. Formation of a Database of Lung Disease Sound Signals. Science and Innovation, International Scientific Journal, vol. 3, issue 9, pp. 90–96, September 2024. 8. Magrupov, T.M., Nurillaeva, N.M., Akhmadjonov, R.R., Zubaydullaev, Sh.Sh., Gaibnazarov, S.S., Semenova, E.A. Algorithmic and Software Implementation of Biomedical Lung Image Classification Technology (Biomedical Engineering). Meditsinskaya Tekhnika, no. 2, pp. 18– 21, 2025.