Classification of pathologies in chest x-rays using deep learning techniques

The lungs are the primary organs of the respiratory system and can lose functionality due to structural disorders, infections, or environmental factors. Lung diseases such as COVID-19, tuberculosis, pneumonia, pulmonary fibrosis, atelectasis, cardiomegaly, and pneumothorax may lead to fatal outcomes if not diagnosed at an early stage. Therefore, early and accurate diagnosis of lung diseases is crucial. Chest radiographs (CXRs) are widely used for this purpose due to their advantages, such as low cost, quick applicability, and low radiation dose. However, factors like overlapping anatomical structures, limited specialists, and high image volume make interpreting CXRs challenging, resulting in a daily error rate of 3–5%. Thus, the demand for automated and reliable diagnostic systems to assist radiologists in interpreting CXRs is increasing. In this thesis, an end-to-end deep learning model was developed for high-performance classification of CXRs, and novel loss functions were proposed to enhance the model's performance. The pre-trained DenseNet201 architecture was used as the backbone. A Convolutional Long Short-Term Memory (ConvLSTM) layer to obtain richer spatial and temporal information from the feature maps obtained from DenseNet201, a Squeeze and Excitation (SE) block to emphasize channel-wise features, and Vision Transformers (ViT) to capture long-range dependencies were integrated into the model. In addition, a Global Average Pooling (GAP) layer was used to preserve important spatial information for classification. The combination of these components enabled high classification performance. The study also examined the often-overlooked role of loss functions in classification performance. Beyond standard functions, hybrid and dynamic structures were designed. A custom function with a penalty term was proposed to improve sensitivity to misclassification. Ensemble approaches combining the strengths of proposed loss functions were implemented. These loss functions were systematically analyzed and tested with various models, demonstrating architecture-independent success. To evaluate the performance and generalization ability of the proposed model and loss functions, a comprehensive dataset was created by combining clinical images from Başkent University Ankara Hospital with open-access datasets. This dataset was divided into subsets containing 7, 8, 9, 10, 12, 14, and 15 classes, and extensive testing was conducted on each subset. Results showed that the proposed method outperforms many existing approaches in the literature, achieving superior accuracy, F1-score, and AUC. The model's consistent performance in various scenarios demonstrated strong generalization. In conclusion, this thesis presents a novel deep learning model and original loss functions that produce high-performance and reproducible results for the automatic classification of CXRs.