Exploring the Impact of Self-Supervised Learning and Swin-B Transformer on Saliency Maps and Performance in Multi-Label Chest X-Ray Classification

Chest X-rays (CXRs) are among the most commonly performed radiological exams world-wide due to their cost-effectiveness and low radiation exposure, making them indispensable for evaluating, diagnosing, and monitoring cardiopulmonary diseases. Despite their widespread use, interpreting CXRs remains challenging because of the complex anatomy and overlapping structures, necessitating decision-support tools for clinicians. Recent advances in deep learning have shown that certain architectures can surpass the diagnostic performance of average radiologists; however, the influence of model architecture choice, combined with or without self-supervised learning, on both diagnostic accuracy and saliency map interpretability remains underexplored. In this study, we systematically evaluated various deep learning architectures with transfer learning and self-supervised learning for CXR classification and quantitatively assessed their interpretability using saliency maps. Our results demonstrate that a shifted window transformer pre-trained via self-supervised learning achieved the highest area under the receiver operating characteristic curve (AUC) of 0.83 and an intersection over union (IoU) of 0.16, indicating superior performance in detecting and classifying CXR pathologies. However, this model was slower during inference and more memory-intensive than DenseNet121. To facilitate clinical translation, we also developed an open-source, user-friendly web application allowing clinicians without programming expertise to test the models on their own data. Overall, these findings underscore the promise of transformer-based architectures in conjunction with self-supervised learning in enhancing diagnostic accuracy while emphasizing the need to balance performance gains with computational efficiency.