XTC-Net: an explainable hybrid model for automated atelectasis detection from chest radiographs

Atelectasis, characterized by partial or complete lung collapse, presents notable challenges in both diagnosis and treatment. Timely identification is essential to avoid further pulmonary complications and to facilitate early intervention. Leveraging artificial intelligence for the automated detection of atelectasis can significantly improve diagnostic efficiency, reduce clinical workload, and enhance patient care. This work presents an interpretable deep learning model that synergistically integrates Xception, Transformer, and Capsule Network components for the accurate detection of atelectasis from chest radiographs. The Xception module is employed to extract spatially rich features, while the Transformer component models long-range dependencies critical for understanding complex anatomical patterns. The Capsule Network further enhances the system's sensitivity to subtle structural variations associated with atelectatic regions. The training and validation of the model were conducted using a publicly accessible chest X-ray dataset, achieving impressive performance metrics: 99.73% accuracy, 99.74% sensitivity, and an F1 score of 99.73%. Furthermore, to evaluate the model's generalizability, external validation was performed using the NIH ChestX-ray dataset, which demonstrated consistent performance and highlighted the applicability of the proposed approach beyond the primary dataset. These results underscore the model's capability for reliable and interpretable automated diagnosis, supporting its future integration into clinical workflows.