Enhancing Explainability in AI-Based Interstitial Lung Disease Detection via Multi-Class Learning Based on Lesion...

Interstitial lung disease (ILD) comprises a group of lung disorders characterized by inflammation and fibrosis of the lung interstitium. Early detection is essential, yet identifying ILD on chest radiographs remains challenging. Recently, deep learning (DL) technologies have shown potential to improve diagnostic performance; however, their lack of transparency remains a major concern. This study employed a four-class classification scheme to develop a trustworthy and robust detection system for ILD. The explainability of three convolutional neural network (CNN) models and two transformer-based models was evaluated using Intersection over Union (IoU) scores and Dice coefficients, which quantified the consistency between actual lesion areas and the activation regions in heatmaps generated by five different visual explanation methods. The proposed four-class classification scheme yielded statistically significant increases in IoU scores and Dice coefficients for four visualization methods applied to the three CNN models, while maintaining equivalent classification performance compared with conventional binary classification. For the transformer-based models, the application of certain visualization methods to the four-class classification models also improved IoU scores and Dice coefficients, with the highest mean IoU and Dice coefficient reaching 0.356 and 0.483, although not across all combinations of visualization methods and DL models. These results demonstrate the utility of multi-class classification that leverages lesion area information to enhance model explainability. Furthermore, they underscore the importance of selecting suitable visualization methods for respective DL models. These findings may provide valuable insights to strengthen clinical decision-making, ultimately facilitating earlier detection and more reliable management of ILD.