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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Interstitial lung disease (ILD) diagnosis from chest radiographs is challenging.
• Deep learning (DL) shows promise for ILD detection but lacks transparency.
• Explainability of DL models is crucial for clinical trust and decision-making.

Purpose of the Study:

• To develop a trustworthy and robust ILD detection system using a four-class classification scheme.
• To evaluate the explainability of convolutional neural network (CNN) and transformer-based DL models for ILD.
• To enhance the clinical utility of DL in ILD diagnosis through improved transparency.

Main Methods:

• Developed a four-class classification scheme for ILD detection.
• Evaluated explainability of CNN and transformer models using Intersection over Union (IoU) and Dice coefficients.
• Quantified heatmap consistency with actual lesion areas across five visualization methods.

Main Results:

• The four-class scheme significantly increased IoU and Dice coefficients for CNN models with four visualization methods.
• Transformer models showed improved IoU and Dice coefficients with specific visualization methods and the four-class scheme.
• Highest mean IoU of 0.356 and Dice coefficient of 0.483 were achieved, demonstrating enhanced explainability.

Conclusions:

• Multi-class classification leveraging lesion area information enhances DL model explainability for ILD detection.
• Selecting appropriate visualization methods is critical for specific DL models.
• Findings support improved clinical decision-making for earlier and more reliable ILD management.