Clinical domain knowledge-derived template improves post hoc AI explanations in pneumothorax classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Thoracic Diseases

Background:

• Deep learning (DL) models are increasingly used for diagnosing pneumothorax, an acute thoracic condition.
• Explainable AI (XAI) methods aim to clarify DL model decisions but often produce explanations diverging from actual lesion areas.
• Improving the clinical relevance of AI-driven diagnostic explanations is crucial for reliable medical applications.

Purpose of the Study:

• To enhance the quality and clinical accuracy of AI-generated explanations for pneumothorax diagnosis.
• To develop a template-guided approach that integrates radiological knowledge into XAI methods.
• To reduce discrepancies between AI model explanations and actual pneumothorax lesion locations.

Main Methods:

• A template-guided approach was developed, using radiologist-delineated lesion areas to create templates for potential pneumothorax regions.
• Templates were superimposed on XAI-generated explanations to filter out irrelevant or extraneous information.
• The efficacy of this method was evaluated by comparing three XAI techniques (Saliency Map, Grad-CAM, Integrated Gradients) with and without template guidance on two DL models (VGG-19, ResNet-50) using two datasets (SIIM-ACR, ChestX-Det).

Main Results:

• The template-guided approach consistently improved the performance of baseline XAI methods across various scenarios.
• Significant performance gains were observed, with average incremental improvements of 97.8% in Intersection over Union (IoU) and 94.1% in Dice Similarity Coefficient (DSC).
• Visualizations demonstrated that template guidance more accurately aligned model explanations with ground-truth lesion areas on radiographs.

Conclusions:

• A novel template-guided approach was proposed to enhance AI model explanations for pneumothorax diagnosis.
• This method successfully integrates clinical insights, improving the alignment of AI explanations with radiological findings.
• The approach shows potential for broader application in explaining AI models for other thoracic diseases, fostering trust through clinical domain expertise integration.