Balancing Performance and Interpretability in Medical Image Analysis: Case study of Osteopenia.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Machine Learning Interpretability

Background:

• Convolutional neural networks (CNNs) show high accuracy in medical diagnosis but often function as "black boxes."
• This lack of transparency can lead to predictions based on irrelevant image features, raising concerns about reliability.
• Identifying and mitigating confounding variables is crucial for trustworthy AI in clinical settings.

Purpose of the Study:

• To explore methods for enhancing the interpretability of CNNs used in medical image analysis.
• To investigate the impact of occluding confounding variables on osteopenia prediction models.
• To assess whether obscuring irrelevant image regions improves the clinical relevance of AI predictions.

Main Methods:

• Utilized the GRAZPEDWRI-DX dataset for osteopenia prediction.
• Developed image masking techniques to occlude identified confounding variables.
• Trained and evaluated CNN models on both original and occluded images using F1-score, precision, and recall.
• Employed GRAD-CAM for visualizing model focus and conducted radiologist preference tests.

Main Results:

• Models trained on non-occluded images generally showed higher performance metrics (F1-score, precision, recall).
• Radiologists, when evaluating model focus via GRAD-CAM, showed a preference for models trained on occluded images.
• Occluding confounding variables shifted model attention to potentially more clinically relevant image regions.

Conclusions:

• Occluding confounding variables in medical images can decrease overall predictive performance but significantly enhance model interpretability.
• This approach encourages AI models to focus on diagnostically relevant features, leading to more trustworthy predictions.
• Balancing predictive accuracy with interpretability is key for the clinical adoption of AI in medical diagnostics.