Intersection of Performance, Interpretability, and Fairness in Neural Prototype Tree for Chest X-Ray Pathology...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep learning models achieve high accuracy in chest X-ray (CXR) analysis but lack transparency, hindering clinical adoption.
• The neural prototype tree (NPT) is introduced as an interpretable classifier combining deep learning diagnostic power with decision tree interpretability for CXR pathology detection.

Purpose of the Study:

• To evaluate the neural prototype tree (NPT) classifier's performance, interpretability, and fairness in CXR pathology detection.
• To examine the interplay between these three dimensions and highlight local and global explanations of the NPT classifier.

Main Methods:

• Utilized Chest X-ray 14, CheXpert, and MIMIC-CXR datasets for training.
• Compared a baseline ResNet-152 with five NPT classifiers of varying interpretability levels.
• Measured performance (ROC AUC), interpretability (IC), and fairness (mean TPR disparity) using linear regression analysis.

Main Results:

• NPT classifier performance increased with interpretability, outperforming ResNet-152 at specific IC levels across datasets.
• Lower interpretability (IC level 1) correlated with higher unfairness (mean TPR disparity), particularly in age-based subgroups.
• Significant positive relationships were found between interpretability and performance (ROC AUC), and a negative relationship between interpretability and fairness (mean TPR disparity).

Conclusions:

• The NPT classifier offers a balance between performance, interpretability, and fairness for CXR pathology detection.
• Findings provide insights for developing effective, interpretable, and equitable deep learning models in medical imaging.
• The study highlights the potential clinical utility of transparent AI in radiology.