A Test Statistic Estimation-Based Approach for Establishing Self-Interpretable CNN-Based Binary Classifiers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) are crucial for medical image classification but often lack interpretability.
• Post-hoc interpretability methods can yield ambiguous or conflicting explanations, hindering clinical adoption.
• High-stakes medical decisions require transparent and reliable AI models.

Purpose of the Study:

• To develop a self-interpretable model for binary medical image classification.
• To address the ambiguity and limitations of existing post-hoc interpretability methods.
• To provide a quantitative and self-contained interpretability solution for black-box DNNs.

Main Methods:

• A decision-theory-inspired approach using a self-interpretable encoder-decoder model.
• Integration with a single-layer fully connected network with unity weights.
• Training the model to replicate the test statistic of a pre-trained black-box classifier.
• Generating an 'equivalency map' for visualization and quantification of feature contributions.

Main Results:

• The proposed self-interpretable model achieved comparable accuracy to the original black-box classifier.
• Equivalency maps provided clear visualizations of image features driving classification decisions.
• The method allowed for quantitative assessment of the relative importance of these features.
• Successful application demonstrated across three distinct medical image binary classification tasks.

Conclusions:

• The novel approach establishes a self-interpretable, quantitative method for DNNs in medical imaging.
• This technique overcomes limitations of post-hoc methods, reducing ambiguity in model explanations.
• The equivalency map offers a powerful tool for understanding and trusting AI-driven medical diagnoses.