Score-Based Counterfactual Generation for Interpretable Medical Image Classification and Lesion Localization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) show promise for clinical decision-making in biomedical imaging.
• High-quality and sufficient data are essential for DNN performance, but medical imaging data is often scarce and imbalanced.
• External factors can introduce uncertainties and unnatural features, impacting data distribution estimation.

Purpose of the Study:

• To propose a novel framework for generating counterfactual medical images to address data scarcity and imbalance.
• To integrate a FuzzyBlock to manage uncertainties from external physical factors.
• To evaluate the framework's effectiveness in both classification and lesion localization tasks.

Main Methods:

• Developed a score-based counterfactual generation (SCG) framework to create synthetic images from latent space.
• Integrated a learnable FuzzyBlock into the classifier to handle data uncertainties.
• Applied the SCG framework to classification and lesion localization tasks in biomedical imaging.

Main Results:

• The SCG framework significantly improved classification task performance.
• An average performance enhancement of 3-5% was observed compared to state-of-the-art methods.
• The framework demonstrated effectiveness in interpretable lesion localization tasks.

Conclusions:

• The proposed SCG framework effectively compensates for data scarcity and imbalance in biomedical imaging.
• The integration of FuzzyBlock helps manage uncertainties, leading to more robust DNN performance.
• The framework offers a promising solution for enhancing DNNs in clinical decision-making and medical image analysis.