An Interpretable and Accurate Deep-Learning Diagnosis Framework Modeled With Fully and Semi-Supervised Reciprocal...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Medical Image Analysis

Background:

• Deep learning classifiers can improve clinical diagnosis but face a trade-off between accuracy and interpretability.
• Existing accurate models often lack transparency, hindering clinical adoption.
• Interpretable models may not achieve competitive diagnostic performance.

Purpose of the Study:

• To introduce InterNRL, a novel deep learning framework designed for high accuracy and interpretability in clinical diagnosis.
• To address the limitations of current deep learning models in balancing classification performance and explainability.
• To enhance the reliability and trustworthiness of AI in medical decision-making.

Main Methods:

• Developed InterNRL, a student-teacher framework utilizing an interpretable prototype-based classifier (ProtoPNet) as the student and a global image classifier (GlobalNet) as the teacher.
• Implemented a novel reciprocal learning paradigm for mutual optimization between student and teacher models.
• Optimized the framework under both fully- and semi-supervised learning scenarios.

Main Results:

• InterNRL achieved state-of-the-art classification performance in breast cancer and retinal disease diagnosis under both supervised learning settings.
• The framework demonstrated superior performance in breast cancer localization and brain tumor segmentation using weakly-labeled images.
• The reciprocal learning approach enabled effective knowledge transfer and model refinement.

Conclusions:

• InterNRL successfully integrates high accuracy and interpretability in deep learning for clinical diagnosis.
• The proposed reciprocal learning paradigm offers a flexible and effective approach for training accurate and interpretable medical AI models.
• InterNRL shows significant potential for advancing automated diagnosis and medical image analysis in clinical practice.