Expert-Guided Explainable Few-Shot Learning with Active Sample Selection for Medical Image Analysis.

Area of Science:

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

Background:

• Medical AI deployment is limited by scarce labeled data and poor model interpretability.
• Few-shot learning (FSL) addresses data scarcity but lacks prediction transparency.
• Active learning (AL) optimizes data acquisition but neglects sample interpretability.

Purpose of the Study:

• To develop a dual-framework solution for interpretable and data-efficient medical image analysis.
• To integrate expert knowledge and explainability into few-shot and active learning paradigms.
• To improve clinical AI deployment by tackling data limitations and enhancing model transparency.

Main Methods:

• Expert-Guided Explainable Few-Shot Learning (EG-FSL) uses radiologist-defined regions-of-interest for spatial supervision via Grad-CAM-based Dice loss.
• Explainability-Guided Active Learning (EG-AL) iteratively acquires samples based on predictive uncertainty and attention misalignment.
• A closed-loop framework synergistically combines explainability for training and sample selection.

Main Results:

• EG-FSL achieved high accuracy: 92% on BraTS, 76% on VinDr-CXR, and 62% on SIIM-COVID, outperforming baselines.
• EG-AL reached 76% accuracy with 680 samples, significantly better than random sampling (57%).
• Grad-CAM visualizations confirmed models focused on diagnostically relevant regions, with cross-modality validation on breast ultrasound.

Conclusions:

• The proposed dual-framework effectively addresses key challenges in medical image analysis.
• EG-FSL and EG-AL offer interpretable and data-efficient solutions for clinical AI.
• The framework demonstrates strong performance across multiple datasets and modalities, paving the way for improved AI deployment.