Human-in-the-Loop Discrepancy Discovery: Systematic Exploration of AI–Clinician Divergences in Diagnostic Imaging

Building on the “investigate deviations from expectations” heuristic and the work of Musthafa et al. (2024) and Elsharkawy et al. (2024), who both aim to align AI explanations with clinical reasoning, this idea proposes a structured, interactive system. Rather than passively measuring XAI “fidelity,” the system would mine real-world diagnostic cases for significant mismatches between AI model explanations (e.g., Grad-CAM or SHAP) and radiologist rationale. These discrepancies would be systematically categorized (e.g., due to data bias, model overfitting, imaging artifact, or novel pathology), and the results used to both improve model architecture/explanation techniques and to inform clinicians about atypical or ambiguous cases. This approach creates a feedback loop, turning unexpected results into learning opportunities and new knowledge, and could accelerate both model development and scientific discovery in biomedical imaging.

References:

1. Enhancing brain tumor detection in MRI images through explainable AI using Grad-CAM with Resnet 50. Mohamed Musthafa M, M. T R, V. V, Suresh Guluwadi (2024). BMC Medical Imaging.
2. A Clinically Explainable AI-Based Grading System for Age-Related Macular Degeneration Using Optical Coherence Tomography. M. Elsharkawy, A. Sharafeldeen, F. Khalifa, A. Soliman, A. Elnakib, M. Ghazal, A. Sewelam, Aristomenis Thanos, Harpal S. Sandhu, A. El-Baz (2024). IEEE journal of biomedical and health informatics.