Concept Bottleneck Optimization Algorithms For High Fidelity Medical Image Interpretation

Abstract

Interpretable prediction in high-stakes medical imaging tasks is a promising paradigm that can be achieved through Concept Bottleneck Models (CBMs). While promising in theory, the current CBMs suffer from conceptual lacunae, annotation limitations, and suboptimal convergence in complex multi-pathology scenarios, all of which reduce diagnostic fidelity. The research proposes the Concept Bottleneck Optimization Algorithm (CBOA), a novel framework that combines adaptive multi-scale feature extraction, semantic concept alignment, gradient-based concept pruning, and a residual concept learner that learns concepts to boost concept detection accuracy and downstream diagnostic performance simultaneously. The CBOA achieves state-of-the-art AUC scores of 95.6% for the NIH Chest X-ray14 and 94.9% for the ISIC 2020 dermoscopy corpus, with concept accuracy scores of 91.8% and 90.6%, respectively. Comparative analyses with standard deep learning classifiers and previous CBM variants demonstrate statistical significance for both CBOA's interpretability and classification fidelity. The algorithm addresses the known problem of concept rigidity in predefined concepts by adding a complementary residual pathway that discovers latent concepts not in the manually annotated list. The results confirm the clinical feasibility of the CBOA system for accurate and transparent medical image diagnosis.