Calibrated Feature Attribution and Concept Bottleneck Models for Interpretable Kidney Pathology Classification: Bridging Spatial Gate Attention and Radiologist Diagnostic Workflows

• Ada John

  ladoke Akintola university of technology

  Author

Deep learning models for kidney pathology classification have achieved strong discrimination performance, yet their clinical adoption remains constrained by limited interpretability, poor probability calibration, and misalignment with radiologist diagnostic reasoning. This study addresses these gaps by proposing an integrated framework that combines calibrated feature attribution with concept bottleneck models (CBMs) and spatial gate attention mechanisms for interpretable kidney pathology classification. Using a multi-center, patient-disjoint abdominal CT dataset comprising four diagnostic categories (Normal, Cyst, Tumor, Stone), we systematically evaluate whether post-hoc temperature scaling and CBM-based concept supervision can improve both calibration and clinical trustworthiness without sacrificing classification accuracy. Our results demonstrate that the proposed framework achieves 89.4% accuracy with an Expected Calibration Error (ECE) of 0.032, substantially outperforming baseline CNN architectures (84.7% accuracy, ECE = 0.089) and standard Grad-CAM attribution methods in anatomical faithfulness. Concept bottleneck supervision improved radiologist-rated explanation utility by 41% compared to saliency-only approaches. These findings establish that calibrated, concept-grounded feature attribution provides a practical bridge between high-performance kidney pathology classifiers and radiologist diagnostic workflows, offering actionable design principles for clinically deployable AI systems.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.