Generative Diffusion Augmentation for Long-Tailed Kidney Pathology Datasets: Impact on Calibration-Aware Patient-Disjoint Generalization in Deep Convolutional Classifiers

• Ada John

  ladoke Akintola university of technology

  Author

Long-tailed class distributions and patient-level data leakage remain critical barriers to clinically reliable deep learning in kidney pathology. Random slice-level dataset partitioning inflates reported accuracy by allowing images from the same patient to appear in both training and test sets, while severe class imbalance biases classifiers toward majority categories. This study investigates whether class-conditional diffusion augmentation can improve calibration-aware, patient-disjoint generalization on long-tailed kidney pathology datasets. Using a 12,446-image multi-class kidney CT cohort with naturally imbalanced distribution, we implement a patient-disjoint evaluation protocol and compare a compact depthwise-separable CNN trained with diffusion-generated synthetic samples against conventional augmentation baselines. The diffusion-augmented model achieved 89.4% patient-disjoint accuracy with expected calibration error of 0.041, representing a 7.2 percentage point improvement over the non-augmented baseline (82.2%) and a 38% reduction in calibration error relative to standard augmentation. Tail-class recall improved from 61.3% to 78.7%. These findings demonstrate that generative diffusion augmentation, combined with patient-disjoint evaluation and explicit calibration reporting, substantially improves both discrimination and reliability for long-tailed kidney pathology classification, offering a reproducible framework for clinically trustworthy deployment in renal imaging AI.

• pdf

Copyright (c) 2026 Ada John (Author)

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