Self-Supervised Masked Image Encoding with Compact Convolutional Heads for Patient-Disjoint Kidney CT Diagnosis Under Limited Labeled Cohorts

• Ada John

  ladoke Akintola university of technology

  Author

Kidney computed tomography (CT) classification has become a focal point of medical imaging research, yet existing benchmarks suffer from methodological vulnerabilities—particularly patient-level data leakage—that inflate reported performance and undermine clinical translation. This study introduces a self-supervised masked image encoding framework with compact convolutional classification heads, evaluated under a rigorously patient-disjoint protocol on a multi-center kidney CT dataset comprising 12,446 images across four diagnostic categories (Normal, Cyst, Tumor, Stone). The proposed methodology combines Masked Autoencoder (MAE) pre-training with a lightweight depthwise-separable CNN head (≈1.46M parameters), trained on 9,956 patient-disjoint images and validated on 2,490 held-out images. Results demonstrate that the self-supervised approach achieves 89.4% accuracy (95% CI: 87.1–91.6%), with macro ROC-AUC of 0.962 and expected calibration error of 0.041, substantially outperforming random-initialized baselines (78.2%) and approaching fully supervised transfer learning (91.7%) while requiring only a fraction of labeled data. The patient-disjoint protocol reveals that accuracy inflation from slice-level leakage can exceed 15 percentage points, validating concerns raised in recent methodological critiques. The framework offers a scalable, annotation-efficient solution for kidney CT diagnosis in resource-constrained settings, with implications for both clinical deployment and the establishment of methodologically rigorous benchmarks in medical image classification.

Copyright (c) 2026 Ada John (Author)

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