Conformal Prediction and Bayesian Uncertainty Quantification in Compact CNNs for Out-of-Distribution Kidney Pathology Detection in Non-Contrast and Contrast-Enhanced CT

• Abilly Elly

  Texas University

  Author

Deep learning models for kidney pathology classification on computed tomography (CT) have demonstrated high discriminative accuracy, yet their clinical deployment remains constrained by overconfidence on out-of-distribution (OOD) inputs and inadequate uncertainty quantification. This study investigates the integration of conformal prediction (CP) and Bayesian uncertainty quantification (BUQ) techniques with a compact depthwise-separable convolutional neural network (CNN) for multiclass kidney pathology detection across non-contrast and contrast-enhanced CT protocols. Using a patient-disjoint, group-stratified benchmark of 12,446 kidney CT images spanning four diagnostic categories (Normal, Cyst, Tumor, Stone), we evaluate the calibration and OOD detection performance of Mondrian inductive conformal prediction, Monte Carlo dropout, and deep ensembles applied to the NephroNet architecture . The framework achieves 89.4% classification accuracy on the held-out test set while maintaining empirical coverage probabilities exceeding 92% across all conformal prediction configurations. Under OOD conditions simulated through Gaussian noise corruption and contrast protocol mismatch, the combined CP-BUQ approach yields AUROC scores of 0.847 for OOD detection, substantially outperforming softmax confidence baselines. These findings demonstrate that statistically grounded uncertainty quantification can be achieved without sacrificing the computational efficiency of compact CNN architectures, offering a practical pathway toward reliable clinical decision support in radiology.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

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