Uncertainty Quantification in Medical Image Classification: A Study on ResNet-50 Model Based on Conformal Prediction

This article presents an innovative study that applies Conformal Prediction to medical image classification tasks. Using the ResNet-50 architecture on the TissueMNIST dataset, it achieves prediction set outputs with statistical guarantees, offering new ideas for the reliability of medical AI decision-making. The study focuses on solving the "black box" problem of traditional deep learning models and improving the safety of clinical applications through uncertainty quantification.

In the field of medical diagnosis, the reliability of AI models is directly related to patient safety. Traditional deep learning classifiers output a single label without a credibility measure, posing clinical risks. Uncertainty quantification is a direction to address this pain point. As a framework with statistical guarantees, Conformal Prediction can add probabilistic guarantees without changing the training process. This study, conducted by Sedef Kjamili, applies it to histopathological image classification, demonstrating the potential of medical AI.

ResNet-50 was selected as the base classifier and trained on the TissueMNIST dataset (which contains 28×28 grayscale images of 8 tissue types including adipose tissue, background, and connective tissue). Three Conformal scoring methods were compared: LAC (compact sets, average ~2 labels), APS (highest coverage rate of 94.39%, average ~2.6 labels), and Top-K (intuitive but with large set fluctuations). A class-conditional Conformal Prediction strategy was adopted to ensure reliable coverage of rare categories.

Evaluation at the 95% confidence level: APS was closest to the theoretical 95% coverage rate, while LAC had the smallest sets. Comparison of calibration strategies showed that a larger calibration set improved the stability of coverage for rare categories. Confidence threshold scan analysis provided a trade-off curve, offering a basis for selecting deployment thresholds. Specific data: LAC actual coverage rate was 89.96% with an average set size of 2.009; APS was 94.39% with 2.589; Top-K was 93.54% with 3.030.

Grad-CAM was integrated to generate attention heatmaps, verifying that the model focuses on pathology-related regions and showing differences in attention across different candidate categories. Analysis of difficult case characteristics: similar tissue morphology, image quality issues, and boundary cases. These cases can be marked for expert review, forming a human-machine collaboration safety net.

This study provides provable safety guarantees for medical AI, facilitating clinical approval. Future directions: multi-scale Conformal Prediction (whole-slide images), temporal Conformal Prediction (longitudinal cases), cross-domain generalization (different hospital equipment), and integration with active learning (guiding annotation resource allocation). Open-source projects promote medical AI from "usable" to "trustworthy".