Enhancing Medical Image Classification via Uncertainty Estimation

Medical image classification plays a critical role in a variety of healthcare applications. In recent years, Deep Neural Networks (DNNs) have achieved remarkable success in this domain. However, their reliance on softmax outputs limits their ability to capture uncertainty in predictions—a crucial factor in medical decision-making. Unlike traditional uncertainty estimation methods, conformal prediction (CP) offers a model-agnostic, distribution-free framework for generating statistically sound uncertainty sets. Despite its theoretical promise, existing exact full CP methods require retraining the DNN for every test sample across all possible labels, which is computationally intensive. Moreover, current approaches often overlook the underlying causes of prediction uncertainty, making it challenging for doctors to interpret the results. To overcome these issues, we propose a novel, efficient approximate full CP approach that tracks gradient updates from training samples. Subsequently, we introduce an interpretation method that leverages these updates to identify the top-k most influential training samples contributing to model uncertainty. We validate the effectiveness of our approach through extensive experiments on real-world medical imaging datasets.

Committee: Dr. Mengdi Huai (major professor), Dr. Tichakorn Wongpiromsarn, and Dr. Liyi Li