Uncertainty Estimation for Out-of-Distribution Detection of Whole Slide Images

Abstract

Whole slide images (WSIs), which are high-resolution, digitized representations of tissue samples, pose significant computational challenges due to their gigapixel scale and multi-resolution format. Recent research has demonstrated that incorporating neighborhood information using graph neural networks (GNNs) can significantly improve cancer classification accuracy in WSIs. However, these models often struggle with out-of-distribution (OOD) data, where discrepancies between training and testing distributions arise. Accurately estimating the prediction uncertainty in such cases is critical, particularly when the model predictions are used in the medical domain. This work presents a benchmark for evaluating uncertainty estimation methods across multiple datasets, comparing their performance on both in-distribution (ID) and OOD data at the whole-slide level. Additionally, this evaluation illustrates that graph multihead approach outperforms traditional uncertainty estimation methods in distinguishing between ID and OOD data.