WSI-BayesUNet: Uncertainty-Aware Deep Learning for Histopathological Image Segmentation with Active Learning

Yijun Cui, Nadieh Khalili

Keywords: Histopathology, Segmentation, Deep Learning, Uncertainty Quantification

Abstract: Histopathological image segmentation is a core task in digital pathology, supporting applications such as cancer detection and subtype classification. Manual annotation is time-consuming and subjective, making automation essential for improving efficiency and consistency in diagnostic workflows. Although deep learning models have significantly automated this process, they still make silent mistakes. Quantifying the uncertainty of the model and using the uncertainty for further improvement is not fully addressed. The most common way to quantify uncertainty is through ensemble methods, which provide empirical uncertainty estimation but face limitations, including high computational costs and theoretical instability. To address these, we propose a Bayesian U-Net framework that employs variational inference for principled probabilistic uncertainty estimation. Leveraging active learning, our Bayesian U-Net iteratively improves segmentation performance by prioritizing the most uncertain samples. Experiments on the TIGER and CAMELYON17 datasets show that Bayesian U-Net outperforms ensemble methods computationally, offering better uncertainty quantification, uncertainty-guided performance gains, and faster convergence. Notably, uncertainty-based sampling consistently surpasses random sampling, significantly reducing annotation effort while maintaining or improving segmentation accuracy.

Submission Number: 41