Navigating Uncertainty in Medical Image Segmentation

TL;DR

This paper evaluates methods for handling uncertainty in medical image segmentation, providing guidelines for model selection and development that incorporate aleatoric and epistemic uncertainties to improve practical application.

Contribution

It offers new guidelines for selecting and developing uncertain segmentation models in medical imaging, based on case studies and analysis of model limitations.

Findings

Simple deterministic models can suffice for minimal annotation variation.

GED has limitations in model selection for lung lesion segmentation.

Guidelines improve the development and evaluation of uncertain segmentation methods.

Abstract

We address the selection and evaluation of uncertain segmentation methods in medical imaging and present two case studies: prostate segmentation, illustrating that for minimal annotator variation simple deterministic models can suffice, and lung lesion segmentation, highlighting the limitations of the Generalized Energy Distance (GED) in model selection. Our findings lead to guidelines for accurately choosing and developing uncertain segmentation models, that integrate aleatoric and epistemic components. These guidelines are designed to aid researchers and practitioners in better developing, selecting, and evaluating uncertain segmentation methods, thereby facilitating enhanced adoption and effective application of segmentation uncertainty in practice.