Optimization with soft Dice can lead to a volumetric bias

TL;DR

This paper reveals that optimizing with soft Dice loss in medical image segmentation can cause a volumetric bias, especially in uncertain cases, despite improving Dice scores.

Contribution

It provides both theoretical analysis and empirical evidence showing how soft Dice optimization can introduce volumetric bias in segmentation tasks.

Findings

Soft Dice optimization can lead to volumetric bias.

Bias is more pronounced in high-uncertainty tasks.

The issue affects clinical applicability of segmentation methods.

Abstract

Segmentation is a fundamental task in medical image analysis. The clinical interest is often to measure the volume of a structure. To evaluate and compare segmentation methods, the similarity between a segmentation and a predefined ground truth is measured using metrics such as the Dice score. Recent segmentation methods based on convolutional neural networks use a differentiable surrogate of the Dice score, such as soft Dice, explicitly as the loss function during the learning phase. Even though this approach leads to improved Dice scores, we find that, both theoretically and empirically on four medical tasks, it can introduce a volumetric bias for tasks with high inherent uncertainty. As such, this may limit the method's clinical applicability.