Label noise in segmentation networks : mitigation must deal with bias

TL;DR

This paper investigates the impact of biased and unbiased label noise on medical image segmentation, emphasizing the need for mitigation strategies that address bias to improve neural network predictions.

Contribution

It highlights the importance of addressing biased label errors in medical segmentation and evaluates the robustness of existing methods to different error types.

Findings

Supervised and semi-supervised methods are robust to unbiased errors.

Methods are sensitive to biased errors, affecting segmentation quality.

Bias in labels significantly impacts model performance.

Abstract

Imperfect labels limit the quality of predictions learned by deep neural networks. This is particularly relevant in medical image segmentation, where reference annotations are difficult to collect and vary significantly even across expert annotators. Prior work on mitigating label noise focused on simple models of mostly uniform noise. In this work, we explore biased and unbiased errors artificially introduced to brain tumour annotations on MRI data. We found that supervised and semi-supervised segmentation methods are robust or fairly robust to unbiased errors but sensitive to biased errors. It is therefore important to identify the sorts of errors expected in medical image labels and especially mitigate the biased errors.