Are robust loss functions still relevant for medical image segmentation with noisy labels?

Vikram Venkatraghavan, Richard A.P. Takx, Niek E. van der Aa, Timo R. de Haan, Nils Planken, Manon Benders, Ivana Isgum

Keywords: Medical Image Segmentation, Noisy labels, Robust loss

Abstract: Creating reference annotations for semantic segmentation in medical images is a labor-intensive process, where experts often disagree on the precise location of the boundary between semantic structures. Hence, noisy labels in medical image segmentation tasks are pervasive. A popular approach to tackling noisy labels is to use robust loss functions that are resilient to noise. Meanwhile, the introduction of nnU-Net has highlighted the critical role that a well-configured combination of data augmentation, model architecture, and inference pipelines play in medical image segmentation. However, the potential of such strong baselines to mitigate the impact of label noise and the additional advantage of using a robust loss function has not been thoroughly explored. Most studies proposing robust loss functions, often with tunable hyper-parameters, have shown their efficacy either with a custom U-Net architecture, or without hyper-parameter optimization. By thorough benchmarking using a standardized nnU-Net framework, along with independent hyper-parameter optimization, we found that stochastic co-teaching based small-loss sample selection and active-passive loss comprising normalized generalized cross-entropy, reverse cross entropy, and Dice loss are useful for robust segmentation in applications with high label noise. For segmentation tasks characterized by minimal label noise, none of the robust loss functions demonstrate performance improvements over a well-established baseline model. Our results highlight the need for benchmarking with strong baseline models, even when proposing a new robust loss function that is architecture and framework independent.

Primary Subject Area: Segmentation

Secondary Subject Area: Learning with Noisy Labels and Limited Data

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 63