Cut to the Mix: Simple Data Augmentation Outperforms Elaborate Ones in Limited Organ Segmentation Datasets

TL;DR

This work tackles the data scarcity hurdle in automated multi-organ segmentation by re-implementing and evaluating four inter-image/object-level data augmentation strategies—CutMix, ObjectAug, CarveMix, and AnatoMix—on two limited datasets (AMOS and DECT) using nnUNetv2. The study shows that CutMix delivers the strongest and most robust gains in dice scores, with macro improvements up to $4.9$ points without traditional data augmentation and substantial additive gains when combined with TDAs; CarveMix and AnatoMix also improve performance, while ObjectAug can be slower and less reliable. Across datasets, inter-image DA methods offer meaningful improvements in limited-data regimes, though the benefits can depend on anatomical plausibility and dataset characteristics; CutMix emerges as the fastest and simplest option with competitive accuracy. The authors provide open-source implementations for future benchmarking and emphasize that even intuitively “wrong” augmentations can enhance learning in DL-based segmentation under data scarcity, highlighting practical implications for clinical deployment where annotated data are scarce.

Abstract

Multi-organ segmentation is a widely applied clinical routine and automated organ segmentation tools dramatically improve the pipeline of the radiologists. Recently, deep learning (DL) based segmentation models have shown the capacity to accomplish such a task. However, the training of the segmentation networks requires large amount of data with manual annotations, which is a major concern due to the data scarcity from clinic. Working with limited data is still common for researches on novel imaging modalities. To enhance the effectiveness of DL models trained with limited data, data augmentation (DA) is a crucial regularization technique. Traditional DA (TDA) strategies focus on basic intra-image operations, i.e. generating images with different orientations and intensity distributions. In contrast, the interimage and object-level DA operations are able to create new images from separate individuals. However, such DA strategies are not well explored on the task of multi-organ segmentation. In this paper, we investigated four possible inter-image DA strategies: CutMix, CarveMix, ObjectAug and AnatoMix, on two organ segmentation datasets. The result shows that CutMix, CarveMix and AnatoMix can improve the average dice score by 4.9, 2.0 and 1.9, compared with the state-of-the-art nnUNet without DA strategies. These results can be further improved by adding TDA strategies. It is revealed in our experiments that Cut-Mix is a robust but simple DA strategy to drive up the segmentation performance for multi-organ segmentation, even when CutMix produces intuitively 'wrong' images. Our implementation is publicly available for future benchmarks.

Figures (2)

• Figure 1: Illustration of the concept of CutMix, ObjectAug, CarveMix and AnatoMix. All DA strategies are originally proposed for either image classification or tumor segmentation task. They are re-implemented for the multi-organ segmentation task and further evaluated in this work.
• Figure 2: Some example slices of the output volumes using the different DA strategies. The first row shows the outputs from the DECT dataset and the lower rows show the outputs from AMOS dataset. The four DA strategies lead to different compliance with the original human anatomy. Red arrows and dashed lines indicate the abnormal regions.