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

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.

💡 Research Summary

This paper addresses the practical problem of improving multi‑organ segmentation when only a small number of annotated CT volumes are available. While deep learning models such as nnU‑Net achieve impressive results on large, well‑annotated datasets, many emerging imaging modalities (e.g., dual‑energy CT) suffer from severe data scarcity, making robust training difficult. Data augmentation (DA) is a standard regularisation technique, but most prior work focuses on traditional intra‑image transformations (rotation, scaling, intensity shifts). Recent computer‑vision research has introduced inter‑image and object‑level augmentations (e.g., Mixup, CutMix, CarveMix, ObjectAug, AnatoMix) that combine information from different samples to create novel training examples. The authors re‑implement four such strategies—CutMix, CarveMix, ObjectAug, and AnatoMix—specifically for multi‑organ segmentation and evaluate them on two limited‑size abdominal CT datasets: the public AMOS dataset (20 training, 100 test images) and a private DECT dataset (20 training, 22 test images).

Methodology
All four augmentations are applied to the training set at three multiplicative levels (×10, ×25, ×50), producing 200, 500, and 1 000 synthetic volumes respectively. The original cases are excluded from the augmented pool to ensure that the network only sees novel combinations. nnU‑Net v2 is trained on each augmented set with identical hyper‑parameters; experiments are performed both with and without the built‑in traditional DA (TDA) of nnU‑Net. Performance is measured using two Dice‑based metrics: micro‑averaged Dice (overall pixel‑wise accuracy) and macro‑averaged Dice (average across organs, more sensitive to small structures).

Key Findings

1. CutMix dominates – Even without any TDA, CutMix raises macro‑averaged Dice by 4.9 % points and micro‑averaged Dice by 2.6 % points on AMOS. When combined with TDA, the gains increase further (≈+3.0 % micro, +4.8 % macro). Similar trends appear on DECT, where CutMix improves macro Dice by 3.1 % points despite already high baseline performance.

2. CarveMix and AnatoMix provide modest gains – Both improve macro Dice by roughly 2 % points but lag behind CutMix in micro Dice. Their performance is more sensitive to the augmentation multiplier; increasing the number of synthetic samples does not consistently raise accuracy.

3. ObjectAug underperforms – The object‑level pipeline, which requires background in‑painting and per‑organ geometric transforms, leads to severe degradation (macro Dice drops to ~7 % points). The added complexity does not translate into useful regularisation for organ segmentation.

4. Computational efficiency matters – CutMix generates a mixed pair in ~0.3 seconds, whereas CarveMix, AnatoMix, and ObjectAug need 15.7 s, 20.9 s, and 40.4 s respectively. For large‑scale training or clinical deployment, the speed advantage of CutMix is decisive.

5. Effect of augmentation multiplier – On AMOS, higher multipliers (×25, ×50) continue to improve macro Dice for CutMix, but on DECT the gains plateau or even reverse, likely because the baseline Dice is already near the ceiling (≈97 %). This suggests diminishing returns when the model already generalises well.

6. Anatomical realism is not essential – Although CutMix can produce anatomically implausible images (e.g., four kidneys, two livers), the network still learns robust features. This challenges the common assumption that synthetic data must strictly respect organ count and spatial relationships.

Implications
The study demonstrates that a simple inter‑image augmentation like CutMix, despite creating “wrong” anatomy, is the most effective and efficient strategy for limited multi‑organ segmentation datasets. Complex object‑level augmentations that aim to preserve anatomical consistency incur substantial implementation overhead without delivering proportional performance benefits. Moreover, the additive effect of CutMix and traditional DA suggests that combining coarse global transformations with fine‑grained inter‑image mixing yields the best regularisation.

Conclusion
For researchers and clinicians working with scarce annotated CT data, integrating CutMix into the training pipeline offers a low‑cost, high‑impact improvement. It outperforms more elaborate methods both in Dice score gains and runtime, and it synergises with existing TDA. Future work could explore CutMix across other modalities (MRI, PET) and investigate how such mixed samples affect model uncertainty estimation and calibration.