Source-linked AI summary
Cut to the Mix: Simple Data Augmentation Outperforms Elaborate Ones in Limited Organ Segmentation Datasets
Chang Liu, Fuxin Fan, Annette Schwarz, Andreas Maier
TL;DR
Limited manually annotated data constrain supervised multi-organ segmentation, especially for novel imaging modalities. The paper evaluates four inter-image and object-level augmentation strategies on two limited organ-segmentation datasets and finds that CutMix, CarveMix, and AnatoMix improve segmentation, with CutMix emerging as the simplest robust strategy despite sometimes producing anatomically implausible images.
Problem
Supervised multi-organ segmentation requires large manually annotated datasets, while such data remain difficult to obtain for novel imaging modalities.
Method
The study re-implements CutMix, ObjectAug, CarveMix, and AnatoMix for multi-organ segmentation and evaluates them on two limited organ-segmentation datasets.
Results
CutMix, CarveMix, and AnatoMix improve macro averaged dsc on AMOS by 4.9, 2.0, and 1.9, respectively, without TDA.
Takeaways & Limitations
CutMix is identified as the best data-augmentation strategy for limited multi-organ segmentation datasets based on performance and implementation complexity.
Takeaways & Limitations
The study challenges the assumption that augmentation should always preserve in-distribution human anatomy, including organ number and relative location.
Abstract
from arXiv · showhide
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.
1 Introduction
Multi-organ segmentation supports clinical care, but supervised deep-learning models require large manually annotated datasets that remain difficult to obtain for novel imaging modalities. Data augmentation, including inter-image and object-level strategies, is investigated as a way to improve limited-data segmentation.
- Multi-organ segmentation is routinely used in clinical care, including radiation-therapy treatment planning.
- Supervised deep-learning segmentation models rely on large-scale images with manual organ annotations.
- Limited datasets remain difficult to assemble for novel imaging modalities such as dual-energy computed tomography.
- Traditional augmentation changes spatial or intensity properties, while inter-image and object-level methods create data by combining or manipulating images and segmented objects.
- The study re-implements CutMix, ObjectAug, CarveMix, and AnatoMix for multi-organ segmentation with limited data.
2 Method
The method re-implements four inter-image or object-level augmentation strategies for multi-organ segmentation and evaluates them with nnUNetv2 on limited organ-segmentation datasets. The strategies construct augmented images by fusing images or organ objects, with AnatoMix additionally matching organs by size and shifting them within background images.
- Overview: Four established strategies—CutMix, ObjectAug, CarveMix and AnatoMix—are re-implemented for multi-organ segmentation and evaluated on two limited organ-segmentation tasks.nnUNetv2 is used to train the segmentation networks.
- Shared fusion operation: The basic fusion operation overlays a selected image or mask onto a background image or mask using a binary mask.The operation is expressed as Ib ⊗(Is, m) = Is · m + Ib · (1 −m).
- CutMix: CutMix creates a random bounding-box mask and fuses a background image with a randomly selected source image.The bounding-box size follows an image-size-proportional ratio sampled from β(0.5, 0.5).
- ObjectAug: ObjectAug removes organ pixels, uses background inpainting, and recombines organs after random scaling, shifting, and rotation.The implemented transformations are 10% scaling, 5-voxel shifts in all dimensions, and 15° rotation.
- CarveMix and AnatoMix: CarveMix applies augmentation to individual organs, whereas AnatoMix matches organs with similar sizes and replaces background organs after shifting them by an optimal offset.AnatoMix addresses the organ-location and organ-size changes that CarveMix does not maintain.
- Experimental setting: The evaluation uses AMOS and a private DECT dataset, varies augmentation to 200, 500, or 1,000 images, and tests compatibility with traditional augmentation.AMOS contains 300 abdominal CT volumes with 16 annotated organs and anatomical structures; the datasets differ in scanning-region diversity.
3 Results and Discussion
The experiments compare four augmentation strategies across AMOS and DECT, revealing that performance gains and anatomical plausibility vary by method and dataset. CutMix is especially effective despite sometimes producing anatomically abnormal outputs.
- CutMix and CarveMix can disturb human anatomy when input scanning regions differ greatly, producing outputs with duplicated organs or misplaced body regions.AnatoMix produced correct organ locations and similar organ sizes in the described examples.
- CutMix, CarveMix, and AnatoMix improved AMOS macro averaged dsc by 4.9, 2.0, and 1.9 without TDA.
- CutMix improved AMOS macro averaged dsc by 7.0 when combined with optimized nnUNet TDA.The reported gains from CutMix and TDA were additive compared with no TDA.
- On DECT, CutMix, CarveMix, and AnatoMix improved macro averaged dsc by 3.1, 2.8, and 2.4 without TDA, while micro averaged dsc did not improve.With TDA, their macro gains were 1.6, 1.6, and 1.8, respectively.
- The experiments challenge the presumption that augmentation outputs must preserve normal organ counts and relative locations for DL segmentation.The study reports that anatomically abnormal images can still be useful for the networks.
4 Conclusion
The paper concludes that several inter-image augmentation strategies enhance limited multi-organ segmentation datasets. Among them, CutMix offers the strongest practical balance of performance and implementation complexity.
- CutMix, CarveMix, and AnatoMix effectively enhanced limited segmentation datasets.
- Combining these strategies with TDA yielded joint improvements in segmentation performance.
- CutMix was identified as the best DA strategy for limited multi-organ segmentation datasets based on metric results and implementation complexity.