Reproduce “CarveMix: A Simple Data Augmentation Method for Brain Lesion Segmentation” on OpenResearch

About this paper

Xinru Zhang, Chenghao Liu, Ni Ou, Xiangzhu Zeng, Xiaoliang Xiong, Yizhou Yu, Zhiwen Liu, Chuyang Ye

CarveMix introduces a simple, lesion-aware data augmentation method for 3D brain lesion segmentation, generating synthetic samples by combining lesion-centric regions from existing images with consistent annotations. This approach consistently improves segmentation accuracy, especially in low-data regimes, outperforming traditional and generic "mix"-based augmentation techniques on stroke lesion datasets.

Key insights

• Data augmentation methods for medical image segmentation greatly benefit from being "lesion-aware" and ensuring consistent label generation, unlike generic approaches designed for classification.
• Directly applying methods like CutMix, which use random regions or linear label mixing, can degrade segmentation performance due to inconsistent or ambiguous labels, especially in data-limited scenarios.
• A simple, task-specific augmentation strategy can significantly enhance CNN robustness and generalization for precise 3D medical image segmentation, balancing diversity with implementation ease.

Summary by alphaXiv.