Mutation-based Data Augmentation for Medical Imaging Segmentation

In medical imaging segmentation, development of effective deep learning models and clinical deployment is significantly hampered by data scarcity and the high cost of labeling. This research addresses this challenge by proposing a novel data augmentation strategy. It minimizes the dependency on large datasets, and accelerates both development and deployment in clinical settings. Our goal is to maximize data diversity through data augmentation. It allows us to reduce the amount of medical imaging data required to train high-performance segmentation models. Specifically, we introduce a Mutation-based data generation method, which uses spatial transformations and deformations to mutate between object and background structures. These transformations enhance model learning by capturing variations in object location, shape, and intensity. Using only 10% of the original training data (39 samples) from the Medical Segmentation Decathlon Task01-BrainTumour dataset (MSD[1]), our method enables training of a simple, lightweight U-Net. It achieves a mean Dice Similarity Coefficient (DSC) of 0.612 on the test set, a result comparable to the state-of-the-art model (SwinUNETR[2]) performance of 0.644 mean DSC. This demonstrates that even with very limited data, our approach may yield high performance. Additionally, for datasets in the MSD collection with fewer than 20 samples, our method significantly enhances model performance. For example, on the MSD Task02-Heart dataset, this approach boosts the DSC by over 0.5, while on the Task04-Hippocampus dataset, it achieves an improvement of 0.3 DSC. Our findings challenge the conventional notion of data quantity in clinical model training, demonstrating that powerful augmentation can enable faster, cost-effective deployment of medical imaging segmentation models. [1]. Antonelli, Michela, et al. "The medical segmentation decathlon." Nature communications 13.1 (2022): 4128. [2]. Tang, Yucheng, et al. "Self-supervised pre-training of swin transformers for 3D medical image analysis." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.