CarveMix: A Simple Data Augmentation Method for Brain Lesion Segmentation

Brain lesion segmentation provides a valuable tool for clinical diagnosis, and convolutional neural networks (CNNs) have achieved unprecedented success in the task. Data augmentation is widely used strategy that improves training of CNNs, design method brain still an open problem. In this work, we propose simple data approach, dubbed as CarveMix, CNN-based segmentation. Like other "mix"-based methods, such Mixup CutMix, CarveMix stochastically combines two existing labeled images to generate new samples. Yet, unlike these strategies based on image combination, lesion-aware, where combination performed with attention lesions proper annotation created generated image. Specifically, from one carve region interest (ROI) according location geometry, size ROI sampled probability distribution. The carved then replaces corresponding voxels second image, replaced accordingly well. way, network information preserved. To evaluate proposed method, experiments were datasets. results show our accuracy compared approaches.