MedAugment: Universal Automatic Data Augmentation Plug-in for Medical Image Analysis

Data Augmentation (DA) has been widely implemented in the field of computer vision to alleviate the data shortage, whereas the DA in Medical Image Analysis (MIA) faces multiple challenges. The prevalent DA approaches in MIA encompass both general DA and generative adversarial network-based DA. However, the former approach is predominantly experience-driven, and the latter approach can be hindered by unquantifiable synthesis quality and mode collapse. Here, we develop a plug-and-use DA method, named MedAugment, to leverage the automatic DA to benefit the MIA field. To address the differences between natural and medical images, we divide the augmentation space into pixel augmentation space and spatial augmentation space. Moreover, a novel operation sampling strategy is proposed when sampling DA operations from the spaces. To demonstrate the performance and universality of MedAugment, we conduct extensive experiments on four classification datasets and three segmentation datasets. The results show that MedAugment outperforms existing DA methods. This work suggests that the plug-and-use MedAugment may benefit the MIA community. Code is available at https://github.com/NUS-Tim/MedAugment.