Deep Learning has been widely applied to medical image segmentation, aiming to make structures clearer in images to help physicians identify unusual patterns and anomalies. Segmentation models face challenges in collecting a large amount of data for training, due to privacy concerns and pathological representation. Data Augmentation (DA) is an alternative to mitigate this challenge, expanding the dataset by applying transformations to the original set or creating new samples using generative methods. Despite the extensive use of DA techniques, there is still limited understanding of their relative effectiveness for medical image segmentation tasks. This work presents a method for evaluating the impact of DA methods, analyzing traditional augmentation techniques, and diffusion models for generating synthetic data in medical image segmentation models.
Uchida et al. (Mon,) studied this question.