CAM-CycleGAN: A Weakly Supervised Segmentation Method for Medical Images Based on Cycle Consistensy Generative Adversarial Networks

Initial segmentation of key parts of a medical image is a critical step for computers to help doctors analyze medical images. Fully supervised deep learning segmentation models rely on large-scale pixel-level labeling; however, the number of medical images with pixel-level labels is small due to the high cost of annotation. Therefore, this paper proposes a weakly supervised medical image segmentation method that requires only image-level label. The method uses high-confidence regions of class activation mapping (CAM) as supervised information and trains a generative adversarial network with added cycle consistency loss to establish the mapping relationship between the original image and the segmentation mask. We name this method CAM-CycleGAN. With this approach, the model can be trained to acquire subregions that are highly correlated with the category and can preserve the fine-grained features of the original image. Finally, we demonstrate the effectiveness of the proposed method in the publicly available colon gland dataset GlaS.