Retinal Vessel Segmentation Using Multi-scale Generative Adversarial Network with Class Activation Mapping

Retinal vessel segmentation plays a significant role in the accurate diagnosis of retinal diseases. However, existing methods commonly omit micro-vessels in retinal images and generate some false-positive vessels. To alleviate this issue, we propose a multi-scale generative adversarial network with class activation mapping to achieve efficient segmentation. For the problem of small amount of data, we introduce a novel data augmentation method, which can generate multiple samples by cutting pixels from other samples. This method increases the diversity of samples and improve the robustness of the model. We compare our method with previous models with several metrics, and experiments show the superiority and effectiveness of our model.