Domain-adapted polyp image semantic segmentation utilizing a generative adversarial network and U-Net framework
Shizhao Ma, Yunhai Gao, Shuquan Feng, Lin Li, Mengyuan Ma · 2024
Medical image segmentation is a crucial area in medical image processing, aiming to help doctors accurately identify and segment different structures and tissues in images, thereby supporting disease diagnosis, treatment planning, and monitoring. A major challenge in this field is the variability among medical institutions, equipment, and datasets, emphasizing the need for domain adaptation. Domain adaptation for medical image semantic segmentation aims to address this issue, allowing models to achieve accurate segmentation results on medical images with different data distributions. In this context, our introduce a method for domain adaptation in medical image semantic segmentation that combines generative adversarial networks (GANs) with the U-Net architecture. The GAN discriminator is integrated into the U-Net encoder for adversarial domain adaptation, aligning features between source and target domains and improving the model's adaptation performance. Notably, this method introduces key steps of feature channel grouping and feature channel alignment to prevent the features of source domain images and target domain images from adapting at high frequencies but not at low frequencies, or adapting at low frequencies but not at high frequencies, improving the model's stability and accuracy. Through extensive experimental validation, this method has shown outstanding performance in domain adaptation polyp semantic segmentation. The findings highlight the effectiveness and importance of this approach in solving domain adaptation issues in medical image segmentation.