Abstract
This study addresses the challenge of training deep neural networks for medical image analysis, where data annotation demands expertise and incurs high costs, and privacy concerns hinder neural network training. We introduce an innovative source-free domain adaptation method for medical image segmentation, leveraging generative adversarial networks. This technique employs a discriminator to identify and assimilate target-domain features, transferring these parameters to the segmentation network's downsampling section. Experimentation with public medical image datasets shows our method effectively reduces segmentation model loss during domain transfer without source data, leading to progressive improvements.