Abstract
In unsupervised transfer learning for medical image segmentation, where existing algorithms face the challenge of error propagation due to inaccessible source domain data. In response to this scenario, source-free domain transfer algorithm with reduced style sensitivity (SFDT-RSS) is designed. SFDT-RSS initially pre-trains the source domain model by using the generalization strategy and subsequently adapts the pre-trained model to target domain without accessing source data. Then, SFDT-RSS conducts interpatch style transfer (ISS) strategy, based on self-training with Transformer architecture, to minimize the pre-trained model's style sensitivity, enhancing its generalization capability and reducing reliance on a single image style. Simultaneously, the global perception ability of the Transformer architecture enhances semantic representation to improve style generalization effectiveness. In the domain transfer phase, the proposed algorithm utilizes a model-agnostic adaptive confidence regulation (ACR) loss to adjust the source model. Experimental results on five publicly available datasets for unsupervised cross-domain organ segmentation demonstrate that compared to existing algorithms, SFDT-RSS achieves segmentation accuracy improvements of 2.83%, 2.64%, 3.21%, 3.01%, and 3.32% respectively.