Enhancing source-free domain adaptation in Medical Image Segmentation via regulated model self-training

Abstract

Source-free domain adaptation (SFDA) has drawn increasing attention lately in the medical field. It aims to adapt a model well trained on source domain to target domains without accessing source domain data nor requiring target domain labels, to enable privacy-protecting and annotation-efficient domain adaptation. Most SFDA approaches initialize the target model with source model weights, and guide model self-training with the pseudo-labels generated from the source model. However, when source and target domains have huge discrepancies (e.g., different modalities), the obtained pseudo-labels would be of poor quality. Different from prior works that overcome it by refining pseudo-labels to better quality, in this work, we try to explore it from the perspective of knowledge transfer. We recycle the beneficial domain-invariant prior knowledge in the source model, and refresh its domain-specific knowledge from source-specific to target-specific, to help the model satisfyingly tackle target domains even when facing severe domain shifts. To achieve it, we proposed a regulated model self-training framework. For high-transferable domain-invariant parameters, we constrain their update magnitude from large changes, to secure the domain-shared priors from going stray and let it continuously facilitate target domain adaptation. For the low-transferable domain-specific parameters, we actively update them to let the domain-specific embedding become target-specific. Regulating them together, the model would develop better capability for target data even under severe domain shifts. Importantly, the proposed approach could seamlessly collaborate with existing pseudo-label refinement approaches to bring more performance gains. We have extensively validated our framework under significant domain shifts in 3D cross-modality cardiac segmentation, and under minor domain shifts in 2D cross-vendor fundus segmentation, respectively. Our approach consistently outperformed the competing methods and achieved superior performance.