FA-SedLoRA: scene-driven fine-tuning and style disentanglement for domain-generalized cardiac MRI segmentation

The advent of deep learning has significantly advanced the state of the art in cardiac magnetic resonance (CMR) image segmentation. However, most models remain task-specific, which hinders the development and validation of approaches generalizable across different clinical centers, imaging protocols, or scanner vendors. Vision foundation models (VFMs), pre-trained on large-scale natural image datasets, offer powerful and transferable representations under the "pre-training and fine-tuning" paradigm. Nevertheless, adapting them to CMR segmentation faces two major challenges: (1) the substantial domain gap between natural and medical images limits transferability; and (2) extracting domain-agnostic features from diverse domain styles represents a key bottleneck for domain generalization with VFMs. To address these issues, we propose a scenario-activated fine-tuning initialization strategy that adapts VFMs to MRI characteristics and employs singular value decomposition to extract principal components for parameter-efficient tuning. This approach enables robust domain-generalized CMR segmentation. Additionally, we apply Haar wavelet transforms to disentangle style information from domain-invariant content. The former helps stabilize scene content, while the latter captures scene style and mitigates its impact on domain-generalized semantic segmentation. Experiments under various CMR domain generalization segmentation settings demonstrate the state-of-the-art performance of our FA-SedLoRA framework and its versatility across different VFMs.