Adaptive transfer learning and multi-fidelity fusion: Complementary strategies for liver MRI semantic segmentation

Domain shift and scarcity in target domain annotations constrain the generalizability of liver magnetic resonance imaging (MRI) segmentation models. To overcome these challenges, we propose two strategies: adaptive unfreezing transfer learning (AU-TL) for cross-domain adaptation and residual-connected multi-fidelity fusion (R-MFF) for integrating low-fidelity public data and high-fidelity (HF) private data. AU-TL and R-MFF reduced the annotation requirements by 78.6% and 71.4%, respectively, while maintaining their competitive segmentation performance. This finding shows that RMFF is particularly advantageous when target domain annotations are extremely limited, whereas AU-TL exhibits a higher performance as more annotated data become available. This finding demonstrates the complementary advantages of these models under different levels of annotation availability.