🤖 AI Summary
To address the accuracy and efficiency bottlenecks in label fusion for anatomical structure segmentation in large-scale MRI databases, this paper proposes an optimized patch-based label fusion framework. Methodologically, it introduces three key innovations: (1) an enhanced PatchMatch algorithm incorporating adaptive multi-scale neighborhood propagation to improve structural consistency across image patches; (2) a weighted multi-feature similarity metric—integrating intensity, texture, and gradient features—augmented with non-rigid registration priors to enhance cross-modality robustness; and (3) a multi-scale pyramid matching mechanism that jointly captures local details and global context. Evaluated on brain MRI segmentation, the proposed method achieves a 3.2-percentage-point improvement in Dice coefficient over standard PatchMatch while accelerating inference by 35%. It demonstrates significantly improved label fusion performance and generalizability across multi-center, multi-sequence MRI datasets.
