Multi-atlas segmentation of knee cartilage via Semi-supervised Regional Label Propagation

Multi-atlas based segmentation techniques have been proven to be effective in multiple automatic segmentation applications. However, mostly they rely on a non-deformable registration model followed by a voxel-wise classification process that incurs a large computational cost in terms of memory requirements and execution time. In this paper, a novel two-stage multi-atlas method is proposed, which combines constructively several concepts, including Semi-Supervised Learning (SSL), sparse graph constructions, voxel’s linear reconstructions via graph weights, and suitable sampling schemes for collecting data from target image and the atlas library. Representative global data sampled from target image are first classified according to SSL, using a newly proposed label propagation scheme. Next, out-of-sample data of yet unlabeled target voxels are iteratively generated through an iterative sampling based on mesh tetrahedralization. A thorough experimental investigation is conducted on 45 subjects provided by the publicly accessible Osteoarthritis Initiative (OAI) repository. Comparative analysis demonstrates that the proposed approach outperforms the existing state-of-the-art patch-based methods, across all evaluation metrics, exhibiting enhanced segmentation performance and reduced computational loads, respectively.