Knee Cartilage Segmentation from MR Images

Quantitative assessment of articular cartilage in 3D knee MR images is central to osteoarthritis research and clinical monitoring, but cartilage is thin, low-contrast, and topologically complex — making reliable segmentation a long-standing challenge. Our work in this area developed a progression of 3D segmentation frameworks built around graph-cut optimization, localized probabilistic modeling, coarse-to-fine strategies, and unified automatic/interactive pipelines, largely evaluated on MR data from the Osteoarthritis Initiative (OAI).

Research directions

  • 3D graph-cut segmentation. Energy-minimization formulations over voxel graphs that balance image data fidelity with spatial smoothness, applied to articular cartilage and extended with iterative local branch-and-mincut for fully automatic bone compartment segmentation.
  • Localized shape and appearance modeling. Joint optimization of local shape and appearance probabilities so that the model adapts to patient-specific anatomy rather than relying on global priors alone.
  • Hierarchical MRF with localized classifiers. Globally consistent localized classifiers integrated into a multi-level MRF, combining spatially specialized training with global topological consistency.
  • Coarse-to-fine bone and cartilage segmentation. Approximate localization followed by iterative boundary refinement, producing accurate 3D bone compartment segmentation in knee MR.
  • Unified automatic + interactive segmentation. Structured patch models that seamlessly bridge automatic segmentation and interactive editing within a single framework.