Hierarchical Decision Framework with a Priori Shape Models for Knee Joint Cartilage Segmentation—MICCAI Grand Challenge

Abstract

A hierarchical decision framework for segmenting multiple cartilage–bone surfaces belonging tomutually interacting bones (objects) of the knee joint is reported. The underlying segmentation approach is based on optimal graph-based surface detection with embedded pat- tern recognition functionality. A novel tibia/femur/patella detection ap- proach for initial bone segmentation uses 3D Haar wavelet features and AdaBoost classifier. Accurate bone and cartilage surfaces are obtained using LOGISMOS1 segmentation. On 40 knee MR images from the 2010 MICCAI Grand Challenge, our method achieved volume overlap error (VOE) of 31.2%±9.1% and 33.5%±11.0%, as well as volumetric differ- ence (VD) of -7.3%±19.3% and 3.6%±19.7% for the femoral and tibial cartilage regions, respectively. The overall femoral and tibial cartilage segmentation quality scores were 65 and 61, respectively.