Efficient Cartilage Segmentation in High-Field Knee MR Images with Voxel-Classification-Driven Region-Growing Algorithm

Abstract

This paper presents a voxel-classification-driven region-growing algorithm for automatically segmenting the whole femoral, tibial, and patellar cartilage tissues in high-field magnetic resonance (MR) images of the knee joint by taking into consideration systems with limited resources in particular. An abundance of background voxels and high dimensionality of the voxel samples were alleviated via various subsampling techniques and selecting fewer significant features, respectively. Experiments were conducted on 33 MR images obtained from the Osteoarthritis Initiative (OAI) database in three-dimensional (3-D) double echo in the steady state standard (DESS). After processing 10 MR images for training, four training models were generated by Gaussian, uniform, cartilage vicinity correlated (CVC) sparse, and CVC dense subsampling techniques. Then, their effect on the final segmentation accuracies of the cartilaginous compartments of interest on the remaining 23 test MR images was investigated. As a result, the training models of the CVC sparse subsampling technique, which reduced background voxels in weak proportion to their distances to the border cartilage voxels, produced the highest segmentation accuracies on average for all compartments.