Atlas based method for the automated segmentation and quantification of knee features: Data from the osteoarthritis initiative

Abstract

This paper presents a fully unsupervised segmentation method for the segmentation of 3D DESS MRI images of the human knee. Five MRI knees manually segmented by human experts are used as reference atlases to automatically segment subsequent MRI images. The five segmentations are averaged to create the knee segmentation. The methodology was tested on the pilot Osteoarthritis Initiative (OAI) image set of MRI DESS sequences. The data includes longitudinal images from healthy normals and subjects with osteoarthritis (OA) scanned twice at baseline and at the 24 month follow-up. The segmentation methodology was able to create precise cartilage segmentations of the knees that were used to extract volume, thickness and subchondral bone plate curvature information of the knee. The quantitative thickness showed precisions ranging from 0.025mm to 0.051mm. The longitudinal reproducibility of the cartilage thickness measurement methodology showed intra-class correlations coefficients (ICC) ranging from 0.39 to 0.79.