Automatic segmentation of bones and inter-image anatomical correspondence by volumetric statistical modelling of knee MRI

The detection of cartilage loss due to disease progression in Osteoarthritis remains a challenging problem. We have shown previously that the sensitivity of detection from 3D MR images can be improved significantly by focusing on regions of `at risk' cartilage defined consistently across subjects and time-points. We define these regions in a frame of reference based on the bones, which requires that the bone surfaces are segmented in each image, and that anatomical correspondence is established between these surfaces. Previous results has shown that this can be achieved automatically using surface-based Active Appearance Models (AAMs) of the bones. In this paper we describe a method of refining the segmentations and correspondences by building a volumetric appearance model using the minimum message length principle. We present results from a study of 12 subjects which show that the new approach achieves a significant improvement in segmentation accuracy compared to the surface AAM approach, and reduce the variance in cartilage thickness measurements for key regions of interest. The study makes use of images of the same subjects obtained using different vendors' scanners, and also demonstrates the feasibility of multi-centre trials.