Our method can be applied to the segmentation of tibial and patellar cartilages, and can be used in morphometric assessment for diagnosis, prognosis and treatment planning of OA.
FIGURE (OPTIONAL)http://abstract.rsna.org/uploads/2015/15016975/15016975_z4f6.jpg
BackgroundAutomated segmentation of cartilage from knee MR images is a prerequisite process for the morphometric assessment of cartilage structure. However, it has suffered from the challenges of cartilage due to the low contrast surroundings and large variations of appearance and shape. We propose an automatic segmentation method of cartilage in knee MRI using multi-atlas-based locally-weighted voting (LWV) and graph cuts.
EvaluationOur method was tested on a dataset consisting of 14 volumetric 3T PD VISTA coronal knee joint MRI scans of osteoarthritis (OA) patients. All images were acquired using Achieva 3.0T Philips Medical systems with a pixel size of 0.31mm, a slice thickness of 0.5mm and 512 x 512 resolutions. For training atlases, all images were manually segmented by three experts and tested by leave-one-out validation method. In atlas selection, coronal slab-average projection images were generated in training and target image and 2D affine registration was performed to select similar training images to the target image. In bone segmentation, bone atlases were aligned to the target bone by volume- and object-based two-stage 3D affine registration. Bone was then segmented with these aligned bone atlases using LWV and graph cuts. In cartilage segmentation, cartilage atlases were aligned to the target cartilage by applying bone's 3D affine transformations. Then shape-constrained weighting of LWV and graph cuts were proposed to restrict the cartilage label near the bone surfaces. For evaluation, our segmentation results were visually assessed and Dice similarity coefficient was measured between automatically and manually segmented femur and femoral cartilages.
DiscussionAtlas selection enabled our method to be robust to the variation of the cartilage shape. Volume- and object-based two-stage alignment reduced the misalignment of target and training atlases without non-rigid registration. Our shape-constrained weighting of LWV and graph cuts enabled our method to avoid the leakage into low contrast surroundings.