Automatic segmentation of the bones from MR images of the knee

We present and validate a hybrid segmentation scheme based around 3D active shape models, which is used to automatically segment the three bones in the knee joint. This scheme is automatically initialised using an affine registration to an atlas. The accuracy and robustness of the approach was experimentally validated using an MR database of 20 fat suppressed Spoiled Gradient Recall images. A median Dice Similarity Coefficient (DSC) of 0.89, 0.96 and 0.96 was obtained for the patella, tibia and femur which illustrates the accuracy of the approach. The robustness of this scheme to initialisation was validated by segmenting each knee image 19 times, each time using a different image in the database as the atlas. An overall segmentation failure rate (DSC < 0.75) of only 3.60% shows that the scheme was robust to initialisation.