Combining binary classifiers for automatic cartilage segmentation in knee MRI

We have developed a method for segmenting tibial and femoral medial cartilage in MR knee scans by combining two k Nearest Neighbors (kNN) classifiers for the cartilage classes with a rejection threshold for the background class. We show that with this threshold, two binary classifiers are sufficient compared to three binary classifiers in the traditional one-versus-all approach. We also show that the combination of binary classifiers produces better results than a kNN classifier that is trained to partition the voxels directly into three classes. The resulting sensitivity, specificity and Dice volume overlap of our method are 84.2%, 99.9% and 0.81 respectively. Compared to state-of-the-art segmentation methods, our method outperforms a fully automatic method and is comparable to a semi-automatic method.