Hippocampus segmentation using a stable maximum likelihood classifier ensemble algorithm

We develop a new algorithm to segment the hippocampus from MR images. Our method uses a new classifier ensemble algorithm to correct segmentation errors produced by a multi-atlas based segmentation method. Our classifier ensemble algorithm searches for the maximum likelihood solution via gradient ascent optimization. Compared to the additive regression based algorithm, LogitBoost, our algorithm avoids the numerical instability problem. Experiments on a hippocampus segmentation problem using the ADNI data show that our algorithm consistently converges faster and generalizes better than AdaBoost.