We develop a semi-automatic technique for segmentation of hippocampal subfields in T2-weighted in vivo brain MRI. The technique takes the binary segmentation of the whole hippocampus as input, and automatically labels the subfields inside the hippocampus segmentation. Shape priors for the hippocampal subfields are generated from shape-based normalization of whole hippocampi via the continuous medial representation method. To combine the shape priors with appearance features, we use a machine learning based method. The key novelty is that we treat the mistakes made by the shape priors as bias, which can be detected and corrected via learning. The main advantage of this formulation is that it significantly simplifies the learning problem by taking full advantage of current segmentations and focusing on only improving their drawbacks. Experiments show that the bias removal approach achieves significant improvement in all subfields. Our bias removal idea is general, and can be applied to improve other segmentation methods as well.