A generative probability model of joint label fusion for multi-atlas based brain segmentation.

Highlights: [•] A generative probability model is proposed to describe the labeling procedure. [•] The labeling dependency is explicitly modeled to achieve largest labeling unanimity among atlas patches. [•] The sparsity constraint is imposed label fusion weights in order to reduce the risk of including misleading atlas patches. [•] EM-based solution is provided to infer the labels from the generative probabilistic model. [Copyright &y& Elsevier]