Fig. 4:

Schematic of our proposed maximum likelihood registration with deep binary classifiers. Based on the initial misalignment in the dataset, we can perform multiple iterations to jointly learn model and transformation parameters. Our framework includes a deep binary CNN classifier, a Spatial Transform Module, and a 3D patch selector. Our classifier architecture is inspired by DenseNet. The aggregated logits signal (over a set of sampled patches) is used for the optimization of the transformation parameters.