Probabilistic Image Registration via Deep Multi-class Classification: Characterizing Uncertainty

摘要

We present a novel approach to probabilistic image registration that leverages the strengths of deep-learning for modeling agreement between images. We use a deep multi-class classifier trained on different classes of patch pairs, including unrelated, registered, and a collection of discrete displacements between patches. The displacement classes alleviate the need for registration-time optimization by gradient descent; instead, posterior probabilities are used to directly predict expected values of displacements on the lattice of sampled locations. These, in turn, are used to update transformation parameters and the process is iterated. We empirically demonstrate the accuracy of our proposed method on deformable cross-modality registrations of brain MRI, and show improved results compared to Mutual Information based method on challenging data that includes simulated resections. Our approach enables local predictions of registration uncertainty and diagnostics that can indicate areas that seem unrelated in the two images. Uncertainty estimates provide end-users with intuitively actionable information on the quality of registration in interventional and surgical settings.