Deep metrics have been shown effective as similarity measures in multi-modal image registration; however, themetrics are currently constructed from aligned image pairs in the training data. In this paper, we proposea strategy for learning such metrics from roughly aligned training data. Symmetrizing the data corrects biasin the metric that results from misalignment in the data (at the expense of increased variance), while randomperturbations to the data, i.e. dithering, ensures that the metric has a single mode, and is amenable to registrationby optimization. Evaluation is performed on the task of registration on separate unseen test image pairs. Theresults demonstrate the feasibility of learning a useful deep metric from substantially misaligned training data, insome cases, the results are significantly better than from Mutual Information. Data augmentation via ditheringis, therefore, an effective strategy for discharging the need for well-aligned training data; this brings deep learningbased registration from the realm of supervised to semi-supervised machine learning.
展开▼
