Improving Similarity Metric of Multi-modal MR Brain Image Registration Via a Deep Ensemble

Abstract

Brain image registration fuses and aligns sets of structural or functional images within individual and population studies. The similarity metric is an image registration component used for detecting the same target region in different images. Multi-modal image registration constitutes one of the greatest challenges in medical imaging as it adds even more variability to the tissue and organ appearance, shape, and positioning. This paper contains two contributions to solve this complex problem: (1) we propose a solution to compute the similarity metric based on a deep ensemble method. It combines multiple traditional and deep similarity metrics into a single improved similarity map; (2) we propose novel evaluation metrics to validate the results. Experiment results in the context of T1- and T2-weighted MR images of the human brain show a major improvement to the state-of-the-art, especially in reducing the false-positive region occurrences.