Estimation of Registration Accuracy Applied to Multi-Atlas Segmentation

Multi-atlas registration-based segmentation has recently become a popular technique in medical imaging. Since the quality of individual atlas segmentations affect the quality of the results, atlas selection and atlas fusion have become important areas of research for multi-atlas segmentation. In this paper, we present an automatic technique that approximately calculates the quality of registration. We applied our method to multi-atlas segmentation and find that our measure correlates strongly ( = 0.79) with the ground truth DICE similarity index. When applied to atlas fusion using a majority vote technique weighted by our measure of registration quality, our algorithm performs statistically better than both an un-weighted majority vote technique and a voting technique weighted by residual normalized mutual information.