Method for evaluation of different MRI segmentation approaches

Abstract

Common comparison indices for evaluating MRI segmentation are usually based on: percentage of volumetric measure of different tissue types; cross-correlation matrix between a segmentation result and the manual segmentation result. Since manual segmentation is not the true reference, the authors proposed a new evaluation method which simulated MR images. Extracted cross covariance inhomogeneity map from a real MRI data, together with a manual segmentation are used to generate MRI simulated images. The result showed that common evaluation method using manual segmentation as reference is subject to intra- and inter-rater variations and requires time consuming manual segmentation step for each study to evaluate, while the proposed method is expected to provide faster and more objective measure for comparing different MR segmentation methods.