A framework for quality control of corpus callosum segmentation in large-scale studies

Abstract

BackgroundThe corpus callosum (CC) is the largest white matter structure in the brain, responsible for the interconnection of the brain hemispheres. Its segmentation is a required preliminary step for any posterior analysis, such as parcellation, registration, and feature extraction. In this context, the quality control (QC) of CC segmentation allows studies on large datasets with no human interaction, and the proper usage of available automated and semi-automated algorithms. New methodWe propose a framework for QC of CC segmentation based on the shape signature, computed at 49 distinct resolutions. At each resolution, a support vector machine (SVM) classifier was trained, generating 49 individual classifiers. Then, a disagreement metric was used to cluster these individual classifiers. The final ensemble was constructed by selecting one representation from each cluster. ResultsThe proposed framework achieved an area under the curve (AUC) metric of 98.25% on the test set (207 subjects) employing an ensemble composed of 12 components. This ensemble outperformed all individual classifiers. Comparison with existing methodsTo the best of our knowledge, this is the first approach to assess quality of CC segmentations on large datasets without the need for a ground-truth. ConclusionsThe shape descriptor is robust and versatile, describing the segmentation at different resolutions. The selection of classifiers and the disagreement measure lead to an ensemble composed of high-quality and heterogeneous classifiers, ensuring an optimal trade-off between the ensemble size and high AUC.