Characterization of discrepancies between manual and automatic segmentation to improve anatomical brain atlases

Abstract

Purpose: To characterize discrepancies between expert manually segmented brain images from Hammers Atlas Database and automatically generated segmentations of the same images; to decide whether they can be attributed to flaws in the automatic segmentation or in the manual segmentation; and to determine general rules that enable these decisions. Theory: Image segmentation plays an important role in clinical neuroscience and experimental medicine for extraction of information from medical images, and it is a fundamental image processing step in medical image analysis. Another important image processing step is image registration that enables quantitative comparison between datasets of different subjects by geometrically aligning one dataset with another. The scientific underpinning of the project is descriptive science combined with inductive reasoning. Method: The study data consisted of 30 T1-weighted 3D MR images along with manual region label volumes from Hammers Atlas Database, and automatically MAPER-generated segmentations of the same images. The comparison of manual and automatic anatomical (semantic) segmentations involves quantitative and qualitative analyses. Image registration was performed with MIRTK to normalize all images into a common space. Discrepancies were then extracted using a custom-designed image analysis process by the program Convert3D. Result: Conclusion: The work has resulted in a model that enables extraction of discrepancies between manual and automatic segmentation into an individual component for quantitative characterization on a per-label basis. A total of 706 465 surface discrepancies were labelled while 1009 holes were found in both manual and automatic segmentations. Probability maps of the discrepancies have been created and can be used as a basis for determining the probability that certain discrepant voxels have been segmented correctly or not. The study yielded insights into how differences between manual and automatic segmentations arise, and how these can be used to develop an improved segmentation that incorporates information from both models.