ATLAS DATABASES OF HUMAN BRAIN ANATOMY: QUALITATIVE AND QUANTITATIVE COMPARISON

Abstract

The purpose of this study was to quantitatively and qualitatively compare the two brain atlases databases, Hammersmith and 2012 MICCAI Multi-Atlas Labelling Challenge data, by developing a quantitative method. Theory Anatomical atlases of the human brain provide reference information about its structure. Researchers and practitioners use them for varied purposes such as automatic image segmentation, biomarker discovery, and identification of relationships between brain structure and function. There is no worldwide agreement on how to segment the human brain, which gives rise to difficulties and differences in the description of brain structures: the brain atlas concordance problem. Two widely used atlas databases are investigated in this study: the Hammersmith (HM) and the 2012 MICCAI Multi-Atlas Labelling Challenge Data (MGC). Both consist of T1-weigthed 3D magnetic resonance (MR) brain images of 30 study participants, with corresponding anatomical label sets. Method The study data consisted of 60 MR brain images (30 from each database) with 120 corresponding segmentations (30 manual and 30 automatically generated, times two databases). The automatic segmentations of the MGC images were based on the HM atlas, and the automatic segmentations of the HM images were based on the MGC atlas. The study was composed of two main parts, a qualitative comparison and a quantitative comparison. The quantitative comparison was developed during the study and was evaluated by juxtaposition with the qualitative results. The quantitative method included calculation of the most frequent coinciding regions, the Jaccard coefficient, and the volume ratio between corresponding regions from each database. The qualitative comparison was composed of predicting differences based on a comparison between the delineation protocols for a subset of regions, a visual analysis of overlaps and a global comparison of region names included in the protocols.Result: Conclusion The main difference between the protocols is that cortical regions only include the actual cortical grey matter in the MGC, whereas HM includes adjacent white matter as part of the region. 73 of the HM regions had matching region names with the MGC regions and 86 of the MGC regions had matching region names with the HM regions. The main differences between the defined regions from the databases were the subdivisions of regions and inclusion of different gyri. 40 and 43 of the HM regions had the matching MGC region as the most frequent coinciding region in the HM images respective the MGC images, and 76 of the MGC regions had the matching HM region as the most frequent coinciding region in both the HM and MGC images. Both atlas databases leave certain brain regions unclassified (assigned the background label). The Jaccard coefficient showed that the greatest overlap occurred between the regions that had matching names with regions in the other protocol. The HM regions generally had larger volumes compared to the corresponding MGC regions, although there were exceptions where MGC regions were almost twice the size of the corresponding HM regions. The quantitative comparison confirmed most of the predictions and revealed multiple additional overlaps and insights that could not be predicted just based on studying the protocols. The two atlas databases differ systematically, reflecting the differences in purpose and priorities that guided the underlying manual segmentation procedures. The quantitative method developed in this project showed to be able to confirm the most important predictions and reveal additional insight that the qualitative analysis could not predict.