A comparison of Atlas- and DL-based brain segmentation in hybrid PET/MR for regional SUV quantification

The analysis of abnormalities in various brain regions requires combining metabolic data from PET with anatomical segmentation from MR due to the relatively low resolution of PET images. For brain MR segmentation, dealing with the intricate morphological characteristics of the Central Nervous System (CNS) represents a fundamental challenge in medical image analysis. While traditional methods such as Atlas-based techniques have been widely used, recent advancements in deep learning (DL) have significantly transformed the field. In this study, we automatically segmented two representative brain areas from hybrid PET/MR scans of twelve PD and three MSA patients using both Atlas- and DL-based methods. Then we compared the Standardized Uptake Values (SUVs) and accuracy of segmentation (DSC, MSD and HD) in the corresponding regions, with manual segmentation used as the ground truth for comparison. The Atlas-based brain segmentation relied on an atlas template from the automated anatomical labeling atlas, which includes 70 segmented regions labeled from 1 to 70. On the other hand, DL-based brain segmentation utilized a 3D transformer model based on MONAI framework for whole brain segmentation, employing a pre-trained model for inferring the whole brain with 133 structures on T1-weighted MR images. Both methods involved the use of SPM12 for the 3D affine registration of T1-weighted MR images to the MNI space. Manual segmentation was carried out by a clinical neuroimaging expert using ITK-SNAP, focusing on two cerebral nuclei containing four sub-regions (left Caudate, right Caudate, left Putamen, and right Putamen). The results of this study indicate that the DL-based method produced superior segmentation accuracy compared to the Atlas-based method. However, there were no significant differences in the SUV_max and SUV_mean across the different methods for the segmentation of Caudate and Putamen. Despite being more accurate for Caudate and Putamen segmentation, the DL-based method had little effect on the calculation of their SUVs in hybrid PET/MR scans, as compared to the widely used Atlas-based method. This comparison and evaluation could potentially extend to other structures based on the algorithms.