Supervised Volumetric Segmentation of White and Gray Matter from Brain Positron Emission Tomography Images Using Magnetic Resonance Labels

Abstract Background: This study evaluates the performance of five deep convolutional neural networks (DCNNs) for supervised segmentation of white matter (WM) and gray matter (GM) in brain positron emission tomography (PET) images using label maps derived from corresponding magnetic resonance (MR) images. The goal is to reduce dependency on hybrid PET/Magnetic resonance imaging (MRI) by extracting anatomical information solely from PET images, thereby providing a potential pathway for MRI-free partial volume correction (PVC). Methods: A total of 300 PET images and their corresponding MR-derived labels from the OASIS-3 dataset were used. The five DCNNs, including the UNet, RegUNet, VNet, SegResNet, and HighResNet, were implemented within the Medical Open Network for Artificial Intelligence (MONAI) framework. The networks were evaluated using Dice score and Intersection over Union (IoU) metrics. Results: Among the networks, VNet demonstrated superior performance for GM and WM segmentation, achieving Dice scores of 61.18% and 76.23%, and IoU scores of 44.15% and 61.62%, respectively. UNet and VNet showed competitive performance for WM segmentation, with no statistically significant differences between them. Conclusions: These findings provide insights into the performance of DCNNs for PET image segmentation, highlighting VNet’s capability and emphasizing the potential for optimizing segmentation techniques for PVC applications.