Abstract
BACKGROUND: Magnetic resonance imaging (MRI) has traditionally been preferred over computed tomography (CT) for segmentation of intracranial structures due to its superior low contrast resolution. However, a reliable CT-based segmentation could improve patient management when MRI is not practical. Despite advancements in CT imaging, such as enhanced tissue differentiation using virtual monoenergetic imaging (VMI) from dual energy CT, volumetric analysis remains underexplored. PURPOSE: The aim was to evaluate the feasibility of using deep learning (DL) models for segmentation of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF)-using virtual monoenergetic images (VMI). METHODS: The study included 26 patients (training/validation: 21, test: 5) who underwent brain imaging on a dual-layer CT and a T1-weighted MR scan. MR-based segmentation of GM, WM, and CSF served as the ground truth for training and testing of the DL models. Models included a baseline U-Net++ trained on 70 keV VMIs and several U-Net and U-Net++ extensions designed to leverage spectral information from multiple VMIs (50, 70, and 120 keV). Model performance was evaluated using Dice Similarity Coefficient (DSC) and volumetric accuracy. RESULTS: The U-Net++ (Aug) model, using VMIs as augmentations of the input data, outperformed the baseline and other models with DSC 0.84, 0.77, and 0.88 for WM, GM, and CSF, respectively. The superiority was significant compared to several of the other models, and most notably compared to the baseline model with DSC of 0.81 for WM (p = 0.002) and 0.75 for GM (p = 0.002). U-Net++ (Aug) had an average volumetric error of 12%, while U-Net (Gated) had the lowest error at 10%. CONCLUSIONS: This study demonstrates the feasibility of CT-based segmentation of intracranial tissue using DL and VMI. The improved accuracy of the U-Net++ (Aug) compared to the baseline model suggests that spectral information may enhance segmentation performance.