Impact of Mri Modality Selection on Glioma Sub-Region Segmentation Using 3D U-Net and Attention U-Net: A Comparative Evaluation Across Ten Mri Sequences Publisher

Summary: Can MRI scans map glioma without contrast dye? A segmentation study favored T1Gd with Attention U-Net overall, while ASL showed promise as a noncontrast option. #BrainCancer #MedicalImaging

Khorasani A
Source: European Journal of Radiology Open Published:2026

Abstract

Background: Accurate segmentation of glioma sub-regions is essential for treatment planning and clinical decision-making. The selection of MRI input modality plays a critical role in the performance of deep learning–based segmentation models, as each weighting captures distinct tissue contrasts and physiological characteristics. This study performed a comparative evaluation of the impact of MRI modality selection on semantic segmentation of glioma sub-regions. Methods: A total of 468 subjects from the UCSF-PDGM dataset were included. Ten MRI weightings were assessed independently: apparent diffusion coefficient (ADC) maps, arterial spin labeling (ASL), diffusion-weighted imaging (DWI), fractional anisotropy (FA), mean diffusivity (MD), susceptibility-weighted imaging (SWI), T1-weighted, T1-weighted post-contrast (T1Gd), FLAIR, and T2-weighted imaging. For each modality, dedicated 3D U-Net and 3D Attention U-Net models were trained. TOPSIS multi-criteria ranking was applied to identify optimal model–modality combinations per sub-region and overall. Results: The 3D Attention U-Net numerically outperformed the standard U-Net across all modalities and sub-regions, with significant differences in several sequence–region comparisons. Sub-region–specific TOPSIS rankings identified ASL as optimal for active tumor segmentation, T1Gd for necrotic core delineation, and FLAIR for peritumoral edema. Overall TOPSIS ranking, aggregated across all sub-regions, identified the Attention U-Net trained on T1Gd as the most consistently high-performing combination. In settings where gadolinium administration is contraindicated, ASL and MD emerged as the strongest non-contrast alternatives. Conclusions: Within the present experimental framework, Attention U-Net trained with T1Gd yielded the most consistently high-performing single-modality results for comprehensive segmentation, whereas ASL emerged as a promising contrast-free candidate warranting further evaluation. © 2026 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC license. http://creativecommons.org/licenses/by-nc/4.0/