Automated Brain Segmentation in Accelerated T2-Weighted MRI: Effects of Deep Learning-Based Reconstruction

Abstract

Background: Automated brain volumetry is increasingly used for clinical and research assessment, but its outputs depend on acquisition and reconstruction settings. We evaluated how parallel-imaging acceleration and vendor deep-learning reconstruction (Siemens Deep Resolve) affect automated volumetry on T2-weighted turbo spin-echo MRI, and whether deep-learning reconstruction stabilises measurements across protocols. Methods: Four healthy volunteers each underwent 14 T2-TSE-TRA acquisitions on a 1.5 T scanner, covering seven protocols (baseline, GRAPPA R = 2, 3, 4, and SMS x2,x3, x4), with each protocol acquired separately once with Deep Resolve off and once with Deep Resolve on, for 56 acquisitions in total. Six pipelines were applied: SynthSeg, OpenMAP-T2, and GOUHFI 2.0 for subcortical parcellation, and TotalSegmentator MRI, HD-BET, and NV-Segment-CTMR for whole-brain masking. A three-rater STAPLE consensus of six subcortical structures on the baseline acquisition served as the expert reference. Reference-relative accuracy (absolute percentage error), cross-protocol dispersion (coefficient of variation), and spatial reproducibility (Dice coefficient and 95th-percentile surface distance) were compared between reconstruction states. Results: Deep Resolve improved reference-relative accuracy most for GOUHFI 2.0 (pooled median absolute percentage error 28.4% to 21.9%), modestly for SynthSeg (16.2% to 15.5%), and had a mixed, structure-dependent effect for OpenMAP-T2. It approximately halved cross-protocol dispersion for all three parcellation tools (coefficient of variation 2.62% to 1.39% for SynthSeg, 4.42% to 2.14% for OpenMAP-T2, and 21.75% to 7.87% for GOUHFI) and improved spatial reproducibility in parallel. It eliminated the systematic GOUHFI under-segmentation seen at high acceleration under conventional reconstruction (median deviation -31.6% at GRAPPA R = 4; 14 failure events under conventional reconstruction, none with Deep Resolve). Whole-brain masking was robust in both states. Conclusions: Deep Resolve improved the robustness of automated brain segmentation in accelerated T2-weighted MRI, reducing cross-protocol variability and improving spatial reproducibility, with the largest reduction in reference-relative error in the most acceleration-sensitive pipeline. The elimination of severe segmentation failures observed under conventional reconstruction highlights its potential to support reliable volumetry at higher acceleration. Reconstruction optimization therefore offers a practical route to more consistent quantitative measurements across acquisition protocols.