# SynthSR-generated 3D T1-weighted MRI from routine 2D clinical images: Validation for VSRAD analysis

## TL;DR

This study shows that 3D MRI images generated from 2D scans using SynthSR can be used for Alzheimer's analysis as effectively as traditional 3D scans.

## Contribution

The study validates the use of SynthSR-generated 3D MRI images for VSRAD analysis in Alzheimer's detection.

## Key findings

- SynthSR-generated 3D volumes showed high rank correlations with standard 3D volumes for key Alzheimer's markers.
- Diagnostic performance of SynthSR-based VSRAD analysis was comparable to standard 3D volumes.
- Absolute volume estimates from SynthSR were larger but clinically relevant indices remained consistent.

## Abstract

The Voxel-based Specific Regional Analysis System for Alzheimer’s Disease (VSRAD), a voxel-based morphometry tool quantifying medial temporal lobe atrophy as region-specific Z-scores, is widely used in clinical practice for detection of Alzheimer’s disease (AD). However, it typically require high-resolution 3D T1-weighted MRI, which is often difficult to acquire in elderly or cognitively impaired patients. This study aimed to evaluate whether 3D volumes generated by SynthSR from 2D T1-weighted MRI can yield volumetric and VSRAD-derived indices that are comparable to those from standard 3D images, by assessing agreement, rank consistency, and diagnostic performance.

In this retrospective single-center study, MRI data from 75 patients were analyzed using both standard 3D T1-weighted images and SynthSR-generated 3D volumes reconstructed from 2D T1-weighted sequences. Regional brain volumes and four key Z-score indices from VSRAD were compared using Wilcoxon signed-rank tests with Bonferroni correction, robust Bland–Altman analysis, Spearman’s rank correlation, and receiver operating characteristic (ROC) curve analysis focusing on Score 1 “Severity.”

All Z-score indices and segmented volumes showed significant absolute differences between the two methods (p < 0.0071), with SynthSR-based data generally yielding larger volume estimates. Despite these differences, Spearman’s ρ remained consistently high (ρ > 0.7) for brain volume and Score 3 “Ratio,” and other clinically relevant indices also demonstrated moderate correlations. ROC analysis demonstrated high value of the area under the curve (AUC) values for both standard 3D volumes (0.90) and SynthSR-generated 3D volumes (0.96), with no statistically significant difference between the two methods (Z = 0.009, p = 0.99, DeLong’s test).

Although SynthSR-based images produced systematically different absolute values, they preserved rank-order correlations and maintained diagnostic performance comparable to that of standard 3D volumes in VSRAD analysis. Considering that conventional 3D acquisitions are often difficult to obtain in elderly patients undergoing dementia screening, SynthSR-based reconstruction may represent a practical alternative in routine clinical practice, particularly for Score 1 “Severity,” the most clinically relevant marker of hippocampal atrophy.