# AI-Based Brain Volumetry Without MPRAGE? Evaluation of Synthetic T1-MPRAGE from 2D T2/FLAIR

## TL;DR

This study shows that AI can generate synthetic brain scans from commonly used clinical imaging sequences, allowing accurate brain volume measurements typically requiring specialized scans.

## Contribution

The novel use of AI to synthesize MPRAGE-like images from 2D T2/FLAIR sequences for accurate brain volumetry.

## Key findings

- Synthetic MPRAGE-like images showed strong correlation with real MPRAGE for total brain volume (r = 0.99 in controls).
- White matter and hippocampal volumes also demonstrated high agreement with real MPRAGE scans.
- The method works reliably for both healthy individuals and multiple sclerosis patients.

## Abstract

Background: Automated AI-based brain volumetry is increasingly used in clinical practice. T1-weighted sequences (e.g., MPRAGE) are considered the current state-of-the art. However, due to faster acquisition and higher in-plane resolution, 2D anisotropic sequences are often preferred in clinical routine. However, these sequences cannot be processed with currently available AI-volumetry software. Thus, we here aimed to evaluate volumetric data from synthetic MPRAGE-like sequences (mprAIge). Methods: We analyzed 412 datasets (206 conventional MPRAGE and 206 T2w/FLAIR) from healthy volunteers (n = 36) and patients with multiple sclerosis (n = 140). Synthetic mprAIge was generated using SynthSR-CNN and assessed via assemblyNET on the volBrain platform. Total brain volume (TBV), gray and white matter volume (GMV/WMV), and key substructures were compared between mprAIge and conventional MPRAGE. Average volume differences (AVDs) and correlations were calculated. Results: Synthetic mprAIge was generated successfully in all 206 cases. Quantitative analysis demonstrated strong correlation and high agreement for key substructures. TBV showed excellent agreement (AVD: 2.75% for controls, 3.90% for MS patients; r = 0.99 and 0.97, respectively). White matter volume exhibited excellent agreement (AVD: −1.92% for controls, 0.28% for MS patients; r = 0.95). Hippocampal volume also demonstrated good to excellent agreement (AVD: 1.13% for controls, −1.92% for MS patients; r = 0.91 and 0.89, respectively). Conclusions: Synthetic mprAIge enables AI-volumetry software application without limitations. Its volumetric assessments align well with conventional MPRAGE, opening new opportunities for volumetric post-processing and mapping of disease progression.