Deep Learning for Synthetic Postcontrast T1-Weighted MRI: A Systematic Review With Targeted Meta-Analysis of Brain...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Deep Learning Applications

Background:

• Gadolinium-based contrast agents are crucial for MRI but pose risks.
• Deep learning (DL) offers a promising method to synthesize postcontrast T1-weighted MRI from precontrast sequences alone.

Purpose of the Study:

• To systematically review DL-based synthesis of postcontrast T1-weighted MRI.
• To characterize DL model architectures and evaluation practices across subspecialties.
• To perform meta-analysis on brain tumor studies using DL for image synthesis.

Main Methods:

• Systematic literature search across major databases (PubMed, Embase, etc.) up to January 2025.
• Independent screening and data extraction by two reviewers on study characteristics, metrics, and reader studies.
• Risk of bias assessment using modified QUADAS-2 and random-effects meta-analysis for brain tumor studies.

Main Results:

• 41 studies met inclusion criteria, predominantly in neuroimaging (59%).
• Generative adversarial networks and convolutional neural networks were common architectures.
• High heterogeneity (I² > 99%) was observed, with pooled SSIM of 0.92 and PSNR of 30.6 dB for brain tumors.
• Pathology-specific evaluation showed lower metrics; only 37% included reader studies.

Conclusions:

• DL-based postcontrast MRI synthesis is technically feasible but faces challenges.
• Substantial heterogeneity in study design and inconsistent evaluation metrics limit clinical translation.
• Standardized evaluation, including reader studies and external validation, is crucial for clinical adoption.