Deep Learning for Synthetic Postcontrast T1-Weighted MRI: A Systematic Review With Targeted Meta-Analysis of Brain Tumor Studies.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, NYU Grossman School of Medicine, New York, NY.
- Department of Radiology, Columbia University Irving Medical Center, New York, NY.
Abstract
<b>Background.</b> Gadolinium-based contrast agents remain essential for MRI but carry risks. Deep learning (DL) methods have emerged as a potential approach for synthesizing postcontrast T1-weighted images from precontrast sequences alone. <b>Objective.</b> The objective of this study was to systematically review DL-based synthesis of postcontrast T1-weighted MRI, characterize model architectures and evaluation practices across subspecialties, and perform targeted meta-analysis where sufficient literature existed. <b>Evidence Acquisition.</b> A systematic search of PubMed, Embase, Cochrane Central, Scopus, and Web of Science (through January 16, 2025) identified peer-reviewed studies using DL to synthesize postcontrast T1-weighted MRI from precontrast sequences in adults. Two reviewers independently screened studies, extractingdata on subspecialty, architecture, quantitative metrics, pathology-specific evaluation, and reader studies. Risk of bias was assessed using modified QUADAS-2. Random-effects meta-analysis was performed for brain tumor studies. <b>Evidence Synthesis.</b> Of 268 records after deduplication, 41 met inclusion criteria. Most studies focused on neuroimaging (n = 24, 59%), followed by breast (n = 7, 17%) and body imaging (n = 6, 15%). Generative adversarial networks (n = 20, 45%) and convolutional neural networks (n = 19, 43%) predominated. Structural similarity index measure (SSIM, n = 31, 76%) and peak SNR (PSNR, n = 28, 68%) were the most common metrics. Fifty-one percent (n = 21) of studies performed pathology-specific evaluation, which showed substantially lower SSIM and PSNR compared with whole-image metrics. Thirty-seven percent (n = 15) included reader studies, 29% (n = 12) released code, and 61% (n = 25) used single-institution data. Meta-analysis of 15 brain tumor studies (30 models) yielded pooled SSIM of 0.92 (95% CI, 0.90-0.93) and PSNR of 30.6 dB (95% CI, 28.6-32.6). Given extreme heterogeneity (I<sup>2</sup> > 99%), pooled estimates should be interpreted as descriptive. <b>Conclusion.</b> DL-based postcontrast MRI synthesis shows technical feasibility across subspecialties but suffers from substantial heterogeneity in study design, inconsistent quantitative metric computation, and limited clinical validation. <b>Clinical Impact.</b> Limited rates of reader studies and external validation represent key barriers to clinical translation of DL-based postcontrast MRI synthesis. Standardized evaluation workflows incorporating whole-image metrics, pathology-specific assessment, and reader studies are essential before these techniques can be translated into clinical practice.