Quality Assessment and Model Training Impact of Synthetic Images in Medical Diagnostic Imaging: Systematic Review.
Authors
Affiliations (6)
Affiliations (6)
- Information Systems Department, Catalan Health Service (CatSalut), Gran Via de les Corts Catalanes, 587, Barcelona, 08007, Spain, +34 93 403 85 85.
- Digitalization for the Sustainability of the Heathcare System (DS3) Research Group, Barcelona, Spain.
- Fundació Tic Salut i Social (FTSS), Barcelona, Spain.
- Process Unit, Directorate of Quality, Processes and Innovation, Vall d'Hebron University Hospital, Barcelona, Spain.
- Faculty of Engineering, Maersk Mc-Kinney Moller Institute, University of Southern Denmark, Odense, Denmark.
- Biostatistics Unit, Medical School, Universitat Autònoma de Barcelona, Bellaterra, Spain.
Abstract
AI has increasingly been used to generate synthetic images to overcome the scarcity of large, regulatory-compliant medical image datasets. However, their quality and actual utility for algorithm development across deep learning architectures, medical settings, and image types have been poorly assessed. This study aimed to review the use of synthetic images for algorithm development and to assess their quality and effects on downstream algorithms trained on these images, according to AI architecture, medical setting, and image type. The authors conducted a systematic review of studies published between January 2020 and September 2025. PubMed, Scopus, IEEE Xplore, and Embase were searched for peer-reviewed studies evaluating synthetic images generated from human medical imaging data. Data were extracted for imaging modalities, generative architectures, image-quality assessment methods, downstream model performance, explainability approaches, reproducibility practices, and risk of bias. Given substantial heterogeneity and incomplete statistical reporting, findings were synthesized descriptively. A total of 160 studies were included. Generative adversarial network-based architectures dominated image generation (119 studies, 74.4%), although hybrid and diffusion-based approaches became increasingly represented in recent years. Image-quality evaluation was highly heterogeneous: 50 studies included human-reader evidence, and 101 reported an automated image-quality assessment. Synthetic-only training generally underperformed real-data training, including a median Dice similarity coefficient change of -10.2% for segmentation. Adding synthetic data to real training sets was associated with median improvements in classification precision (+7.8%), recall (+5.3%), <i>F</i><sub>1</sub>-score (+5.5%), accuracy (+4.3%), and area under the curve (+4.0%), while the median Dice coefficient improvement was +1.2%. Explainability evidence was reported in 68 (42.5%) studies and was rarely validated against expert annotations. Reproducibility reporting was inconsistent, and only 38% (60/160) of studies were judged to have low overall risk of bias. Unlike previous reviews restricted to particular diseases, imaging modalities, or generative architectures, this review provides a cross-domain quantitative synthesis of both synthetic-image quality assessment and downstream model performance. Synthetic images currently appear more valuable as a complement to real training data than as a replacement, with modest improvements in classification performance, smaller gains in segmentation, and generally poorer performance with synthetic-only training. For researchers, developers, and regulators, these findings highlight the need for standardized multimethod image-quality evaluation, complete reporting, and rigorous external validation before synthetic-image pipelines are translated into clinical AI applications.