The application of generative neural radiance fields in medical imaging: A systematic review.

BACKGROUND: Generative neural radiance fields (GNeRF) extend NeRF with adversarial, variational, and pose-aware techniques, enabling 3D reconstructions from sparse 2D medical images. While NeRF reviews exist, no prior synthesis has focused specifically on GNeRF in medical imaging. PURPOSE: This systematic review evaluates the application of GNeRF in medical imaging, summarizing models, modalities, outcomes, and methodological gaps. METHODS: Following PRISMA 2020 guidelines, we searched major scientific databases and digital libraries for studies published between January 2021 and April 2025. Eligible articles applied GNeRF or its variants to medical image analysis. Data on study design, modality, evaluation metrics, and reported outcomes were extracted and synthesized. RESULTS: Of 260 records, 8 studies met inclusion. These studies covered multiple imaging modalities, with most using synthetic DRR datasets derived from CT and others extending to coronary angiography and MRI applications. Architectures varied: MedNeRF and UMedNeRF retained baseline GNeRF backbones, while RepMedGraf, ACNeRF, and imp-MedNeRF introduced adversarial or pose-aware techniques. Metrics such as PSNR, SSIM, and FID were used to benchmark performance. However, all studies relied on small or synthetic datasets, limiting generalizability. CONCLUSIONS: Reviewed studies indicate that GNeRF methods show early promise in reducing radiation dose in ionizing imaging modalities (e.g., CT, x-ray) and in reconstructing anatomy from sparse views; however, translation to clinical workflows is not yet feasible. Progress requires larger multi-institutional datasets, efficient architectures, and prospective validation. This review defines the current evidence base and outlines priorities for future interdisciplinary development.