This paper presents a novel diffusion-based framework for uterine MRI synthesis. To address the challenges of existing diffusion models in generating anatomically accurate female pelvic images, we integrate a conditional and unconditional denoising diffusion probabilistic model (DDPM) and a latent diffusion model (LDM) in 2D and 3D. This approach generates anatomically consistent, high-fidelity synthetic images that accurately mimic real scans, providing a valuable resource for training robust diagnostic models. We evaluate the quality of the synthesis using advanced perceptual and distributional metrics and benchmark it against standard reconstruction methods, demonstrating significant improvements in diagnostic accuracy on key classification tasks. Blinded expert evaluation further validated the clinical realism of the synthetic images. To support reproducible research and advance unbiased AI in obstetrics and gynecology, we release the model, along with privacy protections and a comprehensive synthetic uterine MRI dataset.
