Decoupling Diffusion and Super-Resolution GANs for Low-Resource 3D Breast MRI Synthesis

Abstract

Breast cancer remains a leading cause of cancer-related mortality worldwide, with early detection through MRI playing a critical role in patient outcomes. However, a significant clinical challenge arises from the positional mismatch between MRI acquisition (prone) and surgery (supine), as breast tissue is highly deformable. Developing a biomechanical model for prone-to-supine mapping could enhance surgical precision, yet such models require large datasets of high-quality 3D breast MRI volumes, which are scarce and expensive to acquire. In order to address this data scarcity, this work proposes a two-stage generative pipeline combining diffusion models to generate 3D breast MRI volumes at lower resolutions (64×64 and 128×128) with super-resolution (SR) GANs to upscale samples to clinical-grade resolutions (256×256 and 512×512). The pipeline, which was specifically designed to augment a small private dataset and enable future prone-to-supine modelling, successfully generates realistic and diverse volumes matching the target distribution and on par with clinical standards, demonstrating feasibility even in low-resource settings.