Simultaneous Dual-View Mammogram Synthesis Using Denoising Diffusion Probabilistic Models

A multi-institution research team has developed a novel AI system that generates synthetic, paired mammogram images to address a critical data gap in breast cancer screening. The model, a three-channel Denoocing Diffusion Probabilistic Model (DDPM), is designed to simultaneously create the two standard diagnostic views—the craniocaudal (CC) and mediolateral oblique (MLO)—from a single breast. Its key innovation is a third input channel that encodes the absolute difference between the two views, which guides the diffusion process to learn and preserve the coherent anatomical relationships between the different projections. This ensures the synthetic pairs are geometrically consistent, mimicking how a real breast would appear from different angles.

The researchers fine-tuned a pre-trained DDPM from Hugging Face on a private screening dataset. Evaluation showed the difference-based encoding helped preserve global breast structure across views, producing synthetic CC-MLO pairs that closely resemble real clinical acquisitions. The work, accepted at SPIE Medical Imaging 2025, demonstrates the feasibility of using this guided diffusion approach for high-fidelity medical image synthesis.

This technology's primary application is augmenting incomplete datasets, which currently limit the development of robust AI algorithms for breast cancer detection. Many historical and real-world datasets lack complete, paired views for each patient, hindering the training of models that rely on cross-view analysis. By generating realistic, anatomically consistent synthetic data, this model can help create larger, more balanced training sets for next-generation diagnostic AI, potentially improving early detection rates.