Making DBT Accessible Without a DBT Scanner: A Generative AI Approach to Synthesizing Pseudo-Tomosynthesis from Standard Mammograms

Abstract

This study presents MammoTomo, a physics-constrained generative AI framework that synthesizes pseudo-digital breast tomosynthesis (pseudo-DBT) volumes from standard 2D mammograms, addressing the hardware barrier restricting tomosynthesis access in resource-constrained settings. It uses a pseudo-3D latent diffusion model with depth attention for inter-slice coherence and cross-attention with mammogram features, denoising all slices jointly conditioned on the input image. A differentiable projection consistency loss, derived from the Beer-Lambert attenuation model, enforces that the synthesized volume reproduces the input mammogram under forward projection, and lesion-aware attention injection preserves diagnostically critical regions. Benchmarked on T-SYNTH (9,000 paired DM/DBT images), MammoTomo improves image quality and projection fidelity over GAN and independent-slice diffusion baselines, training a DBT lesion detector to near-real-DBT performance (AUC = 0.821 vs. oracle 0.849). Ablations confirm depth attention, physics consistency, and lesion guidance contribute complementary gains, reaching 96.7% of oracle performance. The findings establish a foundation for low-resource, scanner-free breast cancer screening.