Generative Adversarial Networks for Contrast-Enhanced Mammography: Translating Low-Energy Scans into Dual-Energy Subtracted Images

This study explores the use of generative models, particularly Generative Adversarial Networks (GANs), for generating Dual-Energy Subtracted (DES) images in the context of Contrast-Enhanced Mammography (CEM) for breast cancer screening. Traditional CEM offers enhanced diagnostic performance but comes with certain risks and drawbacks. To mitigate these risks, we propose novel approaches for generating synthetic digital breast images (DES) that maintain the high contrast necessary for effective mass detection, without relying on contrast agent. We explore two main strategies: the first involves guiding the objective function of the GAN using the Contrast-to-Noise Ratio (CNR) in three different ways: SoftCNRloss, PeakCNRloss, and Adaptive Exponential Contrast Loss (AECloss). These approaches refine the GAN’s loss function to emphasize on the contrast of the masses. The second strategy introduces attention mechanisms into the U-Net architecture (IAMUNet) to improve feature extraction and generation, resulting in contrast-enhanced and high-quality image generation. Among these approaches, IAMUNet_256, a U-Net variant enhanced with Channel, Spatial, and Multi-Scale Channel Attention, demonstrates superior performance in generating high-quality DES images, both quantitatively and qualitatively. These results suggest that the proposed methods hold significant promise for improving diagnostic performance while reducing patient exposure to radiation and contrast agent.