A deep learning approach for contrast-agent-free breast lesion detection and classification using adversarial synthesis of contrast-enhanced mammograms

Contrast-enhanced digital mammography (CEDM) has emerged as a promising complementary imaging modality for breast cancer diagnosis, offering enhanced lesion visualization and improved diagnostic accuracy, particularly for patients with dense breast tissues. However, the reliance of CEDM on contrast agents poses challenges to patient safety and accessibility. To overcome those challenges, this paper introduces a deep learning methodology for improved breast lesion detection and classification. In particular, an image-to-image translation model based on cycle-consistent generative adversarial networks (CycleGAN) is utilized to generate synthetic CEDM (SynCEDM) images from full-field digital mammography in order to enhance visual contrast perception without the need for contrast agents. A new dataset of 3958 pairs of low-energy (LE) and CEDM images was collected from 2908 female subjects to train the CycleGAN model to generate SynCEDM images. Thus, we trained different You-Only-Look-Once (YOLO) architectures on CEDM and SynCEDM images for breast lesion detection and classification. SynCEDM images were generated with a structural similarity index (SSIM) of 0.94 ± 0.02. A YOLO lesion detector trained on original CEDM images led to a 91.34% accuracy, a 90.37% sensitivity, and a 92.06% specificity. In comparison, a detector trained on the SynCEDM images exhibited a comparable accuracy of 91.20%, a marginally higher sensitivity of 91.44%, and a slightly lower specificity of 91.30%. This approach not only aims to mitigate contrast agent risks but also to improve breast cancer detection and characterization using mammography.