AI-enabled RF data synthesis for breast ultrasound: efficacy in quantitative ultrasound tissue characterization

Quantitative ultrasound (QUS) methods can derive insightful biomarkers from raw radiofrequency (RF) signals for tissue characterization and monitoring, but their clinical adoption is limited by the inaccessibility and storage burden of RF data. This study is the first to investigate the potential of deep generative models in synthesizing RF data from standard B-mode images and evaluate their efficacy in downstream QUS analysis. Three conditional generative adversarial networks (cGAN), namely Pix2Pix, a shallow ViT‐based cGAN, and a deep ViT‐based cGAN, were adapted and trained on a large paired dataset of RF/B‐mode frames (21,174 training, 3,456 validation, 8,919 test frames) collected from 152 patients (98 patients in the training, 16 in validation, and 38 in the test set) with suspicious breast lesions. The synthesized RF data were assessed using sample-level evaluation metrics, and via a benign-malignant lesion classification task based on the corresponding QUS features. The generative models achieved a structural similarity index measure (SSIM) of 0.82 ± 0.05 on the synthetic RF data and an average peak signal-to-noise ratio (PSNR) of about 33 dB on the corresponding B-mode images, confirming strong reconstruction fidelity. In the lesion classification experiments, a classifier trained on a selected subset of six QUS features derived from the original RF data achieved a test accuracy of 82 ± 6%. In training and testing the classifier with the same subset of QUS features derived from the synthetic RF data, the deep ViT cGAN matched the original model’s performance (accuracy = 82 ± 6%), outperforming the Pix2Pix and shallow ViT cGANs. When the feature selection and classifier training and testing were exclusively performed on the synthetic QUS parameters, the Deep ViT cGAN (accuracy = 81 ± 7%) and Pix2Pix cGAN (accuracy = 81 ± 6%) demonstrated competitive performance, while the Shallow ViT remained slightly lower (accuracy = 79 ± 6%). The promising results obtained in this study demonstrate the feasibility of RF data synthesis from B‐mode images, and therefore, is a step forward towards QUS‐based tissue characterization without the necessity of direct access to RF data.