Near Real-Time Quantitative Ultrasound Parametric Imaging for Breast Tissue Characterization via Conditional...

Quantitative ultrasound (QUS) imaging provides noninvasive, safe, and clinically meaningful biomarkers for breast lesion characterization and treatment response assessment. However, its widespread clinical adoption has been limited by the time-consuming and computationally intensive signal processing pipeline required to generate parametric maps. This study investigates the application of generative deep-learning models to synthesize QUS maps in near real time from raw ultrasound radio frequency (RF) data, enabling rapid and scalable tissue characterization. Ultrasound data from 192 patients with suspicious breast lesions (95 benign, 97 malignant) were used to develop and evaluate two conditional generative adversarial networks (GANs) with single-decoder and dual-decoder generators. Both models process multichannel inputs, including the sample RF slice, lesion mask, and a reference phantom slice, to produce five QUS parametric maps. Synthesized maps were evaluated using pixel-level reconstruction metrics, distributional similarity, and performance in a downstream benign-malignant lesion classification task. Both GANs generated high-fidelity maps closely matching their original counterparts, with the dual-decoder model showing modest but consistent improvements across all metrics. Importantly, the deep-learning approach reduced QUS map generation time from approximately 6 min per RF slice using the conventional method to 1.00-1.58 s on CPU and 113-133 ms on GPU. Classifiers trained/evaluated on synthetic maps achieved patient-level accuracy ( $0.84~\pm ~0.06$ ) and area under the curve (AUC $= 0.90~\pm ~0.05$ ) on par with those trained/evaluated on the original maps, demonstrating preserved diagnostic information. The obtained results indicate that generative deep-learning-based QUS parametric map synthesis offers a fast, accurate, and clinically promising alternative to conventional computation, paving the way for real-time, scalable QUS imaging and broader clinical adoption.