Dual-discriminator Generative Adversarial Network (DD-GAN) for Virtual Elastography Ultrasound in Thyroid Cancer Diagnosis.

RATIONALE AND OBJECTIVES: Ultrasound elastography (EUS) provides tissue stiffness information to improve diagnostic specificity, yet its application is limited by operator dependence, nonreproducibility, and insufficient hardware support in portable devices. This study aimed to develop a dual-discriminator generative adversarial network-based image synthesis model to investigate the feasibility of directly generating virtual elastography ultrasound (V-EUS) images from conventional B-mode ultrasound images, and to evaluate its diagnostic value. MATERIALS AND METHODS: The consistency between V-EUS and real EUS images was quantitatively assessed using strain ratio (SR), with metrics including peak signal-to-noise ratio, structural similarity index, color histogram correlation, and mean absolute percentage error. Clinical utility was evaluated through radiologist blind testing and diagnostic performance analysis, particularly when V-EUS was integrated with thyroid imaging reporting and data system (TI-RADS) classification. RESULTS: Results showed that the diagnostic performance for benign and malignant nodules based on V-EUS SR was not significantly different from that based on real EUS (internal test set area under the curve [AUC]: 0.744 vs. 0.774, p = 0.318; external test set AUC: 0.742 vs. 0.759, p = 0.461; prospective validation set AUC: 0.751 vs. 0.761, p = 0.772). In radiologist blind tests, V-EUS and real EUS were indistinguishable. Incorporating the Tsukuba score derived from V-EUS into the TI-RADS system significantly improved diagnostic AUC for both junior and senior radiologists. CONCLUSION: This technology provides an ancillary solution to overcome the operator- and hardware-related limitations of traditional elastography, facilitating the widespread adoption of multimodal ultrasound assessment for thyroid nodules.