3D cGAN-based deformation-aware synthesis of prostate TRUS from MR imaging

Transrectal ultrasound (TRUS) is widely used in prostate interventions due to its real-time imaging capability, but suffers from limited contrast and deformation artifacts, which affect the accuracy of registration with preoperative MR images. To address this, we propose a deformation-aware 3D conditional generative adversarial network (cGAN) to synthesize TRUS-like images directly from MR. Unlike previous approaches, the proposed method generates TRUS-like images within the MR coordinate frame, capturing realistic deformation patterns relevant for MR-TRUS registration. We introduce a model that combines a VNet-based generator, a 3D PatchGAN discriminator, a novel background anchoring strategy, and a weakly supervised loss function using anatomical masks. The proposed method achieves good image and structural similarity (SSIM of 0.772 ± 0.084 DSC of 0.922 ± 0.026) between generated and ground truth images. This approach supports MR-only workflows, virtual TRUS simulation, and improved multimodal registration without requiring real intraoperative TRUS.