Hybrid GAN-Diffusion Framework for 3D Ultrasound to CT Translation using a Self-Critique Generator and Residual Refinement

The translation of 3D ultrasound (US) to computed tomography (CT) is a significant and unsolved challenge in medical imaging due to the vast domain gap and inherent US artifacts. We propose a novel, two-stage hybrid generative framework to address this problem. The first stage introduces a UNetSelfCritique3D generator, a Generative Adversarial Network (GAN) designed to mitigate mode collapse through an internal self-critique loss, producing a structurally coherent but coarse CT volume. The second stage employs a conditional Denoising Diffusion Probabilistic Model (DDPM) for residual refinement, using the GAN output as a strong prior to synthesize high-frequency textural details. We validated our framework on the clinical TRUSTED dataset. Our baseline 3D Pix2Pix model achieved a Peak Signal-to-Noise Ratio (PSNR) of 11.34 dB. The UN-etSelfCritique3D GAN significantly improved performance to 20.57 dB PSNR, and our final hybrid model achieved 21.10 dB PSNR, representing an 86% improvement over the baseline. This work establishes the first performance benchmark for this task using real 3D clinical data and demonstrates the feasibility of the hybrid GAN-Diffusion approach.