Exploratory thyroid ultrasound–computed tomography synthesis using anatomy-preserving contrastive unpaired translation

Abstract

Abstract Objectives When ultrasound (US) and computed tomography (CT) are both clinically indicated, unpaired image translation may support model development but cannot replace either examination. We evaluated Anatomy-Preserving Contrastive Unpaired Translation (AP-CUT), a CUT-based framework with a CT-source Sobel edge-consistency term for CT → US translation. Methods The study used 832 malignant thyroid US images from TN5000 and 800 cervical CT slices from Head-Neck-CHUM. CycleGAN, CUT, and AP-CUT were compared with a low-dimensional Fréchet proxy and nearest-neighbor SSIM/PSNR. Bounding boxes were used only to prepare local US regions for qualitative US → CT examples. A separate single-run experiment treated 600 CT → US outputs as malignant-class training examples and evaluated three classifiers. Results AP-CUT had the highest CT → US nearest-neighbor SSIM (0.1388), compared with 0.0982 for CUT and 0.1247 for CycleGAN. CycleGAN had the lowest US → CT proxy FID (10.4; lower is better). In the classification experiment, the absolute AUC change was +0.0002 for ResNet-50, +0.0261 for EfficientNet-B4 (0.7297 to 0.7558), and −0.0109 for EfficientNet-B4 + CBAM. Conclusions The CT-source regularizer improved one relative structural proxy over CUT, but absolute structural scores remained low. US → CT outputs are local CT-domain-styled patches rather than reconstructed axial CT images, and the mixed single-run classification results do not establish clinical utility.