Abdominal ultrasound simulation drops the patient CT at inference: semantic labels drive physics-based images, and editing the map controls deformation and lesions
Synopsis
Abdominal ultrasound simulation has largely relied on patient CT for ray-casting, which limits control over deformation and pathology. A two-stage pipeline first synthesizes a simplified physics-based ultrasound image from anatomical semantic labels, then refines it into a realistic scan with anatomy-guided unpaired translation, so no patient-specific CT volume is needed at inference. In stage one the semantic diffusion model beat Pix2Pix on morphological metrics while Pix2Pix scored better on perceptual ones; training still requires CT-derived segmentations and ray-casting simulations.
Interpretation
The inference step of abdominal ultrasound simulation can run without patient CT volumes, producing ultrasound images with physics-based grounding from anatomical semantic labels alone, with deformation and pathology controlled by editing the label map. Earlier simulation methods typically required CT-based anatomical references for ray-casting, which limited how far deformation and pathology variability could be controlled. The two-stage pipeline was evaluated on the public VISCERAL Anatomy3 and IRCAD data, comparing Pix2Pix and the semantic diffusion model in stage one and adding segmentation-guided CycleGAN refinement in stage two.
The pipeline works in two steps: semantic labels are mapped to a simplified physics-based ultrasound image, which is then translated into a realistic scan. Stage one learns CT ray-casting outputs with paired generative models, and stage two performs anatomy-guided unpaired translation for style refinement. The semantic diffusion model led on morphological metrics, with MAE 19.85 versus 21.65, SSIM 0.28 versus 0.24 and mIoU 0.43 versus 0.29; Pix2Pix was better on perceptual metrics, with LPIPS 0.17 versus 0.19, FID 0.32 versus 0.37 and KID 0.25 versus 0.48.
Skipping patient CT at inference comes at the cost of CT during training. CT-derived segmentations and ray-casting simulations remain necessary inputs for training stage one. This condition applies to the trained pipeline, which can then generate healthy and pathological simulations by modifying semantic maps.
Perspective
The demonstrated scope is abdominal ultrasound: training and evaluation use the public VISCERAL Anatomy3 and IRCAD data, inference needs only semantic labels, and training still requires CT-derived segmentations and ray-casting simulations. It targets simulation and training settings that need controllable healthy and pathological ultrasound images.
The two stage-one models each lead somewhere: the semantic diffusion model on morphological metrics and Pix2Pix on perceptual metrics, with no single reported choice between them. The complete source code cannot be released due to project restrictions; representative inputs, test samples and preprocessing information are provided instead.