Improving realism in abdominal ultrasound simulation combining a segmentation-guided loss and polar coordinates...

Ultrasound (US) simulation helps train physicians and medical students in image acquisition and interpretation, enabling safe practice of transducer manipulation and organ identification. Current simulators generate realistic images from reference scans. Although physics-based simulators provide real-time images, they lack sufficient realism, while recent deep learning-based models based on unpaired image-to-image translation improve realism but introduce anatomical inconsistencies.

We propose a novel framework to reduce hallucinations from generative adversarial networks (GANs) used on physics-based simulations, enhancing anatomical accuracy and realism in abdominal US simulation. Our method aims to produce anatomically consistent images free from artifacts within and outside the field of view (FoV).

We introduce a segmentation-guided loss to enforce anatomical consistency by using a pre-trained Unet model that segments abdominal organs from physics-based simulated scans. Penalizing segmentation discrepancies before and after the translation cycle helps prevent unrealistic artifacts. Additionally, we propose training GANs on images in polar coordinates to limit the field of view to non-blank regions. We evaluated our approach on unpaired datasets comprising 617 real abdominal US images from a SonoSite-M turbo v1.3 scanner and 971 artificial scans from a ray-casting simulator. Data was partitioned at the patient level into training (70%), validation (10%), and testing (20%). Performance was quantitatively assessed with Frechet and Kernel Inception Distances (FID and KID), and organ-specific ${\chi }^2$ histogram distances, reporting 95% confidence intervals. We compared our model against generative methods such as CUT, UVCGANv2, and UNSB, performing statistical analyses using Wilcoxon tests (FID and KID with Bonferroni-corrected $\alpha =0.01$ , ${\chi }^2$ with $\alpha =0.008$ ). A perceptual realism study involving expert radiologists was also conducted.

Our method significantly reduced FID and KID by 66% and 89%, respectively, compared to CycleGAN, and by 34% and 59% compared to the leading alternative UVCGANv2 ( $p\ll 0.01$ ). No significant differences ( $p>0.008$ ) in echogenicity distributions were found between real and simulated images within liver and gallbladder regions. The user study indicated our simulated scans fooled radiologists in 36.2% of cases, outperforming other methods.

Our segmentation-guided, polar-coordinates-trained CycleGAN framework significantly reduces hallucinations, ensuring anatomical consistency, and realism in simulated abdominal US images, surpassing existing methods.