Fig. 1.

Pipeline for generating synthetic ultrasound images from a semantic diffusion model (SDM), to be used in segmentation training and testing on real data. SDM Training: The SDM is trained to transform the noise to the realistic image through an iterative denoising process. SDM inference: The trained SDM is inferenced on augmented labels to generate corresponding realistic synthetic ultrasound images. Segmentation: The set of generated synthetic ultrasound images are used to train a segmentation network. The segmentation performance is tested on real data. x denotes the semantic label map; y_t denotes the noisy image at each time step t.