DFCG: A Dual-Frequency Cascade Graph model for semi-supervised ultrasound image segmentation with diffusion model

Semi-supervised semantic segmentation based on deep learning is crucial for ultrasound image analysis. However, the scattering noise of ultrasound images decreases the network performance in segmenting lesions of various shapes, sizes, and locations, undermining efficient global contextual information utilization. Furthermore, obtaining a large number of unlabeled medical ultrasound images remains a significant challenge. We propose a dual-frequency cascade graph model for semi-supervised ultrasound image segmentation with a diffusion model to address these challenges. The framework includes two stages. The generate stage uses the latent diffusion model to generate synthetic medical images, which reduces the burden of data annotation and addresses privacy issues associated with medical data collection. In the Segmentation stage, we combine a Fourier frequency domain space with a multi-scale attention mechanism to reduce the effect of scattering noise. We introduce a graph cascade decoder to capture global contextual information and adaptively weight cancer lesion regions to improve the representation ability. We evaluate the proposed method on two public breast cancer segmentation datasets and a private biliary atresia dataset to compare quantitatively with previous state-of-the-art methods. The experimental results demonstrate that the segmentation accuracy of our method is better than others due to the effective transfer of knowledge of probability distributions from the latent diffusion models to the semantic representation of the medical image. Github