ConUDiff: diffusion model with contrastive pretraining and uncertain region optimization for segmentation of left ventricle from echocardiography

Accurate segmentation of the left ventricle (LV) in echocardiograms plays a crucial role in the diagnosis and treatment of cardiovascular diseases. However, manual segmentation of the left ventricle is time-consuming and subject to inter-observer variability. It is crucial to develop an accurate and automatic segmentation method. In this paper, we propose a novel diffusion-based model, called ConUDiff in short, for LV segmentation in echocardiography. The proposed ConUDiff is based on the denoising diffusion probabilistic model and two modules are introduced, i.e., a contrastive pretrained ResNet-50 encoder and an uncertain region optimization module (UROM). The contrastive pretrained ResNet-50 encoder is employed to extract rich feature representations from the original image and enhance the semantic information contained in the feature maps. The UROM module is designed to optimize uncertain regions in the feature maps. We evaluate our method on two public datasets, i.e., the EchoNet-Dynamic dataset and the EchoNet-Pediatric dataset. The experimental results demonstrate that the proposed ConUDiff outperforms some popular networks, achieving a Dice score of 92.68% on the EchoNet-Dynamic dataset and a Dice score of 90.69% on the EchoNet-Pediatric dataset. Our method shows the potential for echocardiographic left ventricle segmentation.