Semi-supervised echocardiography segmentation via cross-center invariant prior

Abstract

Echocardiography is a clinically significant diagnostic tool, and segmenting heart chambers from echocardiograms holds great clinical importance. Semi-supervised learning can effectively reduce the amount of annotated data required for deep learning segmentation models. However, most existing methods tend to overfit on single-center training data, struggling with cross-center generalization. To alleviate this, we propose a novel semi-supervised framework crafted to comprehensively leverage cross-center transferable image prior that each image can be decomposed into complementary low-frequency content details and high-frequency structural characteristics. Specifically, we decompose each image into high-frequency and low-frequency components, input them parallelly into Mamba-UNet, and enforce consistency between their outputs and the output of the original image input into U-Net. This can be regarded as image-level and network-level perturbations. Additionally, we introduce evidential deep learning to further enhance the robustness of the model. More importantly, our consistency regularization promotes consistency in evidence for predictions between the original image and its decomposed frequency components, aiding in learning image feature embeddings and uncertainty that generalize across centers. Experimental results demonstrate the competitiveness of our proposed method.