A self-supervised semi-supervised echocardiographic video left ventricle segmentation method

Tianxiang Wang, Qun Dai

Abstract: We propose a novel method for semi-supervised video object segmentation, termed FPM4S-VOS, which is designed to segment the left ventricle from echocardiography videos without the need for annotated training data. During our experiments, we observed that the interference caused by speckle noise in ultrasound imaging is not negligible. Therefore, we first considered denoising the video data. Inspired by the outstanding performance of fractional-order partial differential equation-based denoising models in eliminating speckle noise, we have enhanced the diffusion equation. Our improvement integrates the Laplacian of Gaussian theory with Gaussian curvature to preserve edge information, resulting in a new denoising model. Moreover, due to the unique characteristics of echocardiographic data and the limitations of traditional self-supervised learning-based models in handling proxy tasks, standard proxy tasks often fail. To address this, we have developed a preprocessing algorithm based on mathematical morphology and binary segmentation techniques. This algorithm generates binary images (proxy labels/masks) for each echocardiogram, effectively replacing the need for optical flow images or subchannel images in the proxy task of conventional self-supervised networks. On the other hand, to prevent our network from treating the proxy mask as the ground truth, we incorporate the similarity between the original image and its proxy mask into the smooth l1<math><msub is="true"><mi is="true">l</mi><mn is="true">1</mn></msub></math> objective function. We then validate the effectiveness of our proposed self-supervised semi-supervised echocardiographic video left ventricle segmentation model using the EchoNet-Dynamic echocardiographic video dataset. Extensive experiments and objective evaluations demonstrate its superior ability in predicting ejection fraction and segmenting the left ventricle.