Temporally Consistent Segmentations from Sparsely Labeled Echocardiograms Using Image Registration for Pseudo-labels Generation

The segmentation of the left ventricle in echocardiograms is crucial for diagnosing cardiovascular diseases. However, current deep learning methods typically focus on 2D segmentations and overlook the temporal information in ultrasound sequences. This choice might be caused by the scarcity of manual annotations, which are typically limited to end-diastole and end-systole frames. Therefore, we propose a method that trains temporally consistent segmentation models from sparsely labeled echocardiograms. We leverage image registration to generate pseudo-labels for unlabeled frames enabling the training of 3D models. Using a state-of-the-art convolutional neural network, 3D nnU-Net, we delineate the left ventricle (LV) cavity, LV myocardium, and left atrium. Evaluation on the CAMUS dataset demonstrates the quality and robustness of the generated pseudo-labels, serving as effective training data for subsequent segmentation. Additionally, we evaluate the segmentation model both intrinsically, measuring accuracy and temporal consistency, and extrinsically, estimating cardiac function markers like ejection fraction and left ventricular volumes. The results show accurate delineation of the cardiac structures that evolves smoothly over time, effectively demonstrating the model’s accuracy and temporal consistency.