TY - GEN
T1 - Temporally Consistent Segmentations from Sparsely Labeled Echocardiograms Using Image Registration for Pseudo-labels Generation
AU - Tafuro, Matteo
AU - Jansen, Gino
AU - Išgum, Ivana
N1 - Publisher Copyright: © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2023
Y1 - 2023
N2 - 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.
AB - 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.
KW - Echocardiography
KW - Image registration
KW - Left ventricle segmentation
KW - Pseudo-labels
UR - http://www.scopus.com/inward/record.url?scp=85174730198&partnerID=8YFLogxK
U2 - https://doi.org/10.1007/978-3-031-44521-7_19
DO - https://doi.org/10.1007/978-3-031-44521-7_19
M3 - Conference contribution
SN - 9783031445200
VL - 14337 LNCS
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 195
EP - 204
BT - Simplifying Medical Ultrasound - 4th International Workshop, ASMUS 2023, Held in Conjunction with MICCAI 2023, Proceedings
A2 - Kainz, Bernhard
A2 - Müller, Johanna Paula
A2 - Noble, Alison
A2 - Schnabel, Julia
A2 - Khanal, Bishesh
A2 - Day, Thomas
PB - Springer Science and Business Media Deutschland GmbH
T2 - 4th International Workshop of Advances in Simplifying Medical Ultrasound, ASMUS 2023
Y2 - 8 October 2023 through 8 October 2023
ER -