Multi-Task Learning for Motion Analysis and Segmentation in 3D Echocardiography.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate left ventricular (LV) motion estimation is crucial for assessing cardiac function and myocardial injury using 3D+time echocardiography.
• Traditional strain analysis often involves separate segmentation and motion tracking steps, which can be optimized jointly for improved accuracy.

Purpose of the Study:

• To develop a multi-task learning network for simultaneous segmentation and motion tracking of the left ventricle in 3D echocardiography.
• To enhance the accuracy of myocardial strain analysis and clinical parameter estimation.

Main Methods:

• A novel multi-task learning network was designed, integrating two task-specific networks for segmentation and motion tracking.
• Cross-stitch units were employed to learn shared representations between tasks, while a shape-consistency unit ensured coherence between propagated and directly predicted segmentations.
• The model was trained and validated on a combined dataset of synthetic and in-vivo 3D echocardiography data.

Main Results:

• The proposed model achieved excellent estimates of left ventricular motion displacement and myocardial segmentation.
• Image-based strain measurements showed strong correlation with crystal-based measurements and good correspondence with SPECT perfusion mappings.
• Segmentation masks derived from the model accurately estimated ejection fraction and sphericity indices, correlating well with benchmark values.

Conclusions:

• Jointly optimizing segmentation and motion tracking using a multi-task learning network significantly improves left ventricular analysis in 3D echocardiography.
• The developed model demonstrates clinical utility for accurate cardiac function assessment and myocardial injury detection.
• This approach offers a promising advancement for quantitative analysis in cardiac imaging.