TRSA-Net: Task Relation Spatial Co-Attention for Joint Segmentation, Quantification and Uncertainty Estimation on...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Cardiovascular Ultrasound Analysis

Background:

• Accurate segmentation and quantification of the Left Ventricle (LV) from 2D echocardiography are crucial for clinical cardiac assessment.
• Current methods often overlook uncertainty estimation, hindering a comprehensive understanding of model performance in clinical settings.
• There is a need for unified optimization methods that jointly address segmentation, quantification, and uncertainty estimation in echocardiography.

Purpose of the Study:

• To propose a novel multitask model, TRSA-Net, for simultaneous segmentation, quantification, and uncertainty estimation of the Left Ventricle (LV) from paired 2D echocardiograms.
• To leverage spatial correlations between tasks for improved joint learning.
• To enhance the accuracy and reliability of cardiac assessment through integrated model optimization.

Main Methods:

• Development of TRSA-Net, a multitask model incorporating Task Relation Spatial co-Attention (TRSA) for joint learning.
• TRSA mechanism utilizes non-local and co-excitation to learn spatial mappings among task-specific features, integrating spatial information into segmentation and quantification.
• Integration of Boundary-aware Structure Consistency (BSC) and Joint Indices Constraint (JIC) into the multitask learning objective to guide segmentation and quantification.

Main Results:

• TRSA-Net demonstrated competitive performance in segmentation tasks on the public CAMUS dataset.
• The model achieved highly accurate results in Left Ventricle (LV) quantification.
• Ablation experiments confirmed the efficacy of the proposed TRSA-Net architecture and its components.

Conclusions:

• TRSA-Net offers a unified approach for joint segmentation, quantification, and uncertainty estimation in 2D echocardiography.
• The proposed model enhances the accuracy and clinical utility of cardiac assessment from echocardiographic data.
• The novel TRSA mechanism and integrated constraints (BSC, JIC) effectively improve multitask learning for cardiac image analysis.