K-Net: Integrate Left Ventricle Segmentation and Direct Quantification of Paired Echo Sequence

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Comprehensive cardiac assessment relies on left ventricle (LV) segmentation and quantification.
• Traditional methods can suffer from cumulative errors and limited single-plane views.
• Integrating segmentation and direct quantification offers a more robust approach.

Purpose of the Study:

• To develop an end-to-end framework for simultaneous LV segmentation and direct quantification from paired apical echocardiography views.
• To overcome limitations of single-view analysis by integrating orthogonal views for stereoscopic cardiac assessment.
• To improve accuracy in anatomical morphology and contractile function assessment.

Main Methods:

• Proposed K-shaped Unified Network (K-Net) integrating segmentation and direct quantification.
• Utilized Attention Junction for heterogeneous task learning between segmentation and quantification.
• Employed Bi-ResLSTMs for spatial-temporal information extraction and an Information Valve for multi-view information exchange.
• Implemented Evolution Loss for comprehensive sequential data learning.

Main Results:

• K-Net achieved high performance in simultaneous LV segmentation and quantification.
• Achieved a Dice coefficient of 91.44% for segmentation.
• Demonstrated a mean absolute error of 2.74mm for major-axis dimension quantification.
• Showcased significant clinical potential for cardiac assessment.

Conclusions:

• K-Net represents a significant advancement in automated cardiac assessment using echocardiography.
• The integrated approach overcomes limitations of traditional methods, offering improved accuracy and comprehensive analysis.
• The framework holds promise for enhancing clinical diagnosis and patient management.