Correlated Regression Feature Learning for Automated Right Ventricle Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate segmentation of the right ventricle (RV) in cardiac magnetic resonance (MR) images is crucial for quantifying clinical indices like ejection fraction.
• Existing methods may face challenges in robustly and simultaneously determining RV boundary points.

Purpose of the Study:

• To develop and evaluate a novel Regression Convolutional Neural Network (RegressionCNN) for direct and simultaneous segmentation of RV boundary points from cardiac MR images.
• To improve the accuracy and efficiency of RV segmentation for enhanced clinical index quantification.

Main Methods:

• A RegressionCNN was developed, integrating a holistic regression model with a convolutional neural network (CNN).
• The fully connected layers of the CNN served as the holistic regression model, utilizing feature maps from convolutional layers converted to a 1-D vector.
• This approach optimizes convolutional layers to directly learn the holistic regression model, minimizing feature extraction and regression learning mismatches.

Main Results:

• The RegressionCNN demonstrated high correlation with manual RV segmentation, achieving an average boundary correlation coefficient of 0.9827.
• The method showed strong consistency with manual delineations, evidenced by an average Dice metric of 0.8351.
• Performance was evaluated on cardiac MR images from 145 human subjects across two clinical centers.

Conclusions:

• RegressionCNN offers an effective and automated approach for accurate RV segmentation from cardiac MR images.
• The method's ability to directly learn holistic regression models enhances feature learning for segmentation tasks.
• This technique holds potential for improving the robustness of clinical index quantification in cardiology.