Shape regression machine and efficient segmentation of left ventricle endocardium from 2D B-mode echocardiogram

Area of Science:

• Medical image analysis
• Machine learning
• Computational anatomy

Background:

• Segmentation of deformable anatomical structures in medical imaging, such as the left ventricle endocardial wall in echocardiograms, presents significant challenges.
• Current state-of-the-art methods often require exhaustive scanning or complex classification-based detection, limiting efficiency.

Purpose of the Study:

• To introduce a novel machine learning approach, the shape regression machine (SRM), for efficient and accurate segmentation of deformable anatomical structures.
• To address the limitations of existing methods by developing a two-stage regression-based approach for both detection and segmentation.

Main Methods:

• The SRM employs a two-stage process: rigid shape estimation for object detection and nonrigid shape estimation for detailed segmentation.
• Both stages utilize a novel image-based boosting ridge regression (IBRR) method, enabling multivariate, nonlinear modeling and fast evaluation.
• The approach statistically learns interrelations among shape, appearance, and anatomy from an annotated database.

Main Results:

• The SRM achieves efficient segmentation by learning statistical relationships between shape, appearance, and anatomy.
• The first stage provides a regression-based object detection solution, requiring fewer scans than classification-based methods while maintaining comparable accuracy.
• The second stage directly associates nonrigid shape with image appearance using a nonlinear regressor.
• Experiments on segmenting the left ventricle endocardium from echocardiograms demonstrated automatic detection and accurate segmentation in approximately 120ms.

Conclusions:

• The shape regression machine (SRM) offers an efficient and effective solution for segmenting deformable anatomical structures in medical images.
• The proposed IBRR method facilitates accurate and rapid analysis of complex image data.
• SRM shows significant potential for clinical applications, particularly in echocardiography, by providing fast and precise left ventricle endocardial border segmentation.