A Combined Fully Convolutional Networks and Deformable Model for Automatic Left Ventricle Segmentation Based on 3D...

Area of Science:

• Medical Imaging
• Cardiovascular Ultrasound
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of the left ventricle (LV) from 3D echocardiography (3DE) is crucial for assessing LV function.
• Existing methods may face challenges in precision and automation for clinical diagnosis.

Purpose of the Study:

• To develop a novel, automatic method for precise LV segmentation from 3DE data.
• To improve the accuracy and reliability of LV function assessment in clinical settings.

Main Methods:

• A coarse-to-fine framework combining fully convolutional networks (FCN) and deformable models.
• A deep fusion network with residual modules for initial coarse segmentation.
• Geometrical model initialization and a regularization-enhanced deformable model for fine segmentation, preventing inter-chamber leakage.

Main Results:

• The proposed method achieved superior segmentation accuracy on the challenging CETUS benchmark.
• Demonstrated performance exceeding state-of-the-art methods in LV segmentation.
• Indicated potential for practical clinical application in diagnosing LV function.

Conclusions:

• The developed automatic LV segmentation method offers enhanced accuracy and robustness.
• The coarse-to-fine approach effectively refines segmentation, addressing limitations of prior techniques.
• This method shows significant promise for improving clinical diagnosis of left ventricular function using 3DE.