Generalizable fully automated multi-label segmentation of four-chamber view echocardiograms based on deep...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Generalizability of artificial intelligence (AI) platforms for echocardiogram segmentation is a significant clinical translation challenge.
• Current AI methods often struggle to perform reliably across different datasets and patient populations.
• Accurate segmentation of cardiac structures is crucial for quantitative analysis and diagnosis.

Purpose of the Study:

• To develop and validate a novel, generalizable, and efficient fully automatic multi-label segmentation method for four-chamber view echocardiograms.
• To address the generalizability issue in AI-based echocardiogram segmentation using deep fully convolutional networks (FCNs) and adversarial training.
• To introduce generative adversarial networks (GANs) for pixel classification in cardiac imaging, a novel application.

Main Methods:

• Employed deep fully convolutional networks (FCNs) integrated with adversarial training for segmentation.
• Utilized generative adversarial networks (GANs) for pixel classification, a novel approach in cardiac imaging AI.
• Validated the method against manual segmentations (ground-truth) and compared it with a state-of-the-art method on independent datasets (including CAMUS challenge data).

Main Results:

• Achieved high Dice metrics for automatic segmentation of all four chambers: 92.1% (LV), 86.3% (RV), 89.6% (LA), and 91.4% (RA).
• Demonstrated excellent correlation for Left Ventricular (LV) volumes between automatic and manual segmentation (0.94 for end-diastolic volume, 0.93 for end-systolic volume).
• Showed significant improvement over previous FCN-based methods, indicating enhanced generalizability.

Conclusions:

• Generative adversarial networks for pixel classification effectively create generalizable, fully automatic FCN-based networks for echocardiogram segmentation.
• The proposed method demonstrates excellent agreement with reference contours and robust performance even with limited training data.
• This novel approach holds promise for improving the clinical translation of AI in echocardiography.