Comparative studies of deep learning segmentation models for left ventricle segmentation.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Cardiovascular disease is a leading cause of death.
• Left ventricle (LV) segmentation from 2D echocardiography is crucial for heart function analysis and disease detection.
• Accurate LV segmentation is vital for understanding cardiac anatomy and identifying abnormalities.

Purpose of the Study:

• To develop and evaluate automated LV segmentation methods using deep learning.
• To compare the performance of SegNet, Fully Convolutional Network, and Mask R-CNN for LV segmentation.
• To assess the impact of training data size on segmentation accuracy.

Main Methods:

• Implementation of three convolutional neural network architectures: SegNet, Fully Convolutional Network, and Mask R-CNN.
• Generation of an echocardiography image dataset for training and evaluation.
• Performance evaluation using metrics: pixel accuracy, precision, recall, specificity, Jaccard index, and Dice Similarity Coefficient (DSC).

Main Results:

• Mask R-CNN demonstrated superior performance compared to SegNet and Fully Convolutional Network across all evaluation metrics.
• With 4,000 training images, Mask R-CNN achieved a DSC of 92.21%, Jaccard index of 85.55%, and high values for accuracy, recall, precision, and specificity.
• Segmentation performance stabilized with over 4,000 training images.

Conclusions:

• Automated LV segmentation using deep learning, particularly Mask R-CNN, offers a promising approach for clinical applications.
• The Mask R-CNN model provides accurate and reliable LV segmentation, aiding in cardiovascular disease assessment.
• Sufficient training data (over 4,000 images) is essential for achieving stable and high-performance automated segmentation.