Left ventricular segmentation method based on optimized UNet and improved CBAM: ESV and EDV tracking study.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Accurate left ventricular segmentation is crucial for cardiac function assessment.
• Conventional methods like DeepLabv3 face challenges with large model sizes and segmentation precision.
• Automated segmentation requires efficient and accurate deep learning models.

Purpose of the Study:

• To introduce an optimized nested UNet model for automated left ventricular segmentation.
• To improve the accuracy and efficiency of cardiac function assessment using deep learning.
• To address limitations of existing models in terms of size and precision.

Main Methods:

• Utilized the EchoNet-Dynamic dataset with video data and expert annotations.
• Developed a nested UNet architecture with a deeper feature extraction module.
• Integrated CBAM (Attention module) and SimAM (Simple Attention Module) for enhanced feature selectivity.
• Combined binary cross-entropy and Dice loss functions for stable training.

Main Results:

• Achieved a 1.05% increase in the Dice coefficient compared to existing methods.
• Reduced model size to 15% of the original, enhancing computational efficiency.
• Demonstrated significantly improved performance in automated left ventricular segmentation.

Conclusions:

• The optimized nested UNet model offers superior accuracy and efficiency for left ventricular segmentation.
• This approach enhances cardiac function assessment and provides a viable solution for automated clinical diagnosis.
• The model's improvements in accuracy and size offer practical benefits for clinical applications.