Left Ventricle Segmentation in Echocardiography with Transformer.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in cardiology
• Deep learning for medical diagnosis

Background:

• Accurate left ventricular ejection fraction (LVEF) assessment is crucial for diagnosing heart disease.
• Manual echocardiogram analysis is prone to human bias and high labor costs.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has advanced echocardiogram segmentation, but limitations exist due to CNNs' receptive fields.

Purpose of the Study:

• To develop and evaluate novel deep learning models for enhanced left ventricular (LV) segmentation in echocardiograms.
• To address the limitations of CNNs in capturing the large LV region in echocardiography.
• To explore the efficacy of Transformer-based architectures for robust LV segmentation.

Main Methods:

• Two pure Transformer-based models were proposed: one combining Swin Transformer and K-Net, and another using Segformer.
• The models were trained and evaluated on the EchoNet-Dynamic dataset for LV segmentation.
• Quantitative metrics were compared against existing CNN-based models.

Main Results:

• The proposed Transformer models achieved high mean Dice similarity scores of 92.92% (Swin Transformer-K-Net) and 92.79% (Segformer).
• Both Transformer models outperformed most previous mainstream CNN models in LV segmentation accuracy.
• The models demonstrated superior ability to differentiate between the left ventricle and left atrium, even in challenging cases where CNNs failed.

Conclusions:

• Transformer-based models show significant potential for accurate and reliable echocardiographic segmentation.
• These models offer an effective alternative to CNNs, particularly for segmenting large cardiac structures like the left ventricle.
• The improved segmentation accuracy facilitates more precise cardiac function assessment and disease diagnosis.