U-shape-based network for left ventricular segmentation in echocardiograms with contrastive pretraining.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Cardiovascular diseases are a leading cause of death, necessitating accurate cardiac function assessment.
• Manual delineation of the left ventricle on echocardiograms is time-consuming and subjective.
• Current deep learning models often prioritize accuracy over computational efficiency for clinical use.

Purpose of the Study:

• To develop a computationally efficient deep learning model for left ventricle segmentation in echocardiograms.
• To reduce the computational complexity and parameter count of existing segmentation models.
• To improve the performance of left ventricle segmentation while maintaining low resource requirements.

Main Methods:

• Proposed a novel model combining SwiftFormer Encoder and U-Lite Decoder.
• Incorporated Spatial and Channel reconstruction Convolution (SCConv) module.
• Replaced Binary Cross Entropy Loss (BCELoss) with Polynomial Loss (PolyLoss).

Main Results:

• Achieved a Dice similarity coefficient of 0.92714 for left ventricle segmentation on the EchoNet-Dynamic dataset.
• Reported low computational complexity with 4472.55 M FLOPs and 28.96 M Parameters.
• Demonstrated competitive segmentation performance at a significantly reduced computational cost.

Conclusions:

• The proposed model offers an efficient and accurate solution for left ventricle segmentation in echocardiography.
• The integration of SCConv and PolyLoss enhances segmentation performance.
• This approach addresses the need for high-performance computing in clinical cardiac imaging applications.