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

Cardiovascular diseases, characterized by high morbidity, disability, and mortality rates, are a collective term for disorders affecting the heart's structure or function. In clinical practice, physicians often manually delineate the left ventricular border on echocardiograms to obtain critical physiological parameters such as left ventricular volume and ejection fraction, which are essential for accurate cardiac function assessment. However, most state-of-the-art models focus excessively on pushing the boundaries of segmentation accuracy at the expense of computational complexity, overlooking the substantial demand for high-performance computing resources required for model inference in clinical applications. This paper introduces a novel left ventricle echocardiographic segmentation model that efficiently combines the SwiftFormer Encoder and U-Lite Decoder to reduce network parameter count and computational complexity. Additionally, we incorporate the Spatial and Channel reconstruction Convolution (SCConv) module through spatial and channel reconstruction during downsampling and replace the Binary Cross Entropy Loss (BCELoss) with Polynomial Loss (PolyLoss) to achieve superior segmentation performance. On the EchoNet-Dynamic dataset, our network achieves a Dice similarity coefficient of 0.92714 for left ventricle segmentation, with FLoating-point Operations Per Second (FLOPs) and Parameters of just 4472.55 M and 28.96 M respectively. Extensive experimental results on the EchoNet-Dynamic dataset demonstrate that the proposed modifications deliver competitive performance at a lower computational cost.