Estimation of Left Ventricular Ejection Fraction Based on Swin_Uniform

Abstract

A new left ventricular ejection index prediction method was proposed by introducing a block attention mechanism, which helps to reduce overfitting problems and improve the power of the model to process new different samples. Swin Transformer is a Transformer architecture that retains the modeling advantages of self attention mechanism. Its hierarchical design has both global and local modeling capabilities, and can output multi resolution feature maps, which is suitable for EchoCoTr's need to extract multi-scale information from time series medical images. Meanwhile, the resources required of Swin Transformer at runtime are linearly related to images to be processed, which reduces computational complexity and is suitable for processing high-resolution sequence images in EchoCoTr[1]. Swin Transformer has better results than Transformer in both classification and detection tasks, which just goes to show that it is somewhat of a great application in the field of vision. This is beneficial for EchoCoTr's understanding of cardiac ultrasound sequences. Therefore, in the original UniFormer model structure of the EchoCoTr[1] model, a new Swin Transformer Block is added. The new model has stronger modeling capabilities. This means that better results can be obtained under the same computing resources.