Left Atrium Chamber Quantification in echocardiography images using Attention based Convolutional Neural Network

Document Type : Research Paper


1 School of Industrial and Systems Engineering, Tarbiat Modares University, Tehran 1411713116, Iran.

2 Tehran Heart Center, Tehran University of Medical Sciences, Tehran, Iran.



Background and Objective; Left atrium is a heart chamber which volume changes has much importance for identifying, controlling and treatment of cardiovascular diseases. In the current methods, left atrium chamber volume (LAV) is estimated from echocardiography images.
Method: For this purpose, the image segmentation and feature extraction tasks have been performed. The accuracy of these methods highly depends on the quality and performance of the method used for image segmentation and the expertise of the specialist. Therefore, left atrium chamber quantification using automatic image analysis methods is necessitated. In this study, a novel automatic approach by combining convolutional neural network with Convolutional Block Attention Module is proposed for left atrium chamber quantification in echocardiography images with an end-to-end fashion without requiring any prior image segmentation. Two different channel and spatial attention modules are embedded in the designed CNN for identifying the key properties of the output feature map and finding important regions for improving the CNN performance.
Results: The proposed model in this study estimates LAV in end-of-systole and end-of-diastole frames with the average R2 of 96.25% and 88.76%, respectively. Our experimental results show that using attention module in CNN architecture improves the performance of CNN for Left atrium chamber quantification with feature extraction focusing on identifying the key properties and discriminating regions.
Conclusion: The proposed method in this study can be used in computer assisted systems (CAD) for automatic chamber quantification with improving the accuracy and speed compared to manual Left atrium chamber quantification.