Impact of attention mechanism on U-Net architecture for the Left Ventricle segmentation

Echocardiography is the most widely used medical imaging technique for diagnosing cardiovascular diseases. Further, delineation of left ventricle (LV) in 2D echocardiographic images by cardiologists is required for systolic function assessment. However, this daily task is tough to do, takes a lot of time and is not precise due to the mediocre quality of these images. Recently, deep learning methods, and more particularly encoder-decoder networks, have been widely proposed to solve these problems and to automate LV segmentation. A precise segmentation is very important for properly estimating clinical parameters that can be used in diagnostic aid and follow-up of cardivascular pathologies. Therefore, we examine the performance of attention mechanism into two U-Net architectures. These two models created from the original U-Net and adapted for the Cardiac Acquisitions for Multistructure Ultrasound Segmentation (CAMUS) dataset present better accuracies for segmenting echocardiographic images. Experiences demonstrate an improvement in performances when applying attention by achieving a Dice coefficient of 0.940 and a Hausdorff distance of 4.11 with attention U-Net 2.