Segmentation of heart chambers in 2-D heart ultrasounds with deep learning

Echocardiography is a non-invasive image diagnosis technique where ultrasound waves are used to obtain an image or sequence of the structure and function of the heart. The segmentation of the heart chambers on ultrasound images is a task usually performed by experienced cardiologists, in which they delineate and extract the shape of both atriums and ventricles to obtain important indexes of a patient’s heart condition. However, this task is usually hard to perform accurately due to the poor image quality caused by the equipment and techniques used and due to the variability across different patients and pathologies. Therefore, medical image processing is needed in this particular case to avoid inaccuracy and obtain proper results. Over the last decade, several studies have proved that deep learning techniques are a possible solution to this problem, obtaining good results in automatic segmentation. The major problem with deep learning techniques in medical image processing is the lack of available data to train and test these architectures. In this work we have trained, validated, and tested a convolutional neural network based on the architecture of U-Net for 2D echocardiogram chamber segmentation. The data used for the training of the convolutional neural network was the B-Mode 4-chamber apical view Echogan dataset with data augmentation techniques applied. The novelty of this work is the hyperparameter and architecture optimizations to reduce the computation time while obtaining significant training and testing accuraciesObjectius de Desenvolupament Sostenible::3 - Salut i Benesta