AI-Based Evaluation of Multiple Cardiac Diseases using 3D CNN and Apical 4-Chamber Echocardiography

The heart is a vital organ responsible for pumping blood full of oxygen in all the tissues and districts of the body. The cardiac cycle is made up of the alternation of two phases: systole (contraction) and diastole (relaxation). A healthy heart is crucial for the general well-being and heart diseases are the first cause of death in Western countries. Among the most common cardiovascular diseases is heart failure, an impairment in the blood pumping function that prevents the delivery of a sufficient cardiac output. There are more types of heart failure, one of which is the systolic failure, characterized by a reduced ejection fraction (EF) and arises when the left ventricle (LV) loses the ability to contract properly. The LV is crucial in the circulatory system, so its impairment can lead to compensatory mechanisms to increase blood flow that can result in heart failure. Echocardiography, due to its safety, availability, and high temporal resolution, it’s the recommended method for diagnosing and prognosticating most heart diseases. However, it is highly operator-dependent, resulting in high variability between operators. Also, there are several techniques for measuring EF, which increases the high variability in diagnosis. It is therefore useful to develop an automated objective diagnostic system that can assist physicians in interpreting echocardiography and diagnosing reduced left ventricular function (RLVF). AI is an instrument that can allow accurate and repeatable diagnosis and deep learning techniques have shown that they can make effective improvements in automatic recognition of cardiac structures, assessment of cardiac function and diagnosis of heart diseases. Deep learning is a branch of machine learning that uses multi-layered neural networks to automatically extract relevant features from input data. By employing AI, physicians can use a wider array of information leading to more informed while faster diagnosis. The aim of this work is to use deep learning to automatically detect RLVF from echocardiographic videos using a 3D convolutional neural network. First, we built our dataset from raw, unstructured data. A database containing all phenotypic parameters and echocardiographic measurements of the patients was automatically created. Also, unstructured clinical reports were translated into numerical vectors reporting values for 32 labels. Then, all echocardiograms were labeled with the type of view they belonged to, using a convolutional neural network to classify them. After that, we were able to select 180 patients to form the dataset to be used for the classification of RLVF and the temporal validation of the resulting model. We developed a classifier based on R(2+1)D, which accepts video as input and only provides video-level labels as supervision. The data structure created is suitable for any machine-learning task. Our model detects cases of RLVF on echocardiographic videos with an accuracy of 91.3% and a F1-score of 90.0%. Lastly, we also modified the network to obtain a multi-classifier for both RLVF and aortic insufficiency with consistent results. In conclusion, this study showed how the use of a 3D CNN was effective in identifying the previously mentioned cardiac diseases from echocardiographic videos using the A4C view, and this confirms that AI may be a useful tool to support clinician decisions on the detection of RLVF and aortic insufficiency pathology, paving the way for early diagnosis.