Deep learning based classification of left ventricular function from two-dimensional echocardiographic images

Abstract Funding Acknowledgements Type of funding sources: Private company. Main funding source(s): Philips Healthcare Background. Echocardiographic quantification of left ventricular function (LVF) is mainly based on ejection fraction (EF) measurements, which relies on either manual or automated identification of endocardial boundaries followed by calculation of model-based end-systolic and end-diastolic LV volumes. Recent developments in artificial intelligence resulted in computer algorithms that allow fully automated detection of endocardial boundaries and measurement of LV volumes and EF. However, this methodology is prone to errors and inter-measurement variability. We hypothesized that a fully automated deep learning algorithm could be developed, which would accurately classify LVF while avoiding volume and EF measurements. This study was designed to test the accuracy of this approach. Methods. Deep learning algorithm was developed (Philips Research) based on convolutional neural network (CNN) that uses as input dynamic sequences of apical 2- and 4-chamber echocardiographic views obtained without ultrasound enhancing agents. We used for CNN development a database of clinical DICOM studies: a training set of 14,427 studies with normal LV function and 6,135 abnormal, and a validation set of 2,898 normal and 1,081 abnormal studies, based on Philips IntelliSpace Cardiovascular (ISCV) codes found (defined by cardiologists) in the patients’ reports. The CNN was trained to automatically classify LVF into 3 categories: (1) normal, (2) mildly-to-moderately or moderately reduced, and (3) moderately-to-severely or severely reduced. In the validation set, the automated classifications were compared to those in the patients’ reports as a reference standard. Accuracy of the automated classification was tested using contingency tables, from which sensitivity, specificity, and negative and positive predictive values (NPV, PPV) and overall accuracy were calculated for each category of LVF. Additionally, the area under ROC curve (AUC) was calculated to assess the diagnostic accuracy of the automated classification for each LVF category. Results. Automated classification of LVF showed high levels of diagnostic accuracy in identifying cases with LVF in all 3 categories, reflected by high AUC values: (1) 0.94, (2) 0.87 and (3) 0.97 (Figure), and overall accuracy of 0.84 (Table). Conclusions. Deep learning algorithm based on CNN allowed accurate automated classification of LVF, when tested on ∼4,000 clinical studies and compared to ISCV codes found in the patients’ reports. This novel fully-automated methodology may become a useful aid in the interpretation of echocardiographic images by providing the reader with a preliminary assessment of LVF. Abstract Figure.