Improve Detection of Pulmonary Hypertension Using Binary Classification

Pulmonary hypertension(PH) is a type of high blood pressure that affects the arteries in the lungs and the right side of the heart. It is evidenced by an increased pressure in the main pulmonary artery. Doppler echocardiograms have been used to non-invasively detect PH by tracing the images to obtain the maximum velocity of blood flow into the right ventricle. This maximum velocity(Vmax) is used to calculate the right ventricular systolic pressure using the Bernoulli equation which will enable clinicians to determine if a patient has PH. Usually, these images are traced by hand by a clinician which might lead to an overestimation or underestimation of Vmax. In addition to this, the nature of the image obtained from the scan could lead to poor tracing by a clinician. Hence, inaccurate diagnosis of the disease in a given subject. In order to address this issue, a machine could be developed and trained to accurately trace Doppler echocardiograms to obtain Vmax. The first step to this goal will be building an algorithm to classify the images into good or bad images based on previous data. Another algorithm will be used to trace the good images to determine Vmax. This research aimed at classifying images into good and bad images using a convolutional neural network(CNN) and a fully connected neural network(FNN). The models were built, trained and tested on data sets obtained from Stanford hospital. A loss function was plotted for each model to analyze how well it trained. The accuracy of each model in predicting if an image was good or bad was calculated to observe its performance on unfamiliar data. CNN had a great loss function and a 35 percent accuracy as compared to a 90 percent accuracy for the FNN. Results showed that, although FNN had a higher accuracy, it had a poor loss function and performs poorly in extracting features from images. This implies that CNN is most suitable for binary classification of echocardiograms as compared to FNN, holding all other variables constant. For future studies, the CNN obtained in this study, could be optimized by using higher epochs to reduce overfitting and for higher accuracy in predicting which image is good for usage in PH diagnosis.