A promising approach for screening pulmonary hypertension based on frontal chest radiographs using deep learning: A retrospective study

Background To date, the missed diagnosis rate of pulmonary hypertension (PH) was high, and there has been limited development of a rapid, simple, and effective way to screen the disease. The purpose of this study is to develop a deep learning approach to achieve rapid detection of possible abnormalities in chest radiographs suggesting PH for screening patients suspected of PH. Methods We retrospectively collected frontal chest radiographs and the pulmonary artery systolic pressure (PASP) value measured by Doppler transthoracic echocardiography from 762 patients (357 healthy controls and 405 with PH) from three institutes in China from January 2013 to May 2019. The wohle sample comprised 762 images (641 for training, 80 for internal test, and 41 for external test). We firstly performed a 8-fold cross-validation on the 641 images selected for training (561 for pre-training, 80 for validation), then decided to tune learning rate to 0.0008 according to the best score on validation data. Finally, we used all the pre-training and validation data (561+80 = 641) to train our models (Resnet50, Xception, and Inception V3), evaluated them on internal and external test dataset to classify the images as having manifestations of PH or healthy according to the area under the receiver operating characteristic curve (AUC/ROC). After that, the three deep learning models were further used for prediction of PASP using regression algorithm. Moreover, we invited an experienced chest radiologist to classify the images in the test dataset as having PH or not, and compared the prediction accuracy performed by deep learing models with that of manual classification. Results The AUC performed by the best model (Inception V3) achieved 0.970 in the internal test, and slightly declined in the external test (0.967) when using deep learning algorithms to classify PH from normal based on chest X-rays. The mean absolute error (MAE) of the best model for prediction of PASP value was smaller in the internal test (7.45) compared to 9.95 in the external test. Manual classification of PH based on chest X-rays showed much lower AUCs compared to that performed by deep learning models both in the internal and external test. Conclusions The present study used deep learning algorithms to classify abnormalities suggesting PH in chest radiographs with high accuracy and good generalizability. Once tested prospectively in clinical settings, the technology could provide a non-invasive and easy-to-use method to screen patients suspected of having PH.