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

• DOI: 10.1371/journal.pone.0236378
• Publication Date: 2020-07-24T17:28:19Z
• Authors: Xiao-Ling Zou; Yong Ren; Ding-Yun Feng; Xu-Qi He; Yue-Fei Guo; Hai-Ling Yang; Xian Li; Jia Fang; Quan Li; Jun-Jie Ye; Lan-Qing Han; Tian-Tuo Zhang
• Keywords: Adult; Aged, 80 and over; Male; China; Science; Hypertension, Pulmonary; Q; R; Middle Aged; Thorax; Mass Chest X-Ray; 3. Good health; 03 medical and health sciences; Deep Learning; 0302 clinical medicine; Medicine; Humans; Mass Screening; Radiographic Image Interpretation, Computer-Assisted; Female; Research Article; Aged; Retrospective Studies

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