Artificial intelligence-based rapid on-site cytopathological evaluation for bronchoscopy examinations
Dilbar Ai, Qin Hu, Yencheng Chao, Chi-Cheng Fu, Wei Yuan, Lei Lv, Dexian Ye, Chun Li, Maosong Ye, Yong Zhang, Qunying Hong, Jie Hu, Xiaobo Xu, Longfu Zhang, Qiuli Jiang, Xingxing Wang, Fang Qu, Boyang Wang, Yingyong Hou, Xin Zhang
Abstract
Cytological rapid on-site evaluation (ROSE) is becoming an integral technique for improving the performance of bronchoscopic examinations by confirming specimenadequacy and accuracy in real-time. However, the time- and personnel-consuming nature of ROSE limits its application. We constructed an artificial intelligence (AI)-based ROSE model using deep-learning convolutional neural network (CNN) technique to assist in classifying cytologic whole-slide images (WSIs) as malignant or benign. A total of 627 patients with ROSE slides were enrolled, among whom 374 and 91 patients were included and randomly assigned into training and validation groups, respectively. Another 162 patients were selected as a testing group. The malignant-benign classification results of the test group were compared between cytopathologists' results and AI-based ROSE model results. Actual ROSE reports of the test group given on-site were considered as results of junior cytopathologists; the official cytological diagnostic reports of the test group, which were given without time pressure and with reference to more clinical and pathological information by the senior cytopathologist, were considered as results of the senior cytopathologist. The real-world comprehensive diagnosis was considered as the gold standard. The area under the ROC curve (AUC) achieved 0.9846 in the validation group at patch-level. The accuracy achieved by one senior cytopathologist, two junior cytopathologists and the AI-based ROSE model were 96.90%, 83.30%, and 84.57%, respectively. This AI-based ROSE model may have the potential to support the diagnosis and therapeutic management of patients with respiratory lesions.