OA12.05 Noninvasive CT-Based Image Biopsy System (iBiopsy) for Early Stage Lung Adenocarcinoma

Abstract

CT screening programs frequently detect early stage lung adenocarcinoma. Recent studies show that distinct subtypes of lung adenocarcinoma are associated with different prognosis and suggest that treatment should be tailored to histological subtypes as identified in the new WHO Lung Tumor Classification. To develop this personalized approach, it is important to have reliable tools to diagnose tumors before treatment, preferably non-invasively through image analysis. We have developed a CT-image analysis system (iBiopsy) that uses computerized deep learning and artificial intelligence. To validate the accuracy of a noninvasive CT-based image biopsy system (iBiopsy) in differentiating early stage lung adenocarcinoma subtypes of atypical adenomatous hyperplasia (AAH), adenocaricnoma in situ (AIS), minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC). We retrospectively identified 365 eligible patients from Zhongshan Hopsital Fudan University, diagnosed with AAH, AIS, MIA or IAC by surgical pathological diagnosis. The last high definition CT scan prior to the surgery of the lesion was analyzed using the iBiopsy system, blinded to pathological result. Based on a pulmonary nodule image feature set (PNIFS) in combination with classified pattern models, such as R-SVM, all the pulmonary nodules were classified into four groups. For diagnosis efficacy, area under the curve (AUC) of Precision-Recall score (PRS), receiver operating characteristic (ROC) of a classification model were calculated in each group. 365 patients were included in the analysis. The classification recognition rate of the PNIFS was 80.03%. The average value of PRS is 0.92, the mean of ROC is 0.95, and it is more than 0.80 for the cross validation value. iBiopsy system allows the non-invasive imaged based stratification of pulmonary adenocarcinoma nodules into four groups, from AAH to IAC. Our result suggest that iBiopsy system could ultimate facilitate the diagnosis and precision management of pulmonary nodules.