ParticipantsLinlin Qi, Beijing, China (Presenter) Nothing to Disclose
Jianwei Wang, MD, Beijing, China (Abstract Co-Author) Nothing to Disclose
Zhen Zhou, Beijing, China (Abstract Co-Author) Nothing to Disclose
Ning Wu, MD, Beijing, China (Abstract Co-Author) Nothing to Disclose
Shichao Feng, Beijing, China (Abstract Co-Author) Nothing to Disclose
Xiaojuan Xu, Beijing, China (Abstract Co-Author) Nothing to Disclose
Wei Tang, MD, Beijing, China (Abstract Co-Author) Nothing to Disclose
Ming Lei, Beijing, China (Abstract Co-Author) Nothing to Disclose
qdqilinlin@163.com
PURPOSETo validate a novel DeepWise computer-aided detection (CAD) system for automated detection of pulmonary nodules.
METHOD AND MATERIALSThe DeepWise CAD system designed by means of a specialized deep neural network is a novel and more intelligent CAD system to detect pulmonary nodules automatically. A public data set LIDC-IDRI and an in-house data set (a total of about 7,000 nodules) were used as the development data set. One hundred consecutive low-dose CT (LDCT) scans in a screening program and 100 specified and matched nodules in another 60 LDCT scans were independently evaluated by two radiologists (Radiologist 1, 2) and the DeepWise CAD system to identify nodules larger than or equal to 2 mm in average diameter. All the nodules detected by both the radiologists and the system were reviewed jointly by another two chest radiologists, who were experienced in LDCT lung cancer screening, and a "true" nodule count was determined. The performance of the two radiologists and the DeepWise CAD system were compared.
RESULTSRadiologist 1, 2 and the DeepWise CAD system detected 193, 115 and 271 nodules, respectively. Of the 325 separate nodules detected by the three techniques, 282 were classified as true nodules on consensus review. Of the true nodules present, the detection date of the Radiologist 1, 2 and the DeepWise CAD system were 66.3% (187/282), 40.8% (115/282) and 83.0% (234/282), respectively. And the 48 nodules missed by the DeepWise CAD included 41 solid nodules with an average diameter of less than 5 mm and 7 ground glass nodules. 187 (96.9%) of 193 Radiologist 1-detected nodules were true nodules, all 115 Radiologist 2-detected nodules were true nodules, and 234 (86.3%) of 271 of the DeepWise CAD-detected nodules were true nodules. The DeepWise CAD system identified 37 lesions that on consensus review were false-positive nodules, a rate of 0.23 (37/160) per patient.
CONCLUSIONThe novel DeepWise CAD system detected 83.0% of true nodules, which was significantly superior to radiologists. And its false positive rate, 0.23 per patient, was significantly cut down.
CLINICAL RELEVANCE/APPLICATIONWith the popularity of MDCT, the detection rate of lung nodules was significantly improved, which brought great pressure to radiologists. And the false positive nodules detected by traditional CAD software were so much that the clinical application was limited. So we need a more intelligent CAD system to detect pulmonary nodules accurately.