Masahiro Yakami MD, Presenter: Nothing to Disclose

Takeshi Kubo MD, Abstract Co-Author: Grant, Toshiba Corporation

Koichi Ishizu MD, PhD, Abstract Co-Author: Nothing to Disclose

Koji Fujimoto MD, Abstract Co-Author: Nothing to Disclose

Yoshio Iizuka, Abstract Co-Author: Nothing to Disclose

Masami Kawagishi, Abstract Co-Author: Nothing to Disclose

Kaori Togashi MD, Abstract Co-Author: Research funded, Nihon Medi-Physics Co, Ltd Research funded, Eisai Co, Ltd Research funded, Bayer AG Research funded, DAIICHI SANKYO Group Research funded, Covidien AG Research funded, Toshiba Corporation Research funded, Canon Inc Research funded, FUJIFILM Holdings Corporation

00030490-DMT et al, Abstract Co-Author: Nothing to Disclose

A system was successfully built to infer diagnosis of pulmonary nodules from their features scored by radiologists.

Computer inference engines of radiological diagnosis usually use image features calculated by image processing techniques. We investigated the possibility of computer inference from features evaluated by radiologists.

Consecutive 100 CT examinations were selected which had 1 to 5 pulmonary nodules measuring 5 to 30 mm and the diagnosis of which were clinically confirmed to be primary lung cancer, metastatic pulmonary nodules or other benign nodules. A checklist of features (n=51) on pulmonary nodules were created, which mainly consisted of radiological findings reported to be useful for the diagnosis of pulmonary nodules: other miscellaneous features were added to the list so that readers fully described CT features related to pulmonary nodules. A board-certificated radiologist graded and inputted conspicuity of the features in the list on every nodule in these examinations using dedicated software. The conspicuity was rated on 5-point scales (1 to 5). The radiologist also inputted the diagnosis of each nodule (n=222). The diagnosis was selected from primary lung cancer, metastatic pulmonary nodules or other benign nodules. The confidence in the diagnosis was rated on 5-point scales (1 to 5). Another diagnosis was inferred by a Bayesian network that was trained with the conspicuity of features and the confirmed diagnoses of all the other nodules (Leave-One-Out method). Diagnostic accuracy by the radiologist and that by the Bayesian network were 67.6%, 61.3% respectively. The accuracy by the radiologist decreased in proportion to decrement of his confidence whereas that by Bayesian network did not. If the radiologist adopted the computer inferences for nodules diagnosed with lower confidence (1 to 3), the accuracy was 70.3%.  

Computer inference from radiological features may be useful as a second opinion, especially when radiologists are lacking confidence in their diagnosis.  

Yakami, M, Kubo, T, Ishizu, K, Fujimoto, K, Iizuka, Y, Kawagishi, M, Togashi, K, et al, 0, Computer Inference of Diagnosis from Features Scored by Radiologists: Evaluation with 51 Features of 222 Pulmonary Nodules in Consecutive 100 CT Examinations.  Radiological Society of North America 2009 Scientific Assembly and Annual Meeting, November 29 - December 4, 2009 ,Chicago IL. http://archive.rsna.org/2009/8000477.html