Yutaka Emoto MD, PhD, Presenter: Nothing to Disclose

Masahiro Yakami MD, PhD, Abstract Co-Author: Nothing to Disclose

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

Takeshi Kubo MD, Abstract Co-Author: Nothing to Disclose

Ryo Sakamoto, Abstract Co-Author: Nothing to Disclose

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

Gakuto Aoyama, Abstract Co-Author: Nothing to Disclose

Masami Kawagishi, Abstract Co-Author: Nothing to Disclose

Koji Sakai, Abstract Co-Author: Nothing to Disclose

Hiroyuki Sekiguchi, Abstract Co-Author: Nothing to Disclose

Yoshio Iizuka, Abstract Co-Author: Nothing to Disclose

Hiroyuki Yamamoto, Abstract Co-Author: Nothing to Disclose

Computer Aided Diagnosis (CADx) has been expected to help radiologists. Because a CADx does not always suggest the right diagnosis, a radiologist may not agree the suggestions. If a CADx shows why it suggests the diagnosis, a radiologist can make better decision. We developed a CADx system which discloses the reasons of diagnosis of lung nodules in CT images, and evaluated the accuracy of radiologists influenced by the CADx.

We built a database of 491 lung nodules whose diagnoses were clinically confirmed as primary lung cancer, metastatic nodules or benign nodules. The database consisted of CT images, image findings, clinically confirmed diagnosis, clinical information such as laboratory data and patient history. The image findings were described by consensus of two board-certified radiologists. 179 and 312 nodules in the database were used for training the CADx and for evaluation, respectively. The inference model of the CADx was a Bayesian network, which was constructed using the Markov chain Monte Carlo method with the training data set. The CADx derives a set of inference probabilities of each diagnosis (ICAD). In addition to the result, image findings and/or clinical information are indicated as reason of the inference for each case (RCAD). The reason is derived based on influence degree for the diagnosis with the highest inference probability. 11 radiologists, with 5 years’ experience for diagnostic imaging, interpreted the 312 nodules with three different conditions (with no CAD, ICAD, RCAD) with more than 1 month interval. Mean accuracy rates are 0.714, 0.763, 0.766, 0.74 with no CAD, ICAD, RCAD, CAD alone, respectively.  

Radiologists are grouped into 2 groups by the average accuracy rate with no CAD. In the high accuracy group, 2 radiologists are better with ICAD than with RCAD, 2 show no change. In the low accuracy group (LA), 2 are better with ICAD than with RCAD, 5 are better with RCAD than with ICAD. RCAD improves LA better than ICAD.

A CADx which discloses the reason of diagnosis may be effective for radiologist with low accuracy rate of lung nodule diagnosis.

Emoto, Y, Yakami, M, Fujimoto, K, Kubo, T, Sakamoto, R, Togashi, K, Aoyama, G, Kawagishi, M, Sakai, K, Sekiguchi, H, Iizuka, Y, Yamamoto, H, A Computer Aided Diagnosis (CADx) which Discloses the Reason of Diagnosis May Improve a Low Accuracy Group of Radiologist More than a CADx without Reason Disclosure.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13016989.html