Ryo Sakamoto, Presenter: Nothing to Disclose

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

Masami Kawagishi, Abstract Co-Author: Nothing to Disclose

Gakuto Aoyama, Abstract Co-Author: Nothing to Disclose

Takeshi Kubo MD, 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

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

Yoshio Iizuka, Abstract Co-Author: Nothing to Disclose

Yutaka Emoto MD, PhD, Abstract Co-Author: Nothing to Disclose

Hiroyuki Sekiguchi, Abstract Co-Author: Nothing to Disclose

Koji Sakai, Abstract Co-Author: Nothing to Disclose

Hiroyuki Yamamoto, Abstract Co-Author: Nothing to Disclose

Inter-observer variability among radiologists may lead to inappropriate clinical recommendations. With appropriate reasoning, CADx may help reducing variability among radiologists’ diagnosis. The aim of this study was to compare variability of radiologists’ interpretations of lung nodules on CT between with and without using a CADx, which is capable of presenting reasoning for the diagnosis.

With the approval of the institutional review board, we built a database of 491 lung nodules consisted of primary lung cancers, metastases and benign nodules. It includes image features scored by two board-certified radiologists and clinical data sets. We employed a Bayesian network for the inference engine of the CADx system and trained it by 179 nodule data sets. Our CADx system infers the diagnosis of nodule with providing the reasoning. Eleven radiologists with 5 years' experience interpreted 312 nodules under three different conditions: without CADx (NCAD); with inference only (ICAD); with presenting reasoning (RCAD). The level of likelihood for each diagnostic category was recorded in percentages up to 100 %. Inter-observer variability was assessed and compared among three different conditions (NCAD, ICAD and RCAD) by using following evaluation criterion: 1) Multi-rater κ (the level of agreement for diagnosis); 2) Standard deviation of AUC for the diagnosis in ROC analyses (variation in diagnostic accuracy); 3) The variance in radiologists’ output (degree of diagnostic consensus for each nodule).

Multi-rater κ was moderate (κ=0.561, 95%C.I.; 0.558, 0.564) by NCAD and was improved to good agreement by ICAD (κ=0.679, 95%C.I.; 0.676, 0.682) and RCAD (κ=0.692, 95%C.I.; 0.689, 0.694). The variation of accuracy was reduced with RCAD compared to ICAD as well as NCAD. The degree of radiologists’ consensus was also improved significantly by using both ICAD and RCAD, but there was no significant difference between ICAD and RCAD.

CADx reduced the radiologists' variability in interpreting lung nodules. With presenting reasoning, CADx was more effective in the aspect of improving the level of agreement for the diagnosis and variation in diagnostic accuracy.

Sakamoto, R, Fujimoto, K, Kawagishi, M, Aoyama, G, Kubo, T, Togashi, K, Yakami, M, Iizuka, Y, Emoto, Y, Sekiguchi, H, Sakai, K, Yamamoto, H, Can Computer-aided Diagnosis (CADx) System that Presents Reasoning Reduce Radiologists’ Inter-observer Variability?: Evaluation in Interpreting Lung Nodules on Computed Tomography.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13016988.html