Xiangrong Zhou PhD, Presenter: Nothing to Disclose

Huayue Chen, Abstract Co-Author: Nothing to Disclose

Takeshi Hara PhD, Abstract Co-Author: Nothing to Disclose

Ryujiro Yokoyama, Abstract Co-Author: Nothing to Disclose

Masayuki Kanematsu MD, Abstract Co-Author: Nothing to Disclose

Hiroshi Fujita PhD, Abstract Co-Author: Nothing to Disclose

Accurately and efficiently detecting the location of an object of interest (an organ, a lesion, etc.) plays an important role in medical image analysis. Machine-learning has demonstrated potentials for solving object detection problems in 3D CT image analysis. However, machine-learning requires a large number of samples for training and testing to show the real performance. As far as we know, few previous works reported performances on more than 1,000 CT scans and showed the possibility of the proposed algorithm for the localization of all the major organs on 3D CT images.

We proposed a machine-learning approach to accomplish the fast and automatic localization of the major organ regions on 3D CT scans. This approach combines object detections and the majority voting technique to achieve the robust and quick organ localization. We applied this approach to localizing 12 kinds of major organ regions independently on 1,300 torso CT scans. In our experiments, we randomly selected 300 CT scans for training, and then, applied the trained system to localize each of the target regions on the other 1,000 CT scans for the performance testing. The detection results were evaluated subjectively and quantitatively based on the ground truth from the human operators.

Our evaluation based on 1,000 test CT cases showed that the heart location in 983 cases, liver location in 978 cases, stomach location in 952 cases, pancreas location in 943 cases, spleen location in 932 cases, left kidney location in 963 cases, right kidney location in 969 cases, left lung location in 992 cases, right lung location in 988 cases, bladder location in 984 cases, right femur head location in 578 cases, and left femur head location in 551 cases were correct. Those results showed that our approach can accomplish the localization tasks in over 94% CT cases except the femur heads. The failure of the femur head locations was due to the poor performance of the trained results.

An universal approach was demonstrated to localize 12 major organs automatically on 3D CT scans and its efficiency and accuracy were validated by using 1,300 clinical CT scans.

Zhou, X, Chen, H, Hara, T, Yokoyama, R, Kanematsu, M, Fujita, H, Automatic Multi-organ Localizations on 3D CT Images by Using Machine-learning Approach Based on a Large Dataset.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13019104.html