Program Information
Automatic Chest X-Ray Screening with Convolutional Neural Networks
J Kotoku1,2*, T Hirose1 , S Kumagai2 , A Matsushima1 , K Shiraishi1,2 , N Arai2 , A Haga3 , T Kobayashi1 , (1) Teikyo University, Itabashi-ku, Tokyo, (2) Teikyo University Hospital, Itabashi-ku, Tokyo, (3) The University of Tokyo Hospital, Tokyo,
Presentations
SU-F-702-7 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: 702
Purpose: Chest X-ray screening takes a lot of time and efforts for doctors while it is important to find hidden diseases. We propose a novel application of deep neural network for automatic chest X-ray screening.
Methods: The number of chest X-ray radiograph images was 3177. Those raw images were labeled as “Normal” or “Reexamination needed” by a medical doctor and used for training. The images were registered with a mean image using B-spline algorithm. Following the registration, we subtracted the mean image from the deformed images for normalization. As a prediction model, we constructed a convolutional neural network based on AlexNet for deep learning. Cross entropy was used as a loss function.
Results: The predictive sensitivity and specificity was roughly 90 % and 80% respectively. Cardiomegaly, pleural effusion and mass shadow images were successfully labeled as "Reexamination needed” by this model. There was a tendency to mislabel normal fat men as “Reexamination needed”.
Conclusion: Our network shows a potential for successful clinical use to reduce the screening time.
Funding Support, Disclosures, and Conflict of Interest: Funding: This work was partially supported by a Grant-in-Aid from JSPS (Japan Society for the Promotion of Science) KAKENHI JP Scientific Research (C) Grant Number 15K08703.
Contact Email: