Kenta Hamada1,2, Yoshiro Kawahara2, Takayoshi Tanimoto3, Akimitsu Ohto4, Akira Toda3, Toshiaki Aida5, Yasushi Yamasaki6, Tatsuhiro Gotoda6, Taiji Ogawa6, Makoto Abe6, Shotaro Okanoue6, Kensuke Takei6, Satoru Kikuchi7, Shinji Kuroda7, Toshiyoshi Fujiwara7, Hiroyuki Okada6. 1. Department of Endoscopy, Okayama University Hospital, Okayama, Japan. 2. Department of Practical Gastrointestinal Endoscopy, Okayama University Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama, Japan. 3. Business Strategy Division, Ryobi Systems Co., Ltd., Okayama, Japan. 4. Health Care Company, Ryobi Systems Co., Ltd., Okayama, Japan. 5. Okayama University Graduate School of Interdisciplinary Science and Engineering in Health Systems, Okayama, Japan. 6. Department of Gastroenterology and Hepatology, Okayama University Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama, Japan. 7. Department of Gastroenterological Surgery, Okayama University Graduate School of Medicine, Dentistry and Pharmaceutical Sciences, Okayama, Japan.
Abstract
BACKGROUND AND AIM: Recently, artificial intelligence (AI) has been used in endoscopic examination and is expected to help in endoscopic diagnosis. We evaluated the feasibility of AI using convolutional neural network (CNN) systems for evaluating the depth of invasion of early gastric cancer (EGC), based on endoscopic images. METHODS: This study used a deep CNN model, ResNet152. From patients who underwent treatment for EGC at our hospital between January 2012 and December 2016, we selected 100 consecutive patients with mucosal (M) cancers and 100 consecutive patients with cancers invading the submucosa (SM cancers). A total of 3508 non-magnifying endoscopic images of EGCs, including white-light imaging, linked color imaging, blue laser imaging-bright, and indigo-carmine dye contrast imaging, were included in this study. A total of 2288 images from 132 patients served as the development dataset, and 1220 images from 68 patients served as the testing dataset. Invasion depth was evaluated for each image and lesion. The majority vote was applied to lesion-based evaluation. RESULTS: The sensitivity, specificity, and accuracy for diagnosing M cancer were 84.9% (95% confidence interval [CI] 82.3%-87.5%), 70.7% (95% CI 66.8%-74.6%), and 78.9% (95% CI 76.6%-81.2%), respectively, for image-based evaluation, and 85.3% (95% CI 73.4%-97.2%), 82.4% (95% CI 69.5%-95.2%), and 83.8% (95% CI 75.1%-92.6%), respectively, for lesion-based evaluation. CONCLUSIONS: The application of AI using CNN to evaluate the depth of invasion of EGCs based on endoscopic images is feasible, and it is worth investing more effort to put this new technology into practical use.
BACKGROUND AND AIM: Recently, artificial intelligence (AI) has been used in endoscopic examination and is expected to help in endoscopic diagnosis. We evaluated the feasibility of AI using convolutional neural network (CNN) systems for evaluating the depth of invasion of early gastric cancer (EGC), based on endoscopic images. METHODS: This study used a deep CNN model, ResNet152. From patients who underwent treatment for EGC at our hospital between January 2012 and December 2016, we selected 100 consecutive patients with mucosal (M) cancers and 100 consecutive patients with cancers invading the submucosa (SM cancers). A total of 3508 non-magnifying endoscopic images of EGCs, including white-light imaging, linked color imaging, blue laser imaging-bright, and indigo-carmine dye contrast imaging, were included in this study. A total of 2288 images from 132 patients served as the development dataset, and 1220 images from 68 patients served as the testing dataset. Invasion depth was evaluated for each image and lesion. The majority vote was applied to lesion-based evaluation. RESULTS: The sensitivity, specificity, and accuracy for diagnosing M cancer were 84.9% (95% confidence interval [CI] 82.3%-87.5%), 70.7% (95% CI 66.8%-74.6%), and 78.9% (95% CI 76.6%-81.2%), respectively, for image-based evaluation, and 85.3% (95% CI 73.4%-97.2%), 82.4% (95% CI 69.5%-95.2%), and 83.8% (95% CI 75.1%-92.6%), respectively, for lesion-based evaluation. CONCLUSIONS: The application of AI using CNN to evaluate the depth of invasion of EGCs based on endoscopic images is feasible, and it is worth investing more effort to put this new technology into practical use.