DEVELOPMENT AND VALIDATION OF ARTIFICIAL INTELLIGENCE PROGRAM USING THE STANDARD 8 REGION IMAGING METHOD FOR THE QUALITY CONTROL OF ESOPHAGODUODENOGASTROSCOPY

Aims Esophagogastroduodenoscopy (EGD) plays an important role in diagnosis and treatment of upper gastrointestinal disease. Complications of EGD have been increased, and this highlights the necessity of quality control of endoscopy. Complete visualization and photodocumentation of upper gastrointestinal (UGI) tracts is an important measure to prove the performance of each EGD. Based on recent success of AI (artificial intelligence) application in endoscopic images, we developed an AI-drived quality control system for EGD through convolutional neural network (CNN) using documented endoscopic images.

Methods We labeled the stomach location to eight alphabets according to the ESGE photodocumentation methods. The total number of EGD images was 2592 from 250 cases, 200 complete cases and 50 incomplete cases. We removed unnecessary black pads from the original images, and we resized our data into 224 by 224 for modeling. After image preprocessing, we performed two studies using 26 different networks with 5-fold cross-validation: multi-class classification study of images into 8 locations, and binary classification study to determine whether the EGD procedure was performed without missing any location.

Results We used a ResNet101 model pre-trained with Imagenet data for both classification studies and a few data augmentation methods to improve the data. For the multi-class classificiation, the model we used classified the location with 98% accuracy, 98% positive predictive value, and 97% sensitivity. For the binary classification, our model showed 89% of accuracy. We also used class activation mapping to be more transparent of our study results and to explain how the model works.

Conclusions We were able to classify the images to the correct anatomical locations and evaluate the completeness of EGD study in terms of visualization.