Sa1913 THE DETECTION AND DIFFERENTIAL DIAGNOSIS FOR COLORECTAL LESIONS DURING ROUTINE COLONOSCOPY WITH AN ARTIFICIAL INTELLIGENCE ASSISTANCE.

We developed a novel artificial intelligence algorithm to assist the detection and differential diagnosis of colorectal lesions during routine colonoscopy in real time (AIAC). A framework using convolutional neural network, one of deep learning approaches was applied to establish AIAC. In this study, the detection rate and accuracy values of histological type prediction (neoplastic or non-neoplastic) with AIAC at the current status were evaluated using recorded non-magnifying endoscopic images. In order to prepare a training dataset for AIAC, 7,281 lesions and 29,341 still images (including movie frames) were extracted from the database of colonoscopy of Jikei University Hospital stored since April 2014. 35,238 training data were created by enclosing a diseased area with a bounding box. Annotations of histological diagnosis were then added on each of the boxed areas. In the assessment of the detection rate of AIAC, three types of endoscopic images were used; Test data 1: 3,229 still images regardless of morphology type and size of the lesions, Test data 2: 105 still images containing flat and diminutive lesions, Test data 3: 12 non-edited video clips of total colonoscopy. In order to assess the accuracy of the differential diagnosis, 280 non-magnified still images of the lesions (236 neoplastic lesions, 44 non-neoplastic lesions) were used and the results were compared with the diagnosis made by an endoscopist in real-time. Sensitivity and positive predictive value (PPV) of the lesion detection with AIAC were 98.0% and 91.2% in the test data 1and 93.7% and 96.7% in the test data 2. In the implementation of movie clips with the test data 3, endoscopists detected 34 lesions in 12 cases, whereas AIAC detected 41 lesions including 7 lesions which were undetected by endoscopists. The accuracy values of the differential diagnosis were sensitivity 90.7%, specificity 65.9% and accuracy 86.8%, respectively. The results of this pilot study demonstrated that the detection of colonic lesions would be improved with the AIAC assistance. In addition, differential diagnostic accuracy between neoplastic and non-neoplastic lesion was also comparable with that of endoscopists.