O6 Artificial intelligence using convolutional neural networks for detection of early Barrett’s Neoplasia

Abstract

Initial results from THE PAIGE PROJECT Portsmouth’s Project on Artificial Intelligence in Gastrointestinal Endoscopy Introduction Endoscopic detection of early Barrett’s neoplasia remains very challenging, with significant inter-observer variation in identifying and assessing these lesions. Artificial intelligence is proposed to help with computer aided detection in this field and could have significant clinical and cost implications. We aim to develop and validate a deep learning (DL) algorithm using Convolutional Neural Networks (CNN) for detection of Barrett’s neoplasia. Methods We collected 132 high definition white light endoscopy images from 46 lesions of histologically confirmed Barrett’s neoplasia. These images were marked and annotated using specially designed software, and reviewed by two experts on advanced assessment and management of Barrett’s neoplasia. Another 119 images of non dysplastic Barrett’s were collected from 20 patients and used as control. Both dysplastic and non dysplastic images were divided into three datasets and used for training, validation and testing of CNN algorithm. We used SegNet segmentation architecture. Graphic processing unit used was ‘GeForce RTX 2080 Ti. We collected metrics on processing speed, sensitivity, specificity and global accuracy at different score thresholds. Results Image processing speed by the algorithm was 33 ms/image. This is much faster than the average human visual response latency which is estimated at 70–100 ms. The algorithm was able to detect Barrett’s neoplasia with sensitivity of 93%, specificity of 78% and global accuracy of 83% (see figure (1) below for examples of algorithm detection). Conclusions We developed and validated an early AI algorithm with high sensitivity and reasonable specificity when compared with PIVI criteria. The ultra short image processing time would suggest this algorithm may be suitable for real time detection of Barrett’s neoplasia. We will develop this model further for use during real time endoscopy.