403 A HIGHLY SENSITIVE AND HIGHLY SPECIFIC CONVOLUTIONAL NEURAL NETWORK-BASED ALGORITHM FOR AUTOMATED DIAGNOSIS OF ANGIODYSPLASIA IN SMALL BOWEL CAPSULE ENDOSCOPY

Abstract

Capsule endoscopy (CE) has become a standard non-invasive tool for small bowel (SB) examination. However, with an average number of 50,000 SB still frames per CE video, lesions can be missed and CE reading remains a time-consuming activity. Therefore, the development of computer-aided algorithms for lesions’ detection has become an active research area in CE. Gastro-intestinal angiodysplasias (AGD) are the most common SB vascular lesions with an inherent risk of bleeding. This study aimed to develop a computer-assisted diagnosis (CAD) tool for SB-AGD detection in CE. A French national database (CAD-CAP) was created: 5296 deidentified third generation SB-CE videos (Pillcam® SB3 system, Medtronic) were collected from 13 centers; 5427 pathological still frames were extracted from 1369 SB-CE with at least one pathological findings; 20,000 normal (control) still frames were extracted from 200 normal SB-CE videos. Six hundred still images with typical AGD were selected from the database by expert readers. Six hundred normal still frames were randomly extracted as controls. The 600 selected AGD were annotated on each individual pathological frame, as “ground truth” reference. Two sets of 600 SB-CE still frames (300 normal frames and 300 frames with AGD, per set) were created. Different learning-based algorithms were tested including color-based approaches and deep feature extraction. The first set of still frames was used for the machine learning process and the second bank for the validation process. Among the tested algorithms, the one with the best diagnostic performance characteristics used a semantic segmentation images approach associated with an artificial convolutional neural network (CNN) for deep feature extraction (figure 1 from left to right: original image, annotated image, CNN approach, CNN approach on original image). This algorithm reached a 100% sensitivity, a 95,83% specificity, a 96.15% positive predictive value, and a 100% negative predictive value for AGD detection. Reproducibility was perfect (Kappa = 1.0). This study shows that a computed algorithm based on a CNN has very good reproducibility, diagnostic sensitivity (100%), and specificity (95.83%), for detection of AGD on still frames. This study opens the way to automated SB-CE reading softwares, and calls for prospective evaluations of CNN at the video level.