P161 Computer aided detection of Crohn's disease small bowel lesions in wireless capsule endoscopy. (16th January 2018)

• Record Type: Journal Article
• Authors: de Maissin, A; Gomez, T; Le Berre, C; Normand, N; Mouchere, H; Trang, C; Bourreille, A
• Is Part Of: Journal of Crohn's and colitis. Volume 12:Number 1(2018:Jan.)Supplement 1
• Page Start: S178
• Page End: S179
• Publication Date: 2018-01-16
• Subjects: Inflammatory bowel diseases -- Periodicals; 616.344005
• DOI: 10.1093/ecco-jcc/jjx180.288
• Languages: English
• ISSNs: 1873-9946

Abstract

Abstract: Background: Wireless capsule endoscopy (WCE) is the most efficient exam to detect small bowel (SB) mucosal lesions of Crohn's disease (CD). Unfortunately, videos reading is time consuming. The aim of this study was to develop a computer aided model able to detect CD lesions in SB using WCE. Methods: Forty-five pathologic videos corresponding to 35 patients have been selected among the 250 WCE performed between 2013 and 2017 in patients with a known CD. Three-hundred-sixty-seven pathologic frames have been annotated as follow: aphtoïd ulceration (AU), ulceration 3–10 mm (U3), ulceration >10 mm (U10), oedema (E), stenosis (S) and fistula (F). Several transformations were applied to each pathologic frame (rotation (R), symmetry (S), elastic transformation (ET)) up to increase the number of images (24 times, N = 17664). The complete original dataset was composed of 736 images, 367 positive and 369 negative (randomly extracted from the same videos than positive frames). They were then used to train our model, a convolutional neural network. The whole 736 original images were randomly split into three groups: 80% for the training phase, 10% for the validation phase and 10% for the test phase. The training phase was performed 20-times with random split of data to get a robust 20 folds cross-validation. Results: Sensitivity, specificity, positive predictive value and accuracy for the detection of all types of lesions together were 62.18%, 66.81%, 66.85%, 64.63% respectively. The sensitivity for the detection of all types of ulcerations was 56.84% ( N = 277), 87.29% for stenosis ( N = 50), 56.35% for oedema ( N = 29) and 84.44% for pseudopolype ( N = 11). None fistula was identified in our dataset. Performances of our model have been increased using several images transformations together (rotation, symmetry, and elastic deformation): without data augmentation sensitivity was 33.9% and specificity was 67.02%. The application of rotations alone increased the accuracy to 54.62%. Conclusions: The study demonstrated the feasibility of a computer aided model for automatic detection of SB lesions in patients with CD using WCE. The improvement of the model depends mostly to the number of positive and negative frames of the original dataset.