P188 Artificial Intelligence and Colon Capsule Endoscopy: Automatic detection of colonic mucosal lesions and blood using a convolutional neural network. (27th May 2021)

• Record Type: Journal Article
• Authors: Afonso, J; Mascarenhas, M; Ribeiro, T; Cardoso, H; Ferreira, J; Andrade, P; Jorge, R; Lopes, S; Macedo, G
• Is Part Of: Journal of Crohn's and colitis. Volume 15(2021)Supplement 1
• Page Start: S257
• Page End: S258
• Publication Date: 2021-05-27
• Subjects: Inflammatory bowel diseases -- Periodicals; 616.344005
• DOI: 10.1093/ecco-jcc/jjab076.315
• Languages: English
• ISSNs: 1873-9946

Abstract

Abstract: Background: Conventional colonoscopy is the standard criterion for the diagnosis and staging of colonic disease, including inflammatory bowel disease (IBD). However, it has the risk of complications, including bleeding and perforation. Colon capsule endoscopy (CCE) is a minimally invasive alternative for patients refusing conventional colonoscopy or for whom the latter is contraindicated. However, reading CCE images is a time-consuming task and prone to diagnostic error. Detection of colonic blood is relevant for the diagnosis and assessment of the activity of IBD, particularly ulcerative colitis. An accurate and early diagnosis is essential as it directs subsequent treatment. Our aim was to develop an Artificial Intelligence (AI) algorithm, based on a multilayer Convolutional Neural Network (CNN), for automatic detection of blood and other significant lesions in the colonic lumen in CCE exams: Methods: A total of 24 CCE exams ( PillCam COLON 2 ® ) from a single centre, performed between 2010–2020, were analysed. From these exams 7640 images (2915 normal mucosa, 3065 blood and 1660 mucosal lesions) were ultimately extracted. Two image datasets were created for CNN training and testing. These images were inserted in a CNN model with transfer of learning. The output provided by the CNN was compared to the classification provided by a consensus of specialists (Figure 1). Performance marks included sensitivity, specificity, positive and negative predictive values (PPV and NPV, respectively), accuracy and area under the receiving operator characteristics curve (AUROC). Results: Blood was detected with a sensitivity and specificity of 97.2% and 99.9%, respectively. The AUROC for detection of blood was 1.00 (Figure 2). Detection of other findings had a sensitivity and specificity of 83.7% and 96.7%, respectively. The overall accuracy of the CNN was of 95.4%. Figure 1 - Output obtained from the application of the convolutional neural network. N: normal mucosa; B - blood or hematic residues; ML – mucosal lesions. Figure 2 – Receiver operating characteristic analyses of the network's performance in the detection of normal mucosa, blood and colon mucosal lesions. AUROC – are under the receiver operating characteristic curve; N – normal colonic mucosa; B – blood; ML – mucosal lesions. Conclusion: We developed a pioneer CNN for CCE which demonstrated high levels of efficiency for the automatic detection of blood and lesions with high clinical significance. The development of tools for automatic detection of these lesions may allow for minimization of diagnostic error and the time spent evaluating these exams.