DOP80 Automatic detection of ulcers and erosions in PillCam™ Crohn's capsule using a convolutional neural network. (27th May 2021)

• Record Type: Journal Article
• Authors: Saraiva, M; Ribeiro, T; Afonso, J; Cardoso, H; Ferreira, J; Andrade, P; Parente, M; Jorge, R; Lopes, S; Macedo, G
• Is Part Of: Journal of Crohn's and colitis. Volume 15(2021)Supplement 1
• Page Start: S111
• Page End: S112
• Publication Date: 2021-05-27
• Subjects: Inflammatory bowel diseases -- Periodicals; 616.344005
• DOI: 10.1093/ecco-jcc/jjab073.119
• Languages: English
• ISSNs: 1873-9946

Abstract

Abstract: Background: Capsule endoscopy (CE) plays a central role in the management of patients with suspected or known Crohn's disease (CD). It is indicated for the diagnosis, classification, monitoring of the response to treatment, and prognostic prediction. In 2017, PillCam ™ Crohn's Capsule (PCC) was introduced. It has demonstrated greater accuracy in detecting and evaluating the extent of lesions in these patients. However, this new tool produces thousands of images, whose revision is time-consuming and prone to errors, since lesions can be restricted to a small number of images. In the last decade, several Artificial Intelligence (AI) algorithms were developed, and demonstrated potential to mitigate some of the drawbacks of CE. Among AI tools, Convolutional Neural Networks (CNN) display the best performance for imagery analysis. This study aims to develop an AI algorithm based on an CNN for the automatic detection of ulcers and erosions of the small intestine and colon in PCC images. Methods: A total of 8 085 PCC images were extracted from a single tertiary centre between 2017–2020. This pool of images was constituted by 2 855 images depicting ulcers, 1 975 erosions; the remaining with normal enteric and colonic mucosa. For the automatic identification of these findings, this pool of images was split into training and validation datasets. A CNN model with transfer learning using tensorflow and keras tools was constructed. The performance of the network was subsequently assessed in an independent test set. Results: After optimizing the different layers of the CNN, our model was able to detect and distinguish small intestinal or colonic erosions or ulcers with a sensitivity and specificity of 90.0% and 96.0%, respectively. The precision and accuracy of this model were 97.1% and 92.4%, respectively (Figure 1). Particularly, the CNN detected ulcers with a sensitivity of 83% and specificity of 98%, and erosions with sensitivity and specificity of 91% and 93%, respectively. Conclusion: Our group developed, for the first time, a CNN capable of automatically detecting ulcers and erosions of the small intestine and colon in PCC images with high sensitivity and specificity. These findings are extremely important since they pave the way for the development of systems for the automatic detection of clinically significant lesions, optimizing diagnostic performance and efficiency of monitoring CD activity.