P113 Automatic detection of colonic ulcers and erosions in colon capsule endoscopy images using a convolutional neural network

Ribeiro, T.(1);Mascarenhas Saraiva, M.(2);Afonso, J.(1);Cardoso, H.(1);Ferreira, J.(3);Andrade, P.(1);Jorge, R.(3);Lopes, S.(1);Macedo, G.(1);

(1)Centro Hospitalar Universitário de São João, Department of Gastroenterology, Porto, Portugal;(2)Centro Hospitalar Universitário de São João, Gastroenterology, Porto, Portugal;(3)Faculdade de Engenharia da Universidade do Porto, Mechanical Engineering, Porto, Portugal


Conventional colonoscopy is gold standard for the diagnosis and monitoring of patients with suspected or known inflammatory bowel disease (IBD). Nevertheless, it is a potentially painful procedure with risk of complications, including bleeding and perforation. Colon capsule endoscopy (CCE) is a minimally invasive alternative for patients unwilling to undergo colonoscopy or when it is contraindicated or unfeasible. However, CCE produces thousands of frames and their revision is a time-consuming and monotonous task.
The detection of ulcers and erosions is paramount for the diagnosis and assessment of the activity of IBD. However, these lesions may appear in a very small number of frames, thus increasing the risk of overlooking significant lesions.
Our aim was to develop an Artificial Intelligence (AI) algorithm, based on a multilayer Convolutional Neural Network (CNN) for automatic detection of ulcers and erosions in CCE exams.


A total of 24 CCE exams (PillCam COLON 2®) performed at a single centre between 2010-2020 were analysed. From these exams, 3 635 frames of the colonic mucosa were extracted, 770 containing colonic ulcers or erosions and 2 865 showing normal colonic mucosa. These images were used for construction of training (80% of the frames) and validation (20% of the frames) datasets. For automatic identification of these lesions, these images were inputted in a CNN model with transfer learning. The output provided by the CNN (FigureFigure 1- Convolutional neural network output. N – normal colonic mucosa; U – colonic ulcers. The bars represent the predicted class probability outputted by the algorithm.1) was compared to the classification provided by a consensus of specialists. Performance marks included sensitivity, specificity, positive and negative predictive values, accuracy and area under the receiving operator characteristics curve (AUROC). 


After the optimization of the neural architecture of the CNN, our model automatically detected colonic ulcers and erosions with a sensitivity of 90.3%, specificity of 98.8% and an accuracy of 97.0% (Figure 2). The AUROC was 0.99 (Figure 3). The mean processing time for the validation dataset was 11s (approximately 66 frames/s).
Figure 2 - Confusion matrix and performance marksFigure 3 – Receiver operating characteristic analysis of the network’s performance. AUROC – area under the receiver operating characteristic curve; U – Colonic ulcers.


We developed a CNN model which demonstrated high levels of efficiency for the automatic detection of ulcers and erosions in CCE images. Our study lays the foundations for the development of effective AI tools for application to CCE exams. These systems may enhance the diagnostic accuracy and reading efficiency of CCE, thus expanding the role of minimally-invasive colonic exploration in the management of IBD patients.