Automated detection of ulcers and erosions in capsule endoscopy images using a convolutional neural network

Capsule endoscopy (CE) is an important tool in the management of patients with known or suspected inflammatory bowel disease. Ulcers and erosions of the enteric mucosa are prevalent findings in these patients. They frequently occur together, and their identification in CE is crucial for an accurate evaluation of disease severity. Nevertheless, reviewing CE images is a time-consuming task, and the risk of overlooking lesions is significant. Over the last decade, artificial intelligence (AI) has emerged as a means for overcoming these pitfalls. Of all AI methods, convolutional neural networks (CNN), due to their complex multilayer architecture present the best results in medical image analysis, particularly capsule endoscopy. Therefore, we aimed to develop a CNN for the automatic identification of ulcers and erosions in the small bowel mucosa. A total of 1483 CE exams (PillCam SB3®) performed at a single center between 2015 and 2020 were analysed. From these exams, a total of 6130 frames of the enteric mucosa were obtained, 4233 containing enteric ulcers and erosions, and the remaining containing normal mucosa or other findings. Ulcers and erosions were stratified according to Saurin’s classification for bleeding potential: P1E—erosions with intermediate bleeding risk; P1U—ulcers with intermediate bleeding risk; P2U—ulcers with high bleeding risk. For automatic identification of these lesions, these images were inserted into a CNN model with transfer learning. The pool of images was divided for constitution of training and validation datasets, comprising 80% and 20% of the total number of images, respectively. The output provided by the CNN was compared to the classification provided by a consensus of specialists. After optimizing the neural architecture of the algorithm, our model was able to automatically detect and distinguish ulcers and erosions (any bleeding potential) in the small intestine mucosa with an accuracy of 95.6%, sensitivity of 90.8%, and a specificity of 97.1%. We believe that our study lays the foundation for the development and application of effective AI tools to CE. These techniques should improve diagnostic accuracy and reading efficiency. Graphic abstract Schematic representation of the workflow and summary of the results.