Development of a Convolutional Neural Network for Detection of Erosions and Ulcers With Distinct Bleeding Potential in Capsule Endoscopy

Background and aims The use of capsule endoscopy (CE) is paramount for the detection of small bowel ulcers and erosions. These lesions are responsible for a significant part of obscure gastrointestinal bleeding cases. The interpretation of CE exams is time-consuming and susceptible to errors. This study aims to develop a convolutional neural network (CNN) model for identification and differentiation of ulcers and erosion with distinct hemorrhagic potential in CE images. Methods A CNN based on CE images was developed. This database included images of normal small intestinal mucosa, mucosal erosions, and ulcers with distinct bleeding potential. The hemorrhagic risk was assessed by the Saurin's classification. For CNN development, 23,720 images were ultimately extracted (18,045 normal mucosa, 1765 mucosal erosions, 1300 images of ulcers with uncertain bleeding potential—P1 ulcers; and 2610 ulcers with high bleeding potential—P2 ulcers. Two image datasets were created for CNN training and validation. Results Overall, the network had a sensitivity of 86.6% and a specificity of 95.9% for detection of ulcers and erosions. Mucosal erosions were detected with a sensitivity and specificity of 73.1% and 96.1%, respectively. P1 ulcers were identified with a sensitivity of 71.5%, and a specificity of 97.8%. P2 ulcers were detected with a sensitivity and specificity of 91.4% and 98.8%, respectively. Conclusion Our algorithm is the first deep learning-based model to accurately detect and distinguish enteric mucosal breaks with different hemorrhagic risk. CNN-assisted CE reading may improve the diagnostic of these lesions and overall CE efficiency.