Deep learning and capsule endoscopy: automatic identification and differentiation of small bowel lesions with distinct haemorrhagic potential using a convolutional neural network

Capsule endoscopy is a major tool for the evaluation of patients with obscure gastrointestinal bleeding. Although this endoscopic modality provides a significant progress in the evaluation of the small bowel, its reading remains an exhaustive process which is error-prone. Our group developed a deep learning model able to detect and differentiate small bowel lesions and automatically classify them according their bleeding potential using a validated score (Saurin’s scale). This tool provides accurate differentiation of lesions according ten distinct categories (normal mucosa, lymphangiectasia, xanthomas, blood and protruding lesions, vascular lesions, and ulcers or erosions with uncertain or high bleeding potential. Overall, our model reached an accuracy of 99%, with a sensitivity of 88% and a specificity of 99%. This work was published as an original article in BMJ Open Gastroenterology.

This is the first artificial intelligence algorithm to simultaneous provide detection and differentiation of the most significant types of small bowel lesions, also classifying the findings according to their bleeding potential. 

The article can be found here.

In Press // September 2021