Skip to Main content Skip to Navigation
Conference papers

Réseau de neurones récurrent à attention pour la détection de lésions intestinales

Abstract : Crohn's disease (CD) is a chronic inflammatory bowel disease , affecting young subjects, causing mucosal damage to the small intestine : erosions, ulcerations, edema and stenosis. The wireless capsule endosopy (WCE) is the best examination for their detection. The WCE generates approximately 50,000 images that are time-consuming for gastroenterologists to analyse. The objective of our work is therefore to develop a tool for the automatic recognition of mucosal lesions of CD in the small intestine. The algorithm is based on a network of convolutional neurons with attention. This was trained on a public database, GIANA, containing images of normal WCEs and with inflammatory and vascular lesions. Another database was used separately , CROHN-IPI, consisting of normal images and MC le-sions annotated by gastroenterologists from Nantes University Hospital. Preliminary results show that the algorithm trained on the 1800 GIANA images, is able to detect with an accuracy of 99,77% the pathological images. Concerning CROHN-IPI, the accuracy obtained on the 2500 images from 39 patients is 80.36%. This difference can be explained by the way the images in the database were selected (images of more obvious lesions in GIANA) or by the under-representation of certain pathologies in CROHN-IPI. In the future, a larger scale WCE annotation application will be developed to enrich CROHN-IPI.
Complete list of metadatas

Cited literature [20 references]  Display  Hide  Download
Contributor : Rémi Vallee <>
Submitted on : Wednesday, September 11, 2019 - 10:43:36 AM
Last modification on : Wednesday, June 24, 2020 - 4:19:41 PM
Long-term archiving on: : Friday, February 7, 2020 - 6:24:32 PM


Files produced by the author(s)


  • HAL Id : hal-02283753, version 1


Rémi Vallée, A de Maissin, Harold Mouchère, A Boureille, Antoine Coutrot, et al.. Réseau de neurones récurrent à attention pour la détection de lésions intestinales. ORASIS, May 2019, Saint-Dié-des-Vosges, France. ⟨hal-02283753⟩



Record views


Files downloads