P219 An application of artificial intelligence in Inflammatory Bowel Disease detection
Rasmagina, I.(1);Bakulin, I.(1);Mariya, S.(1);
(1)North-Western State Medical University n.a. I.I.Mechnikov, Chair of the Propedeutics of Internal Diseases- Gastroenterology and Dietology n.a. S.M. Riss, Saint-Petersburg, Russian Federation
Inflammatory bowel disease (IBD) is verified on the basis of complex assessment of clinical, laboratory, endoscopic and morphological data. However, due to the lack of a ‘golden’ standard and similarity of the clinical picture, it usually takes more than 1 year to diagnose ulcerative colitis (UC) and more than 2 years for Crohn`s disease (CD). The aim of the study is to optimize the existing methods in diagnosis of IBD with an application of an artificial neural network (ANN).
216 patients with verified IBD (163 – UC, 53 – CD of large bowel) and 34 patients without endoscopic findings underwent colonoscopy in the Peter the Great Clinic of Mechnikov North-Western State Medical University (Saint-Petersburg, Russia). During the endoscopic examination 856 images were obtained: 268 images of “normal mucosa”, 194 ones of CD of large bowel and 394 ones of UC. The study included images of mild, moderate and severe activity of IBD. Mayo endoscopic score was used to assess endoscopic activity of UC. For CD the scoring scales were not provided, the activity of the disease was based on the length of the affected segment, presence of erosions, ulcers and complications such as strictures, fistulae and stenosis. All images were captured in high-resolution with the same endoscope (Olympus) and stored in a jpeg format. Poor-quality images resulting from blurs, defocus and poor air insufflation were excluded. We also created the ANN using the Neural Network Toolbox application from the MATLAB application package. The ANN consisted of 2 models: the first one was multilayer perceptron which determined a presence of pathology, the second was convolutional neural network which differentiated the types of IBD (UC or CD).
All the endoscopic images were downloaded to the created ANN. It was trained 3 times due to the decrease of its accuracy. The average accuracy of the developed model in differentiation of pathology was 77%. This result could be associated with an insufficient number of CD images or ANN training on continuous data set of images with UC and CD features without clustering them to particular patient.
We conclude that the developed ANN shows moderate accuracy in recognition of different forms of IBD. This study has a limitation as the images were taken only with endoscope Olympus. Enlarged sample size and inclusion of morphologic images are needed in order to improve the accuracy of ANN.