P609 Artificial intelligence in the image recognition of pouchoscopies in patients who have undergone restorative proctocolectomy with ileal pouch anal-anastomosis (IPAA)

Abstract

Abstract Background Restorative proctocolectomy with ileal pouch anastomosis (IPAA) has become the gold standard in the surgical treatment of ulcerative colitis, with pouchitis being the most common late complication of IPAA. Pouchoscopy is the diagnostic tool of choice to examine the pouch in cases of pouchitis. Artificial intelligence (AI)-based image recognition programmes can assist the examiner in making a diagnosis, support in the education of physicians in training and potentially improve patient outcomes. The use of artificial intelligence has not yet been demonstrated in patients with IBD. The aim of this study is the development of an image recognition algorithm based on Convolutional Neural Networks (CNN) that reliably detects the endoscopic findings of pouchitis analogue to the PDAI score in endoscopic images. Methods The dataset consisted of 10 pouchoscopy videos, which were split into individual frames and assessed using a specially developed labelling tool. In the training process, the labelled images were divided into training, validation and test datasets (ratio 60:20:20). Pre-trained networks and hyperparameters were selected by cross-validation. The defined networks were then fine-tuned through tenfold cross-validation and evaluated for accuracy, specificity, precision F1 score and AUC. Results Our dataset consisted of 7130 images, with 1961 images in the „Inflammation“ group, 2326 images in the „Not assessable“ group and 2843 images in the „Healthy“ group. The final model has been cross-validated 10 times and has been shown to be able to correctly assess the endoscopic features of pouchitis to a certain extent. However, its performance is not yet sufficient for clinical use. Conclusion The work has shown that it is possible to develop an image recognition programme using artificial intelligence that is capable of detecting endoscopic features of pouchitis to a certain extent. Our model is not yet able to make a reliable classification, so it is not yet ready for practical use. However, it should be noted that this project was a feasibility study with a small number of cases and a limited patient population. Further research and more data are needed to sufficiently train the model. Once this is done, the model could be further developed to make reliable classifications and be used in clinical practice.