Development of a convolutional neural network for the endoscopic classification of pouchitis in patients after restorative proctocolectomy.

BACKGROUND AND AIMS: The aim of this prospective single-center study is to train convolutional neural networks (CNNs) to detect the presence of pouchitis in two-dimensional (2D) images acquired during pouchoscopies and test its feasibility. METHODS: Two separate networks were constructed. The goal of network 1 was to detect whether an inflammation was present. Network 2 was designed to classify endoscopic findings of pouchitis, according to the pouchitis disease activity index (PDAI) score. The dataset was divided into three distinct sets: a training set, a validation set, and a test set. The performance was quantified using a tenfold cross-validation approach. RESULTS: For the detection of inflammation, sensitivity was 71.78% with a specificity of 90.35%. When differentiating the six endoscopic findings according to the PDAI score, the sensitivity ranged from a low of 38% for the 'ulceration' class to a high of 67.18% for the 'friability' class, with a specifity of 94.12% ('ulceration') and 96.57% ('friability'). CONCLUSIONS: This study shows that an artificial, intelligence-based image recognition software can be trained to recognize the endoscopic features of pouchitis with reasonable accuracy. The results, although encouraging, confirm that artificial intelligence (AI) performance in this context remains below human expert level. A larger dataset, human benchmarking and more appropriate endoscopic markers are required to reach clinically relevant performance. Trial registration This trial was registered in the 'ClinicalTrials.gov' database on 26 April 2021 (NCT04864587).