Artificial Intelligence-assisted Confocal Laser Endomicroscopy For Grading The Inflammation Degree Of Ulcerative Colitis

Background and aimsPersistent inflammatory activity in ulcerative colitis(UC)increases the risk of disease progression and poor prognosis.Probe-based confocal laser endomicroscopy(pCLE)is a novel high-resolution endoscopic technique that can accurately assess the inflammatory activity of UC,but its clinical applicability is limited by a long learning curve and difficulty in image interpretation.Artificial intelligence(AI)can improve the operability and efficiency of pCLE.The aim of this study was to develop an AI-assisted pCLE model for grading the degree of UC inflammation.MethodsThis study reviewed the endoscopy center database of Qilu Hospital,Shandong University,the pCLE images of UC patients collected from January 4,2013 to February 4,2022 were standardally labeled by pCLE specialists based on the previously established CLE grading criteria of UC inflammation degree(four classifications of A-D and two classifications of A/B and C/D)in our research center,which were used to train a convolutional neural network(CNN)-based AI model to grade the degree of UC inflammation.The pCLE images and videos of UC patients who satisfied the inclusion criteria from February 5,2022 to October 25,2022 were prospectively collected to further test the sensitivity,specificity,accuracy,and area under the receiver operating characteristic curve(AUROC)of the AI model for graded diagnosing the degree of UC inflammation and to assess the graded diagnostic performance of the model.ResultsIn this study,a total of 18,431 pCLE images from 215 UC patients were retrospectively collected for model training,and 9,761 pCLE images from 19 UC patients and 263 pCLE offline videos from 30 UC patients were prospectively collected for model testing.In the prospective image testing set,applying the four classification criteria,the sensitivities of the AI model for correctly predicting Grades A,B,C,and D of UC inflammation degree were 97.38%,84.56%,86.22%,and 97.74%;the specificities were 99.25%,99.36%,99.90%,and 97.60%,respectively;accuracies were 99.04%,98.49%,98.68%,and 97.62%;overall accuracy was 95.97%.Applying the two classification criteria,the sensitivity,specificity,and accuracy of the AI model for correctly predicting Grade A/B were 94.70%,99.32%,and 98.53%;the sensitivity,specificity,and accuracy for predicting Grade C/D were 97.04%,98.82%,and 98.34%,respectively;with an overall accuracy of 97.39%.In the prospective offline video testing set,applying the four classification criteria,the sensitivities of the AI model for correctly predicting Grades A,B,C,and D of UC inflammation degree were 75.56%,94.00%,92.31%,and 95.24%;the specificities were 98.17%,93.25%,98.00%,and 98.10%,respectively;accuracies were 94.30%,93.54%,97.72%,and 96.96%;overall accuracy was 91.25%.Applying the two classification criteria,the sensitivity,specificity,and accuracy of the AI model for correctly predicting Grade A/B were 100.00%,98.31%,and 99.24%;the sensitivity,specificity,and accuracy for predicting Grade C/D were 98.31%,100.00%,and 99.24%;with an overall accuracy of 99.24%.The AUROC of the prospective testing set was all greater than 0.9600;and the quadratic weighted kappa coefficients,a measure of agreement between the model and the endoscopy experts on grading the degree of UC inflammation,were all above 0.920.ConclusionsIn this study,an AI-assisted pCLE model for grading the degree of UC inflammation was designed and developed based on CNN,and we confirmed the good grading diagnostic performance of both the four-classification and two-classification models for UC inflammation degree.The model had high grading agreement with pCLE specialists and could accurately identify the degree of inflammation in UC.