| Background: Inflammatory Bowel Disease(IBD)mainly includes Ulcerative Colitis(UC)and Crohn’s Disease(CD).Both UC and CD are chronic inflammatory diseases of digestive tract.Intestinal Tuberculosis(ITB)is an intestinal infectious disease caused by bacillus tuberculosis,and its also a chronic inflammation of the digestive tract.The clinical symptoms and endoscopy manifestations of UC,CD and ITB are very similar,the diagnosis usually should considered many situations,like clinical manifestations,laboratory examination,endoscopy,and histopathology results.Both UC and CD are lack of "gold standard" for diagnosis,and the detect rate of tuberculosis infection is extreme low.Therefore,the diagnosis and differential diagnosis of these three diseases were a challenge in clinical practice.In recent years,Artificial Intelligence(AI)technology has made great achievements in medical related fields.There are also a lot of researches focus on AI and digestive system diseases.However,most of the current researches on AI and digestive system diseases are mainly focused on the field of neoplastic or precancerous lesions.Although some studies have focused on the application of AI in IBD,they are mainly limited to the assessment of disease conditions and prognosis prediction,and few studies focus on the differential diagnosis between IBD and ITB.Furthermore,most current researches focuses on the performance and effects of AI models.But there is not enough explanation for the AI model’s interpretability.Objective:The purpose of this study are using colonoscopy images and clinical data to develop Deep learning(DL)and Machine learning(ML)model for differential diagnosis among UC,CD and ITB,and to explain the prediction principle of AI model.To further compare the performance of different AI models and verify the clinical application of AI model.Method:1.Colonoscopy images of UC,CD and ITB were collected,DL methods were used to develop the DL model for differential diagnosis of UC,CD and ITB.The Grad-CAM method was used to locate the lesions predicted by the DL model.In order to explain the DL model,the feature information of the colonoscopy images with the highest predicted value were analyzed.2.Clinical data of UC,CD and ITB were collected,ML methods were used to develop the ML model for differential diagnosis of UC,CD and ITB.SHAP and LIME methods were used to analyze and explain the prediction features of the ML models.3.The performance and consistence of DL model and ML model were compared by the colonoscopy images and clinical data matched data-set.4.A prospective trial was conducted to verify the performance of the ML model in the identification of CD and ITB,and to compare the performance and consistence of ML model and MDT.Result:1.DL model(1)Differentiation diagnosis:(1)Xception performed best in the identification of UC and ITB by using colonoscopy images,the AUC and accuracy were 0.819 and 0.849.(2)Xception performed best in the identification of CD and ITB by using colonoscopy images,the AUC and accuracy were 0.761 and 0.857.(3)VGG19 performed best in the identification of UC and CD by using colonoscopy images,the AUC and accuracy were 0.864 and 0.791.(2)Grad-CAM: The Grad-CAM method could locate the regions of lesions,the interested areas of DL model,precisely.(3)DL model predictive features:(1)DL model(UC vs ITB): The frequency of mucosal fragility and bleeding in UC images were higher then ITB images(60.0% vs 13.3%,P=0.021),and the frequency of ring ulcer in ITB images were higher then UC images(40.0% vs 0.0%,P=0.017).(2)DL model(CD vs ITB): The frequency of longitudinal ulcers and cobblestone in CD images were higher then ITB images(53.3% vs 0.0%,P=0.002 and 73.3% vs 0.0%,P<0.001),and the frequency of scars in ITB images were higher then CD images(33.3% vs0.0%,P=0.042).(3)DL model(CD vs UC): The frequency of longitudinal ulcers and cobblestone in CD images were higher then UC images(53.3%vs 0.0%,P=0.002 and 73.3% vs 0.0%,P<0.001).2.ML model(1)Differentiation diagnosis:(1)XGBoost performed best in the identification of UC and ITB by using clinical data,the AUC and accuracy were 0.981 and 0.936.(2)XGBoost performed best in the identification of CD and ITB by using clinical data,the AUC and accuracy were 0.946 and 0.884.(3)XGBoost performed best in the identification of UC and CD by using clinical data,the AUC and accuracy were 0.963 and 0.891.(2)ML model predictive features:(1)The top 3 features of ML model(UC vs ITB)were T-spot,rectal involved,and OB.(2)The top 3features of ML model(CD vs ITB)were T-spot,tuberculosis,and age of onset.(3)The top 3 features of ML model(CD vs UC)were terminal ileum involved,age of onset,and granulation hyperplasia.3.DL model vs ML model(1)Differentiation diagnosis(in colonoscopy images and clinical data matched data-set):(1)The accuracy,sensitivity and specificity of CD and ITB differential diagnosis DL model were 88.1%,95.8%,and 74.6%.(2)The accuracy,sensitivity and specificity of CD and ITB differential diagnosis ML model were 89.2%,89.0%,and 89.6%.(2)Contrast experiment:(1)The ML model performed better than DL model in specificity(89.6% vs 74.6%,P=0.024).(2)The agreement rate between DL model and ML model was 88.1%,and the Kappa coefficient was 0.724(P < 0.001).(3)The fusion model of ML and DL model:The accuracy,sensitivity and specificity of CD and ITB differential diagnosis fusion model were82.70%、81.25%,and 73.13%.4.Prospective trial result(1)Differentiation diagnosis(in the prospective trial data-set):(1)The accuracy,sensitivity and specificity of ML were 86.0%,83.3% and 87.1%.(2)The accuracy,sensitivity and specificity of MDT were 95.3%,91.7%,and 96.8%.(2)Contrast experiment: The agreement rate between MDT and ML model was 90.7%,and the Kappa coefficient was 0.780(P < 0.001).Conclusion:1.The DL models and ML models for differential diagnosis of UC,CD and ITB all have good performance and interpretability.2.The DL model could differential diagnosis between CD and ITB patients by only using colonoscopy.It is suitable for colonoscopist to establish a preliminary differential diagnosis.However,the ML model performed better than DL model in specificity.3.The ML model also has good performance in clinical practice and is highly consistent with the MDT.It is suitable for clinicians to make differential diagnosis between CD and ITB patients based on clinical data.Figures: 60,Tables: 18,References:113... |
PDF Full Text Request