Research On Subtypes And Severity Classification Of Inflammatory Bowel Disease Based On Deep Learning

Inflammatory bowel disease(IBD)is a group of chronic non-specific gastrointestinal inflammatory diseases with unclear etiology,which mainly includes two subtypes: Ulcerative colitis(UC)and Crohn’s disease(CD).In the diagnosis,treatment and follow-up monitoring of IBD,the endoscopists identifies UC and CD by colonoscopy and assesses their severity in order to determine effective follow-up treatment strategies and clinical care programs.However,such identification and evaluation mode relies heavily on the subjective experience and level of endoscopists,which increases the work burden of endoscopists,and long-term review fatigue will also affect the accuracy of diagnosis.The uneven distribution of medical resources in China and the lack of experienced endoscopists in primary hospitals or remote areas pose great challenges for accurate diagnosis and assessment of such diseases.Therefore,the purpose of this study is to develop a deep learning-based auxiliary diagnostic tool for the identification of UC and CD,and to grade their severity.In the differential diagnosis of IBD subtypes and the severity assessment based on deep learning,the following difficulties exist.(1)The lesions of UC and CD display the characteristics of spatial diffusion in static colonoscopy images and are staggered with normal tissues with unclear boundaries.In addition,the similarity between UC and CD in colonoscopy and the difference within the class aggravate the difficulty of differential diagnosis.The attention mechanism is a dynamic selection process that adaptively weights features according to the importance of inputs,which helps suppress background noise and focus on areas of difference.(2)UC severity assessment is a fine-grained classification,and subclassification often requires more detailed and comprehensive feature extraction.The gradual shooting features make the colonoscopy images present a variety of presentation.Many shooting points not only contain the cross-sectional features of the intestinal cavity,but also include the features of the intestinal cavity depth.Remote relationship modeling and local feature capture simultaneously are helpful in extracting effective identification information.(3)In the fine-grained classification task of CD severity assessment,the characteristics of transmural inflammation and ulceration induced by CD are not easily reflected by colonoscopy images,which are different from those of UC,which mainly focus on mucosal lesions.Computed-tomography enterography can better reflect the inflammatory degree of intestinal wall,which is beneficial to judge the activity of CD.Due to the lack of color information in CT images,the representation of higher-order information can make up for the limitations of first-order information modeling and help to capture the subtle differences between subclasses.To solve the above problems,the following studies were mainly conducted in this paper:(1)In view of the interclass similarities and intraclass differences between UC and CD,and the spatial diffuseness of the lesions in static colonoscopy images,spatial bilinear attention network was designed to recognize UC and CD based on ResNet50.The spatial attention module was applied to select spatial focal areas adaptively,the invalid information was suppressed and the most important local differential features were focused.The bilinear attention module was introduced to capture second-order statistics and establish the interdependence between the local and global second-order information,so as to realize effective feature screening and enhance the recognition ability of discriminant features.The experimental results displayed that adding attention mechanism can promote the classification performance of the model and the overall accuracy was 92.67%,0.73% higher than that of the baseline model.(2)Aiming at the fine-grained classification requirements of UC severity assessment,a two-branch hybrid feature fusion framework was proposed.The framework combined pyramid vision transformer(PVT)and ResNet50,where PVT can learn high resolution representation while having high operation efficiency.The feature fusion module solved the feature fusion problem that ResNet50 and PVT change with the feature resolution pyramid,and effectively combined the local feature with the global representation in an interactive way.Finally,second-order statistics are applied to enrich the discrimination information.The experimental results showed that the framework was effective for improving the classification performance of the model,and the overall accuracy was improved by 2.90% and 1.62%,respectively,compared with the single branch ResNet50 and PVT.(3)According to the characteristics of CD transmural inflammation and ulceration and the intestinal mucosal information provided by endoscopic images is not comprehensive,a hybrid high-order asymmetric convolutional network was proposed based on CT enterography(CTE)images to evaluate the severity.The intestinal wall part in CTE images was extracted as the input of the neural network,in order to capture features such as intestinal wall thickness,strengthening mode and degree.In the proposed model,the asymmetric convolutional module was used to extract intestinal wall features and the high-order hybrid module was applied to achieve real high-order information characterization,while the classification accuracy was 85.41%,and the performance was improved by 0.76%～4.22%compared with other models.In addition,data augmentation based on generative adversarial network was explored,and experimental results expressed that to same extent,this method can improve the classification performance of the model.In summary,this paper conducted a systematic study on subtypes recognition and severity assessment of IBD based on deep learning algorithm,providing technical support for the development and application of artificial intelligence in the recognition of digestive diseases.