AI-Assisted MRE for Intestinal Fibrosis in Crohn's Disease

Crohn Disease (CD)

MRE, Artificial Intelligence, Crohn Disease, Fibrosis

Intestinal fibrotic strictures represent a severe complication of Crohn's disease (CD), affecting over half of the patients. Despite the continuous emergence of novel medications, effective treatment options remain scarce. Endoscopy fails to identify the full-thickness fibrosis of the bowel wall, and standardized assessment for cross-sectional imaging has yet to be established. Previous studies have demonstrated that radiomics models based on computed tomography and deep learning models exhibit commendable diagnostic capability. Thus, this project seeks to conduct a prospective multicenter study, with plans to recruit 234 CD patients requiring bowel resection from five medical centers. The aim is to develop and validate a deep learning model based on magnetic resonance enterography (MRE) to accurately characterize intestinal fibrosis.

Quality Assurance Plan The registry implements a comprehensive quality assurance (QA) plan to validate data and maintain protocol adherence. This includes routine site monitoring, regular audits, and verification of data consistency. Sites participating in the registry are periodically reviewed for compliance with the established operational standards. Data Entry and Summary Process Management will enforce access regulations to ensure only authorized personnel can enter or query data. Patient information meeting inclusion criteria will be entered into a Tencent form from the hospital's medical record system. Research assistants will supplement this form with details about intestinal surgical specimens, including condition, quantity, and storage, and summarize all specimens. Researchers will summarize the MRE imaging data for the relevant patients. No one may delete, alter, copy, print, or output confidential data without management's consent. Verification System During patient enrollment, information collection, and specimen collection, two or more research assistants or researchers will confirm the process. Relevant information will be verified again during specimen collection, labeling, and storage. In the analysis phase, researchers will recheck the accuracy of imaging, patient information, and specimens. Management will conduct a random audit every three months to verify patient inclusion criteria and confirm specimen information accuracy. Data Dictionary A comprehensive data dictionary is used to define each variable collected within the registry. It includes the source of the variable, coding standards and any relevant normal ranges for clinical measures. This data dictionary serves as a reference to ensure uniformity in data collection and analysis. Standard Operating Procedures (SOPs) The registry follows established Standard Operating Procedures (SOPs) for various registry functions, including patient recruitment, data collection, management, and analysis. SOPs also cover reporting procedures for adverse events, including guidelines for data reporting and event classification. Change management processes are in place to address any amendments or updates to registry protocols. Sample Size Assessment A statistical sample size calculation has been performed to ensure that the registry is adequately powered to detect meaningful differences or effects. This calculation takes into account the expected incidence of the event of interest, anticipated variability, and the desired statistical power. The required number of participants or participant years is specified based on the primary and secondary objectives of the study. Plan for Missing Data The registry has a clear policy for handling missing data, including cases where data may be unavailable, uninterpretable, or missing due to inconsistencies (e.g., out-of-range results). A specific protocol is followed for imputing missing values or excluding incomplete data from analysis, ensuring the final dataset remains reliable and valid for statistical analysis. Statistical Analysis Methods Automatic recognition and segmentation of intestinal lesions in images, based on multi-parametric MRI data and artificial intelligence models, are used to evaluate intestinal fibrosis and assist in clinical decision-making. Specifically, the process includes: performing VOI annotation to generate 3D VOI; normalizing and resampling MRE images, cropping voxel intensity and applying min-max normalization; decomposing each 3D MRE lesion image into patches, and applying 5-fold data augmentation as input to the network; developing a deep learning segmentation algorithm using the nnU-Net model for automatic recognition of intestinal lesion images, with performance evaluated using the Dice similarity coefficient; constructing a ResNet model to accurately assess different degrees of intestinal fibrosis, with output as a predicted probability between 0 and 1; collecting multi-parametric MRI data prior to model construction and extracting features not affected by intestinal inflammation; excluding relevant features during model development, retaining only those reflecting intestinal fibrosis; after model construction, grouping patients based on inflammation severity and re-evaluating the AI model's recognition capability. Through these steps and the integration of multi-omics data, molecular subtyping and related prognostic analysis of patients are achieved to assist in clinical treatment decision-making.

No related papers found yet.

None listed

China