A Study On Automatic Segmentation Method Of Abdominal Adipose Tissue On CT And MRI And Its Application To Inflammatory Bowel Disease

Inflammatory bowel disease is a chronic and progressive immune-mediated disease.Early diagnosis and treatment are crucial for improving patient outcomes.The high heterogeneity of inflammatory bowel disease requires multidimensional clinical information to stratify the severity of the disease and develop individualized treatment plans.Obtaining multidimensional clinical information is extremely costly,therefore it is a pressing issue in clinical practice to obtain highly consistent and accurate quantitative indicator without increasing the cost of the patient’s disease.Adipose tissues,as highly dynamic metabolic and endocrine organ,is closely related to the treatment decision-making and disease management of inflammatory bowel disease.However,there is currently a lack of automated methods for measuring adipose tissue in different regions.Computed tomography(CT)and magnetic resonance imaging(MRI)are common diagnostic and follow-up imaging techniques for inflammatory bowel disease,and extracting quantitative parameters for adipose tissue and muscle content from images may provide more evidence for clinical diagnosis and evaluation of inflammatory bowel disease.However,the segmentation of abdominal adipose tissue in the inflammatory bowel disease population is limited by the high variability of visceral adipose tissue intensity values and morphological features,as well as the insufficient amount of groundtruth annotations of adipose tissues in different regions,making it difficult to achieve automated quantitative analysis of regional adipose tissue distribution.The dissertation focuses on the necessity and challenges of regional adipose tissue quantification in the inflammatory bowel disease population.It analyzes the intensity characteristics of adipose tissues in CT and MRI images in the inflammatory bowel disease population,and explores the rapid and accurate acquisition of groundtruth annotations in the automated segmentation process,as well as optimizing image preprocessing and postprocessing.It also investigates automated quantitative methods for abdominal adipose tissues in the inflammatory bowel disease population and verifies the role of quantitative parameters in the diagnosis and treatment of inflammatory bowel disease.Firstly,a novel CT-based automatic segmentation pipeline is proposed in this study.The proposed solution employs a semi-automatic data augmentation labeling strategy to significantly reduce manual labeling time and workload,while simultaneously achieving fast and accurate acquisition of regional adipose tissue and muscle labels.The pipeline uses adjacent slices of two-dimensional CT images with high structural similarity as input to the Trans UNet network,allowing for data augmentation without image transformation.By integrating diverse training data sets,mask-based image preprocessing and postprocessing,the resulting model can achieve both high-accuracy abdominal region positioning and automatic segmentation of regional adipose tissue.This solution incorporates core technologies such as fast and accurate acquisition of groundtruth annotations,image data pre-processing and post-processing strategies,abdominal region positioning,and multi-task synchronous learning of adipose tissue and muscles in different regions.Then,a retrospective clinical study was conducted.Multi-dimensional clinical information reflecting the disease state and diagnosis of inflammatory bowel disease was obtained.Combining masks of visceral adipose tissue,subcutaneous adipose tissue,and skeletal muscle with original CT images,various parameters reflecting the content and distribution of adipose tissue were extracted.The ability of regional adipose tissue and muscle features in the discrimination diagnosis of inflammatory bowel disease and the short-term prognosis prediction of Crohn’s disease was confirmed.Secondly,due to the characteristics of no ionizing radiation and the ability to assess the nature of lesions,MRI has a different application scenario than CT in management of inflammatory bowel disease.Different principles of MRI lead to differences in prior information of image intensity values of adipose tissues between MRI sequences and between MRI and CT images.The same tissue has different intensity values in MRI images due to static magnetic field inhomogeneity,thus requiring more manual intervention in the annotation process.In this study,we combine the aforementioned CT segmentation model with transfer learning to achieve high-precision segmentation of three-dimensional abdominal adipose tissue using few-shot annotated MRI images.Results from the prediction experiment of T2 sequences not involved in model training,demonstrated that the transfer learning model can accurately identify intensity features of muscle and subcutaneous adipose tissue in MRI images.Additionally,we propose a segment scan merging algorithm that solves the problem of repeated labeling and inaccurate volume quantification caused by repeated layers in the MRI sequence of inflammatory bowel disease populations divided into multiple three-dimensional segments.Through MRI co-registration sequence data,we achieve a four-fold increase in the amount of input data without additional labeling,effectively improving the performance of three-dimensional automatic adipose tissue segmentation in the abdominal region.Finally,we included individuals from the inflammatory bowel disease population who underwent both CT and MRI scans within a two-week period,and compared volume parameters of adipose tissues in different region and muscle predicted by the CT and MRI automatic segmentation models.Our results demonstrate that the parameters obtained by both models are consistent,supporting their use for longitudinal evaluation of distribution of adipose tissues and muscle in the inflammatory bowel disease population.In conclusion,this study has achieved automatic segmentation approach of abdominal subcutaneous adipose tissue,visceral adipose tissue,and muscle on CT and MRI images of the inflammatory bowel disease population through three-dimensional automatic segmentation methods.The study has also investigated the role of fat and muscle features in discrimination and prognosis evaluation of inflammatory bowel disease.The research can be used to monitor changes in adipose tissue and muscle distribution in the long-term course of inflammatory bowel disease to evaluate the effects of drug intervention and nutritional therapy on the disease.