M.Sc Thesis


M.Sc StudentZiselman Yaniv
SubjectDeep-Learning based Automatic Crohn's Disease Related
Biomarker Assessment
DepartmentDepartment of Biomedical Engineering
Supervisor PROF. Moti Freiman


Abstract

Crohn's disease is a chronic inflammatory bowel disease of which prevalence is on the rise in recent years. Even though the condition is not completely understood and there is no cure, there are treatment options that can help patients lead normal lives if they are properly diagnosed before significant damage was caused to their bowel. To decide on the optimal course of treatment for Crohn's patients, physicians must first diagnose them and assess the severity of their disease. Magnetic resonance enterography based indices such as the simplified MARIA are widely used to assess the severity of Crohn’s disease. The manual calculation of these indices by radiologists is both time consuming and subject to bias. In this work, we focus on automatically detecting wall thickening in the terminal ileum. This is one of the steps required in automating the calculation of the simplified MARIA score. We present a Unet-based deep-learning model combined with a Support-Vector-Machine classifier to automatically segment the terminal-ileum and assess its wall-thickening from T2-weighted MRE data. To deal with the challenge of having a small dataset, we explored the use of multiple spatially correlated segmentation targets and the use of the empirical marginal distribution of the segmentation mask as an additional input channel. Since the signal intensity in T2-weighted MRE data can teach us about the structure of the bowel wall, we used the bins of the normalized histogram of the voxels belonging the terminal ileum as features for the classifier. We show that given manual segmentations our system achieves good classification results. We demonstrate the benefit of using the marginal distribution of the segmentations by improving the segmentation results over the baseline and by approaching the performance of the system using the manual segmentations. We evaluated our segmentation model on MRE scans from the ImageKids study using a 3-fold stratified cross-validation and our classifier using a leave-one-out method. Our method achieved a dice score of 0.635 for the segmentation (N=180) and an F1-score (Cohen’s kappa) of 0.933 (0.6) for wall-thickening classification (N=18). Our proposed method can be part of a system that automatically calculates the MARIAs score of Crohn's patients.