Image registration for assessment of Crohn's disease severity

Abstract

Crohn’s disease (CD) is a chronic inflammatory bowel disease (IBD) that affects millions of people in Europe alone. It is important to accurately assess the disease severity in a safe and non-invasive manner in order to improve the treatment of patients with CD. Furthermore, the ideal assessment must be objective, reproducible, quantitative and comprehensive. In clinical practice, ileocolonoscopy is the standard technique used for assessing CD activity, for instance by means of the Crohn’s Disease Endoscopic Index of Severity (CDEIS). However, ileocolonoscopy is invasive and the scoring is not comprehensive. To overcome these drawbacks, we investigated the use of magnetic resonance imaging (MRI) as an alternative. MRI is a non-invasive imaging technique that allows examination of the bowel wall and extraenteric soft tissues rather than only bowel surface as in ileocolonoscopy. A wide variety of MRI features has already been studied for measuring the disease activity. However, to the best of our knowledge, all the proposed MRI features were manually obtained by clinicians. Apart from being labor intensive, these measurements are inaccurate, irreproducible and non-objective due to the large intra- and inter-observer variability. The Virtual Gastrointestinal Tract (VIGOR++) project aims to deliver a better disease scoring which fulfils all requirements of the ideal assessment of CD severity. The VIGOR++ imaging involved a suite of MRI modalities to be able to quantify the degree of disease activity based on MRI features. An accurate and precise spatial alignment of all MRI modalities is required for optimally measuring those features and have an implicit correspondence. This spatial alignment is commonly referred to as image registration. This thesis presents four registration methods to take up three different challenges. We evaluated our method based on the VIGOR++ data, which uses the CDEIS as the reference standard. The first challenge is related to the respiratory motion that is inherent in free-breathing DCE-MRI. The discontinuities in the deformation field caused by respiratory motion make it difficult to reliably derive features from DCE-MRI. We proposed an expiration-phase, template-based registration method to reduce the discontinuities to the largest extend. Signal enhancement automatically derived from the registered DCE-MRI showed a significantly better correlation to CDEIS than manually measured features. The second challenge derives from local differences in contrast between DCE-MRI and post-contrast MRI caused by tissue-specific uptake of the contrast agent and the MR bias field during imaging. We proposed a method called autocorrelation of local structural information (ALOST). This method overcomes the contrast problems by using the mean phase and phase congruency of the monogenic signal. This method produced better registration results in a comparison with state-of-the-art techniques. It facilitates combining features from post-contrast MRI (e.g. thickness of the bowel wall) and DCE-MRI (signal enhancement). As an alternative, we proposed an efficient registration pipeline based on the Structure Tensor to the local Phase (STOP) which also gave better registration result on our data in comparison with state-of-the-art methods. The third challenge is posed by the large local deformations due to peristalsis and (depth of) respiration in combination with local contrast variations between pre- and post-contrast MRI. We designed a hybrid method coupling discrete descriptor matching with ALOST. The former solves the problem of large local deformations and the latter makes the method insensitive to local contrast changes. The technique called DM-ALOST (descriptor matching ALOST) facilitates semi-automatic extraction of features such as the relative contrast enhancement (RCE) between pre- and post-contrast MRI. The DM-ALOST method gave comparable correlation with CDEIS as the manual methods. The proposed methods paved the way for automatically extracting MRI features for disease assessment from the different MRI modalities. These automatic features were used in a system that gave improved assessment of Crohn’s disease. What is more, the use of our methods is not limited to VIGOR++ project. Particularly, the methods are very useful to cope with breathing motion in other applications, e.g. in liver imaging. Furthermore, they are applicable to other problems that involve large local geometric deformation and intensity variation, e.g. imaging Alzheimer’s disease.