Magnetic resonance enterography (MRE) based indices are used to assess the severity of Crohn's disease (CD) in patients. Manual calculation of these indices by radiologists is both time consuming and subject to bias. Specifically, terminal-ileum (TI) wall-thickening is challenging to estimate, especially from gadolinium-free T2-weighted MRI data. We present a machine-learning-based system to automatically segment the TI and classify wall-thickening from T2-weighted MRE data. We introduce an anatomically-constrained U-net segmentation model to cope with the inherently heterogeneous appearance, large anatomical variability and limited availability of data for TI segmentation from MRE data. We evaluated the added-value of our anatomically-constrained segmentation model on 180 T2-weighted MRE scans collected as part of the ImageKids study using k-fold cross-validation experimental setup. The anatomically-constrained segmentation model improved segmentation results compared to a vanilla U-Net by means of both Dice score and Housdorff distance (17.931mm vs. 19.97mm, p<0.05). We used histogram-based features with an SVM classifier to distinguish between patients with and without TI high wall thickness (> 8mm vs. < 3mm, N=18). Leave-one-out experimental demonstrated high agreement with the radiologists' assessment (F1 score of 0.933 and a Cohen's kappa score of 0.6). The proposed approach has the potential to facilitate objective assessment of CD from MRE data.