Fully Automated Colon Delineation and Volume Estimation in T2-Weighted MRI with a 2D U-Net

Artificial intelligence has the potential to provide an objective, accurate and fast evaluation of the colon for clinical assessment of gastrointestinal diseases, such as chronic constipation. Accurate colon segmentation is a labor-intensive task, prone to interobserver variability. This study presents a method for automatic colon segmentation in T2-weighted MRI images, using 297 MRI scans from 72 healthy subjects and a neural network based on the 2D U-Net architecture. The presented network provided a Dice similarity coefficient (DSC) of 0.82 ± 0.04 when compared to ground truth data (p=0.722), allowing for accurate evaluations of colonic physiology. The estimated colonic volumes were significantly correlated with the ground truth volumes with no tendency for over- or underestimation (p<0.001). This study is, to our knowledge, the first to fully automatically segment the colon using MRI data, thus providing a method for objective colon evaluation without radiation exposure or disturbance of its physiology.