The assessment of small bowel motility with attentive deformable neural network

Abstract

The small bowel is the longest part of the gastrointestinal tract and quick assessment of its motility using Cine-MRI is conducive to the diagnosis of gastroenteric diseases. Because of the complex shape changes that occur frequently in the small bowel, approaches involving human designed features and simple convolutional neural network (CNN) methods fail to achieve satisfactory performance on massive datasets. To meet the challenge of assessing small bowel motility automatically, we propose the integration of deformable convolutional networks into attentive encoder–decoder. With the help of deformable convolution, a tailored CNN can track small bowel segments in different shapes from each MR image of a Cine-MRI sequence. The proposed attentive encoder–decoder performed significantly better than conventional recurrent neural network (RNN) in the assessment of small bowel motility. Experimental results demonstrate that the proposed method not only automatically assesses small bowel motility correctly, but also outperforms state-of-the-art methods. Furthermore, it provides useful information about the physiology of small bowel motility patterns, which can be used in the diagnosis of gastroenteric diseases.