AUTOMATED SEGMENTATION OF STOMACH CONTENTS IN MRI IMAGES WITH THE USE OF DEEP LEARNING.

This AI solution contributes to more knowledge on gastric digestion and thereby the development of optimized foods for target groups such as elderly, athletes and infants. It also sets the stage for better characterization of gastrointestinal physiology and disorders involving altered gastric emptying rate and gastric juice production.

Research on food digestion and gastric emptying can help in optimizing nutrient absorption by assessing in vivo digestibility and by providing knowledge that can be used to design food products with optimal digestive characteristics so as to promote satiety or protein breakdown dynamics.  Important target groups are older adults, critically ill, athletes and individuals with gastrointestinal diseases. 

Stomach content
The physical properties of food in the stomach can affect the dynamics of gastric digestion, stomach emptying, intestinal digestion, and nutrient absorption. MRI is a very versatile, non-invasive, medical imaging technique that can be used to monitor several aspects of gastrointestinal transit and digestion.  Although there have been a few attempts to automate the measurement of stomach content volume, this has not been very successful and all major research groups still rely on extremely time-consuming manual segmentation; stomach content must be delineated on every slice of the 3D abdominal MRI images.

Deep learning
Previously, automated segmentation was not readily achievable because of the large variations in the shape, size and texture of the stomach contents. This has changed with the availability of so-called convolutional neural networks (CNNs), which provide a powerful approach that can deal with more complex image segmentation problems.  WUR and TU/e have conducted exploratory work to assess the potential of such a ‘deep learning’ approach for the segmentation of the stomach contents on MRI images. The project objective is to build and validate an AI-based software tool that can perform automatic segmentation of the stomach contents on abdominal MRI images. 

Future 
The tool can be further improved by adding data and its functionality expanded with dedicated characterization and analysis of the stomach contents, e.g. calculating image texture metrics and shape characteristics which can then be linked to food and individual characteristics. This provides knowledge that can be used in fundamental and applied research and food development for different target groups. In addition, the tool could be further developed to become part of clinical screening and diagnosis for gastric disorders.

Research team

Maureen van Eijnatten (TU/e), Paul Smeets (WUR), Elise van Eijnatten, (WUR), Guido Camps (WUR), Duco Veen (UU), Wilbert Bartels (UMC Utrecht), Jan Monkelbaan (UMC Utrecht),