Deep learning for automatic volumetric bowel segmentation on body CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate segmentation of the gastrointestinal (GI) tract is crucial for quantitative analysis.
• Estimating large bowel length (LBL) can aid in diagnosing conditions like constipation.
• Current methods for LBL measurement can be invasive or lack precision.

Purpose of the Study:

• To develop a deep neural network (DNN) for automated bowel segmentation on CT images.
• To assess the DNN's accuracy in segmenting the entire GI tract and its compartments.
• To evaluate the DNN's utility in estimating LBL and its correlation with constipation.

Main Methods:

• Utilized 3D nnU-Net models trained on contrast-enhanced and non-enhanced CT scans.
• Segmented the GI tract into esophagus, stomach, small bowel, and large bowel.
• Evaluated segmentation accuracy using Dice Similarity Coefficient (DSC) and compared LBL between constipated and non-constipated groups.

Main Results:

• The DNN achieved high segmentation accuracy for the GI tract (mean DSC 0.985) and its four compartments (DSC > 0.95), outperforming existing methods.
• External testing confirmed the model's robust performance.
• Height-corrected LBL was significantly longer in individuals with constipation compared to those without.

Conclusions:

• The developed 3D nnU-Net model provides accurate segmentation of the GI tract and its major components from CT data.
• This AI-driven approach enables noninvasive estimation of LBL, showing potential for clinical use in diagnosing and managing constipation.