A neural network algorithm for detection of GI angiectasia during small-bowel capsule endoscopy

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Gastrointestinal angiectasia (GIA) is a prevalent small-bowel vascular lesion associated with bleeding risks.
• Small-bowel capsule endoscopy (SB-CE) is the standard diagnostic method for GIA.
• Current diagnostic procedures for GIA can be time-consuming.

Purpose of the Study:

• To develop and evaluate a computer-assisted diagnosis (CAD) tool for detecting GIA.
• To leverage deep learning for enhanced GIA identification in SB-CE.
• To improve the efficiency and accuracy of GIA diagnosis.

Main Methods:

• A dataset of small-bowel capsule endoscopy (SB-CE) still frames, including annotated GIA and normal controls, was utilized.
• A semantic segmentation approach combined with a convolutional neural network (CNN) was employed for feature extraction and classification.
• Two distinct datasets were curated for machine learning training and algorithm validation.

Main Results:

• The GIA detection algorithm demonstrated high diagnostic performance: 100% sensitivity, 96% specificity, 96% positive predictive value, and 100% negative predictive value.
• The algorithm exhibited optimal reproducibility.
• The estimated reading time for a complete SB-CE video using this tool was 39 minutes.

Conclusions:

• The developed CNN-based algorithm achieves high diagnostic accuracy for GIA detection in SB-CE still frames.
• This technology shows promise for future automated SB-CE reading software.
• The CAD tool has the potential to streamline the diagnostic process for GIA.