Toward automated small bowel capsule endoscopy reporting using a summarizing machine learning algorithm: The SUM UP...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) excels at detecting small bowel (SB) vascular lesions in capsule endoscopy (CE), reducing reading time.
• Characterizing lesion bleeding potential (high P2, intermediate P1, low P0) is crucial for diagnosis.
• Improving DL performance and automating report generation are key research goals.

Purpose of the Study:

• To enhance DL performance for SB CE by using a machine learning (ML) classifier to characterize vascular abnormalities.
• To select the most relevant images for inclusion in reports, reducing manual review time.
• To improve the diagnostic accuracy of vascular lesion detection and characterization in SB CE.

Main Methods:

• A dataset of 75 SB CE videos with 1,525 vascular lesion images was used for training.
• Image classification algorithms were tested to differentiate typical angiodysplasia (P2/P1) from other lesions (P0).
• The best-performing algorithm was validated on an independent test set of 73 SB CE videos.

Main Results:

• A random forest (RF) classifier achieved 91.1% specificity and 84.2% accuracy in discriminating P2/P1 from P0 lesions.
• The RF method reduced the number of reported images by 83.2%, with a 91.6% decrease in the test set.
• The RF algorithm showed 98.0% agreement with human reporting, enhancing characterization and image selection.

Conclusions:

• The ML-enhanced DL approach improves characterization and selection of relevant SB vascular abnormalities for CE reporting.
• This method maintains diagnostic accuracy while significantly reducing the volume of images for review.
• These findings support the potential for automated SB CE reporting.