Development of a deep learning-based software for calculating cleansing score in small bowel capsule endoscopy

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Standardized small bowel (SB) cleansing scales are lacking.
• Accurate assessment of SB preparation is crucial for capsule endoscopy efficacy.
• Current subjective grading methods can be inconsistent.

Purpose of the Study:

• To develop automated software for calculating SB cleansing scores using deep learning.
• To establish an objective criterion for adequate bowel preparation.
• To correlate cleansing scores with diagnostic yield for small lesions.

Main Methods:

• Deep learning model trained on a 5-step scoring system using capsule endoscopy images from three hospitals.
• Software performance evaluated on a validation set.
• Comparison of automated scores with clinician-based grading (A-C).

Main Results:

• Automated cleansing scores correlated inversely with clinical grading (A=4.1, B=3.5, C=2.9).
• 91.7% of cases achieved adequate preparation based on automated scoring.
• A cut-off score of 3.25 demonstrated high accuracy (AUC 0.977) for identifying adequate preparation.
• Higher cleansing scores were associated with increased diagnostic yield for small lesions.

Conclusions:

• Novel software provides objective, automated SB cleansing scores.
• The proposed cut-off value serves as a standard criterion for adequate bowel preparation.
• Automated scoring enhances quality assessment and potentially improves diagnostic outcomes in capsule endoscopy.