Development of an Artificial Intelligence System for Objective Assessment of Bowel Preparation Quality in Colonoscopy

Aims Bowel preparation quality is critical for effective colonoscopy, yet its evaluation is often subjective and prone to variability among observers. This study aimed to develop and test an Artificial Intelligence (AI) system capable of objectively assessing the quality of bowel preparation in colonoscopy, providing a standardized and reproducible measure [1] [2] [3].

Methods We trained a neural network using video frames from 10 colonoscopies, annotated with color-coded labels for four regions of interest: intestinal mucosa, residuals, artefacts, and colonic lumen. Color fingerprints for each region were created and used to automatically annotate 109,163 colonoscopic frames. A neural network was trained for semantic segmentation with the best 486 annotated frames.

To validate the system, we analysed 7 colonoscopies independently scored by two expert gastroenterologists. They assigned Boston Bowel Preparation Scores (BBPS) for each segment and the total colon and marked the approximate locations of the hepatic and splenic flexures. The trained neural network was applied on the same 7 colonoscopies to separate the regions of interest in each colonoscopy frame. Using a trial and error procedure, frames containing more than 34% artefacts were excluded as non-conform. An equivalent AI score based on detected residual content was compared to the BBPS assigned by experts.

Results The total BBPS ranged from 3 to 7 across the 7 colonoscopies. Interobserver variability was stated in 5 of the 7 cases, with score differences of 1–2 points. The AI system identified an average residual percentage greater than 10% in segments scored≤2 and less than 10% in segments scored 2 or 3.

A correlation was noted between the total BBPS assigned by experts and the AI system's overall residual percentage evaluation. Specifically, a total BBPS of≤5 corresponded to an average residual percentage exceeding 10%. This proof-of-concept AI system demonstrated reasonable alignment with expert evaluations and highlighted the potential for a more objective and consistent assessment of bowel preparation.

Conclusions Our AI-based system successfully correlated residual percentages with BBPS assigned by expert gastroenterologists, demonstrating its potential as a tool for objective bowel preparation assessment. While still in its early stages, this system underscores the need for an AI-driven scale to standardize the evaluation of bowel preparation quality in colonoscopy. Future research will focus on refining the AI model and integrating it into clinical workflows.

Publication History

Article published online:
27 March 2025

• References

• 1 Vulpoi R.-A., Luca M., Ciobanu A., Olteanu A., Barboi O.-B., Drug V.L.. “Artificial Intelligence in Digestive Endoscopy—Where Are We and Where Are We Going?,”. Diagnostics 2022; 12 (no. 4) p 927
• 2 Ciobanu A., Luca M., Vulpoi R., Drug V.. Anatomic Landmarks Detection by Deep Learning in Colonoscopy, Advances in Digital Health and Medical Bioengineering, Ed. H.-N. Costin, R. Magjarević, G. G. Petroiu. pp. 271-278 2024.
• 3 Luca M., Ciobanu A., Vulpoi R.A., Drug V.L.. Deep Learning for Relevant Findings in Colonoscopy. In: Ono, Y., Kondoh, J. (eds) Recent Advances in Technology Research and Education. Lecture Notes in Networks and Systems. 939. Pp. 283-293 Springer; Cham: 2024.