ABSTRACT
Background and Aims
The impact of computer‐aided detection (CADe)‐assisted colonoscopy on gastroenterology fellow trainees remains incompletely defined. In this study, adenoma detection rates (ADR) with and without CADe were assessed among gastroenterology fellows.
Methods
This quality improvement study compared 1580 colonoscopies performed with and without CADe. The primary outcome was the ADR in fellow‐assisted colonoscopies, with and without CADe, stratified by fellowship year. ADRs were compared among procedures assisted by first‐year fellows; those assisted by second‐ and third‐year fellows (combined); and those performed by faculty without a fellow.
Results
Age, birth sex, race, body mass index, procedure indication, non‐neoplastic resection rate, and withdrawal time were similar. CADe‐assisted colonoscopy significantly improved the ADR for first‐year fellows (66.2% vs. 48.6%, p < 0.001). However, for colonoscopies assisted by second‐ and third‐year fellows, the ADR was not significantly different (61.5% vs. 58.7%, p = 0.39). Although there was an upward trend in ADR for CADe‐assisted colonoscopies performed by faculty without fellow participation (61% vs. 50%), it was not statistically significant (p = 0.24).
Conclusion
CADe‐assisted colonoscopy improved ADR in first‐year fellow‐assisted procedures. The bounding box created by the system's artificial intelligence may enhance polyp detection by improving peripheral gaze patterns in novice endoscopists. Further studies are needed to validate this hypothesis.
Keywords: adenoma detection rate, artificial intelligence, computer assisted detection system, endoscopy training
Abbreviations
- ADR
adenoma detection rate
- AI
artificial intelligence
- APC
adenoma per colonoscopy
- BMI
body mass index
- CADe
computer‐aided detection
- CRC
colorectal cancer
- GI
gastroenterology
- IRB
institutional review board
1. Introduction
Interval development of colorectal cancer (CRC) after a normal colonoscopy is a feared outcome. Increased adenoma detection rate (ADR) has been shown to reduce interval CRC [1, 2]. The development and integration of artificial intelligence (AI) with computer‐aided detection (CADe) aids in colon polyp detection [3]. The Food and Drug Administration approved GI Genius (Medtronic) as a validated real‐time CADe system [4].
Computer‐aided detection (CADe)‐assisted colonoscopy is widely used; however, its impact and role in the training of gastroenterology (GI) fellows have not been defined. This study assessed the ADR among GI fellows with and without CADe assistance.
2. Methods
This single‐center cross‐sectional quality improvement study was approved by the institutional review board (IRB) at the Central Virginia Veterans Affairs Health Care System. Data were collected for Veterans undergoing colonoscopy for a year prior to (6/8/2020–6/7/2021) and a year after (6/8/2021–6/7/2022) implementation of the CADe system, GI Genius.
Colonoscopies were performed by a GI fellow supervised by GI faculty or by GI faculty alone (i.e., without a fellow). All 19 fellows and 12 faculty had no prior CADe exposure. First‐year fellows had no endoscopy experience prior to fellowship. Their exposure by the end of the academic year was equivalent. Second‐ and third‐year fellows had identical endoscopic experience. GI Genius was initiated upon withdrawal from the cecum in patients who underwent CADe‐assisted colonoscopy. The system provided an auditory and visual cue (“bounding box”) signaling an area of interest, such as a polyp.
Veterans undergoing high‐definition white light colonoscopy for screening, post‐polypectomy surveillance, or follow up screening after a positive fecal immunohistochemical test were evaluated. The flow diagram (Figure 1) demonstrates the selection of procedures.
FIGURE 1.
Flow diagram of selected procedures. A total of 3228 procedures were evaluated. 1648 were excluded. 960 colonoscopies assisted with CADe and 620 without CADe were analyzed. The average withdrawal time for all included colonoscopies without intervention was at least 6 min. Exclusion criteria included subjects with a history of inflammatory bowel disease (IBD), colorectal cancer (CRC), prior colon resection, inability to interrupt antithrombotic agents precluding polyp resection, polyposis syndromes, diagnostic procedures, and incomplete procedures due to inadequate bowel preparation or inability to intubate the cecum. Indications for diagnostic colonoscopy included iron deficiency anemia, gastrointestinal bleeding, abnormal imaging, suspected inflammatory bowel disease, history of diverticulitis, chronic diarrhea, personal history of colorectal carcinoma in situ endoscopically resected, intended polyp resection not removed on prior colonoscopy, change in bowel habits, and unexplained weight loss. Adequacy of bowel preparation was determined by the Boston Bowel Preparation Scale [5]. CADe, computer‐aided detection; CRC, colorectal cancer; IBD, inflammatory bowel disease.
Baseline information including patient demographics, colonoscopy indications, and procedure‐related metrics was compared for colonoscopies with and without CADe assistance. Year of fellowship training was recorded for all fellow‐assisted procedures. The decision to remove a polyp was at the discretion of the GI faculty.
The primary outcome was adenoma detection rate (ADR), stratified by fellowship year and CADe use. ADR was defined as the percentage of colonoscopies in which at least one adenoma was removed. This was compared among procedures assisted by first‐year fellows, those assisted by second‐ and third‐year fellows (combined), and procedures performed by faculty without a fellow. Additional analyzed outcomes were non‐neoplastic resection rate, withdrawal time, and adenoma per colonoscopy (APC). Non‐neoplastic resection rate was defined as the proportion of resected polyps submitted for pathology that were histologically not an adenoma or a sessile serrated lesion. Withdrawal time was defined as the time spent withdrawing the colonoscope from the cecum to the anus while inspecting mucosa in procedures without biopsy, polypectomy, or other therapeutic interventions. APC was defined as the number of adenomas removed per procedure.
Descriptive statistics were used to summarize the sample characteristics. Continuous variables were assessed for normality using the normal quantile plots and by visual inspection of histograms. Normally distributed variables are presented as mean ± standard deviation (SD), while non‐normally distributed variables are presented as median and interquartile range (IQR). To compare two independent groups for normally distributed continuous variables, the independent samples t‐test was used. For non‐normally distributed continuous variables, the Wilcoxon rank‐sum test was applied. For overall group comparisons, ANOVA with the Tukey–Kramer HSD test was used for post hoc pairwise comparisons among groups, which adjusts for multiple comparisons and accounts for unequal sample sizes. No formal correction for multiple comparisons was applied to nonparametric analyses because they were exploratory and intended to identify potential differences rather than make confirmatory conclusions. Categorical variables were summarized as frequencies and percentages, and group comparisons were conducted using the Pearson chi‐square test. A p value less than 0.05 was considered statistically significant. Statistical analyses were performed using JMP Pro, Version 17 (SAS Institute Inc., Cary, NC, 1989–2023).
3. Results
The total analyzed cohort included 1580 colonoscopies, 960 assisted with CADe and 620 without CADe. Baseline information including patient demographics, colonoscopy indications, and procedure‐related metrics was compared (Table 1).
TABLE 1.
Baseline information comparing colonoscopies with and without CADe.
| Colonoscopy without CADe (n = 620) | Colonoscopy with CADe (n = 960) | p‐value | |
|---|---|---|---|
| Age, years (median IQR) | 66 (59.2–71.8) | 65.8 (58.7–72.2) | 0.65 |
| Birth sex | 0.8 | ||
| Male | 549 (88.5%) | 854 (89%) | |
| Female | 71 (11.5%) | 106 (11%) | |
| Race | 0.14 | ||
| Black | 344 (55.5%) | 485 (50.5%) | |
| White | 266 (42.9%) | 455 (47.4%) | |
| Other | 10 (1.6%) | 20 (2.1%) | |
| BMI, kg/m2 (median IQR) | 30 (26.5–33.2) | 29.7 (26.6–33.6) | 0.96 |
| Indication | 0.17 | ||
| Screening | 155 (25%) | 240 (25%) | |
| Post‐polypectomy surveillance | 356 (57.4) | 584 (60.8%) | |
| Positive fecal immunohistochemical test | 109 (17.6%) | 136 (14.2%) | |
| ADR | 52.58% | 64.58% | 0.0001 |
| APC (mean ± standard error) | 1.31 ± 0.08 | 1.72 ± 0.07 | 0.0004 |
| Non‐neoplastic resection | 9.68% | 9.38% | 0.86 |
| Withdrawal time a , min (median, 25–75 IQR) | 12.15 (9.4–16.2) | 11.55 (9.02–15.09) | 0.64 |
Abbreviations: ADR, adenoma detection rate; APC, adenoma per colonoscopy; BMI, body mass index.
Withdrawal time was analyzed only in procedures without intervention (n = 227 without CADe; n = 243 with CADe).
Age, birth sex, race, body mass index (BMI), procedure indication, non‐neoplastic resection rate, and withdrawal time were similar. ADR (64.6% vs. 52.6%, p = 0.0001) and mean APC (1.72 ± 0.07 vs. 1.31 ± 0.08, p = 0.0004) were significantly higher in CADe‐assisted colonoscopy.
Adenoma detection was compared among fellow‐assisted and faculty‐alone procedures, with and without CADe (Table 2). First‐year fellow CADe‐assisted colonoscopies had a significantly higher ADR (66.2% vs. 48.6%, p < 0.0001) and mean APC (2 ± 0.12 vs. 1.11 ± 0.11, p < 0.0001). For second‐ and third‐year fellow‐assisted colonoscopies (combined), neither ADR (61.5% vs. 58.7%, p = 0.39) and nor mean APC (1.54 ± 0.12 vs. 1.61 ± 0.13, p = 0.71) differed significantly. Faculty‐alone procedures showed nonsignificant upward trends in ADR (61% vs. 50%, p = 0.24) and mean APC (1.28 ± 0.12 vs. 1.06 ± 0.17, p = 0.29) with CADe assistance.
TABLE 2.
Comparison of adenoma detection among three cohorts with and without CADe‐assisted colonoscopy.
| Colonoscopies without CADe | Colonoscopies with CADe | p‐value | |
|---|---|---|---|
| First‐year fellow‐assisted (n = 707) | n = 284 | n = 423 | |
| ADR | 48.6% | 66.2% | < 0.0001 |
| APC (mean ± standard error) | 1.11 ± 0.11 | 2 ± 0.12 | < 0.0001 |
| Second‐ and third‐year‐fellow assisted (n = 564) | n = 252 | n = 312 | |
| ADR | 58.7% | 61.5% | 0.39 |
| APC (mean ± standard error) | 1.61 ± 0.13 | 1.54 ± 0.12 | 0.71 |
| Faculty without a fellow (n = 219 a ) | n = 78 | n = 141 | |
| ADR | 50.0% | 61.0% | 0.24 |
| APC (mean ± standard error) | 1.06 ± 0.17 | 1.28 ± 0.12 | 0.29 |
Abbreviations: ADR, adenoma detection rate; APC, adenoma per colonoscopy.
309 colonoscopies were performed by 12 faculty alone. However, two faculty members performed a total of 90 colonoscopies exclusively in a single cohort (i.e., either with or without CADe assistance). These were excluded due to a lack of crossover. The remaining 219 colonoscopies in the faculty without a fellow analysis were performed by 10 faculty.
4. Discussion
CADe‐assisted colonoscopy significantly improved ADR only for first‐year fellows. No significant difference was observed in second‐ and third‐year fellow‐assisted or faculty‐alone procedures.
Colonoscopy is a specialized procedure relying on technical and cognitive aspects that require different skills: insertion and withdrawal. Initial insertion requires precise advancement to safely navigate. Careful withdrawal requires cognitive recognition and eye‐hand coordination to assess and treat. These skills are acquired through structured training.
The ability to analyze mucosa on a screen requires an intentional visual gaze. There is a natural tendency to focus on the center of the screen leading to inadequate peripheral mucosal assessment [6]. A methodical spiraling motion of the colonoscope during withdrawal brings the periphery into the central field of view. The process of peripheral gaze can be taught. Trained endoscopists with intentional peripheral gaze have a higher adenoma detection [7, 8, 9].
Early in GI training, fellows are primarily focused on “getting to the cecum” and appropriately with less emphasis on mucosal examination. With experience, control of the endoscope improves as does the cognitive ability to assess pathology. Faculty have acquired the skills of peripheral gaze and endoscope manipulation for improved mucosal evaluation. It can be reasoned that the impact of CADe on adenoma detection would be marginal compared to advanced trainees and attending faculty.
We suggest that the CADe “bounding box” reinforces the skill of an intentional visual gaze. During GI training, fellows are taught to use an organized, intentional gaze that surveys the visual field, including the periphery. This learned visual glance triggers peripheral lesions to be brought into the central field of view. It helps shift what is initially seen peripherally into the more discerning central gaze. This enhances polyp detection within the peripheral visual field and reduces miss rates [10, 11].
The study has several strengths and limitations. It was conducted at a single center during the SARS‐CoV‐2 pandemic. This led to a cohort discrepancy which did not affect the primary aim. Fellow involvement per procedure was not recorded. To accommodate trainee involvement, procedure time slots were extended. The indication for colonoscopy and year of fellowship training were evaluated in a univariate analysis. This was not included in a multivariable model to assess the impact of ADR.
5. Conclusion
To our knowledge, this is the first U.S. study to identify an improvement in GI fellows' ADR with the use of AI. The greatest benefit of CADe‐assisted colonoscopy was observed among less experienced endoscopists. We propose that CADe facilitates development of the lateral gaze function, enhancing peripheral vision and positioning peripheral findings into the central field for precise evaluation. Further studies are warranted to validate this hypothesis and clarify the role of CADe in colonoscopy training.
Funding
The authors have nothing to report.
Ethics Statement
All procedures performed were in accordance with the ethical standards of the institution (IRBNet ID# 1658135). In the context of this quality improvement project, informed patient consent for the study was not obtained.
Conflicts of Interest
The authors declare no conflicts of interest.
Acknowledgments
This material is the result of work supported with resources and the use of facilities at the Central Virginia Veterans Affairs Health Care System.
Chadha N., Healey M., Puri P., et al., “Computer‐Aided Detection‐Assisted Colonoscopy Teaches Gaze to Improve Adenoma Detection in First‐Year Gastroenterology Fellows,” JGH Open 10, no. 3 (2026): e70379, 10.1002/jgh3.70379.
The content does not represent the views of the U.S. Department of Veterans Affairs or the United States Government.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
References
- 1. Rajivan R. and Thayalasekaran S., “Improving Polyp Detection at Colonoscopy: Non‐Technological Techniques,” World Journal of Gastrointestinal Endoscopy 15, no. 5 (2023): 354–367, 10.4253/wjge.v15.i5.354. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Shaukat A., Tuskey A., Rao V. L., et al., “Interventions to Improve Adenoma Detection Rates for Colonoscopy,” Gastrointestinal Endoscopy 96, no. 2 (2022): 171–183, 10.1016/j.gie.2022.03.026. [DOI] [PubMed] [Google Scholar]
- 3. Soleymanjahi S., Huebner J., Elmansy L., et al., “Artificial Intelligence–Assisted Colonoscopy for Polyp Detection,” Annals of Internal Medicine 177, no. 12 (2024): 1652–1663, 10.7326/ANNALS-24-00981. [DOI] [PubMed] [Google Scholar]
- 4. Hassan C., Wallace M. B., Sharma P., et al., “New Artificial Intelligence System: First Validation Study Versus Experienced Endoscopists for Colorectal Polyp Detection,” Gut 69, no. 5 (2020): 799–800, 10.1136/gutjnl-2019-319914. [DOI] [PubMed] [Google Scholar]
- 5. Calderwood A. H., Schroy P. C., Lieberman D. A., Logan J. R., Zurfluh M., and Jacobson B. C., “Boston Bowel Preparation Scale Scores Provide a Standardized Definition of Adequate for Describing Bowel Cleanliness,” Gastrointestinal Endoscopy 80, no. 2 (2014): 269–276, 10.1016/j.gie.2014.01.031. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Tatler B. W., “The Central Fixation Bias in Scene Viewing: Selecting an Optimal Viewing Position Independently of Motor Biases and Image Feature Distributions,” Journal of Vision 7, no. 14 (2007): 4, 10.1167/7.14.4. [DOI] [PubMed] [Google Scholar]
- 7. Nagai M., Ishibashi F., Okusa K., et al., “Optimal Visual Gaze Pattern of Endoscopists for Improving Adenoma Detection During Colonoscopy (With Video),” Gastrointestinal Endoscopy 101, no. 3 (2025): 639–646.e3, 10.1016/j.gie.2024.09.028. [DOI] [PubMed] [Google Scholar]
- 8. Karamchandani U., Erridge S., Evans‐Harvey K., Darzi A., Hoare J., and Sodergren M. H., “Visual Gaze Patterns in Trainee Endoscopists—A Novel Assessment Tool,” Scandinavian Journal of Gastroenterology 57, no. 9 (2022): 1138–1146, 10.1080/00365521.2022.2064723. [DOI] [PubMed] [Google Scholar]
- 9. Edmondson M. J., Pucher P. H., Sriskandarajah K., et al., “Looking Towards Objective Quality Evaluation in Colonoscopy: Analysis of Visual Gaze Patterns,” Journal of Gastroenterology and Hepatology 31, no. 3 (2016): 604–609, 10.1111/jgh.13184. [DOI] [PubMed] [Google Scholar]
- 10. Cold K. M., Vamadevan A., Vilmann A. S., Svendsen M. B. S., Konge L., and Bjerrum F., “Computer‐Aided Quality Assessment of Endoscopist Competence During Colonoscopy: A Systematic Review,” Gastrointestinal Endoscopy 100, no. 2 (2024): 167–176.e1, 10.1016/j.gie.2024.04.004. [DOI] [PubMed] [Google Scholar]
- 11. Yao L., Li X., Wu Z., et al., “Effect of Artificial Intelligence on Novice‐Performed Colonoscopy: A Multicenter Randomized Controlled Tandem Study,” Gastrointestinal Endoscopy 99, no. 1 (2024): 91–99.e9, 10.1016/j.gie.2023.07.044. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.