Abstract
Background
The impact of computer-aided detection (CADe) systems on lesion detection remains uncertain. We evaluated whether a CADe system improves adenoma detection rate (ADR) in a colorectal cancer (CRC) screening program.
Methods
This multicenter randomized clinical trial included patients aged 40–79 years who underwent colonoscopy after a positive fecal immunochemical test (FIT) or for surveillance. Patients were randomized to either conventional or CADe-assisted colonoscopy. The primary end point was ADR. Secondary end points were detection rates for serrated lesions (SLDR), polyps (PDR), advanced adenomas (AADR), and advanced serrated lesions (ASLDR). We performed subgroup analysis according to indication, endoscopist ADR, and bowel cleansing. Effect estimates were reported as adjusted risk ratios (aRRs) and 95%CIs.
Results
857 patients were randomized and 827 were included in the analysis. Baseline characteristics were comparable between groups (mean age: 61.7 [SD 6.5]; male 490; positive FIT: 552). We did not find statistically significant differences between CADe-assisted and conventional colonoscopy for ADR (60.8% vs. 57.7%; aRR 1.05, 95%CI 0.94–1.17). No significant differences were observed for SLDR (8.8% vs. 7.2%; aRR 1.04, 0.65–1.64), PDR (69.6% vs. 68.8%; aRR 1.00, 0.91–1.10), AADR (23.6% vs. 24.4%; aRR 0.90, 0.71–1.14), or ASLDR (4.9% vs. 4.8%; aRR 0.93, 0.51–1.68). Subgroup analysis showed no significant differences.
Conclusion
In colonoscopies performed in a CRC screening program, CADe-assisted colonoscopy did not increase ADR compared with conventional colonoscopy.
Abstract
Graphical abstract
Graphical abstract for this article.
Introduction
Colorectal cancer (CRC) is the third most common malignancy worldwide and the second leading cause of cancer-related mortality 1 2 . In average-risk populations, adenomatous and serrated lesions may grow within the colonic mucosa and progress to CRC 3 . Population-based screening programs aim to prevent this disease through two strategies: the fecal immunochemical test (FIT) and colonoscopy 4 5 . Early detection and removal of precancerous lesions have demonstrated significant reductions in CRC incidence and mortality 4 5 6 . FIT screening is more widespread in Europe, whereas colonoscopy is more common in the United States 5 . In Spain, individuals are invited to complete FIT screening every 2 years, and those with a positive result undergo colonoscopy to detect and resect colorectal lesions 4 5 . Although both strategies effectively reduce CRC risk, colonoscopy remains the first-line procedure for the detection and resection of colorectal lesions 4 5 .
Missed preneoplastic lesions during colonoscopy contribute to the development of post-colonoscopy CRC 7 . Several quality indicators have been evaluated in order to improve colonoscopy performance and, thus, reduce the risk of post-colonoscopy CRC. Among these metrics, the adenoma detection rate (ADR) is the most widely validated indicator of colonoscopy quality, reflecting the endoscopist’s ability to identify adenomatous lesions during the procedure 8 9 10 . Each 1% increase in ADR is associated with an approximately 3% reduction in CRC risk and a corresponding decrease in the incidence of post-colonoscopy CRC 11 . However, the ADR level is variable among endoscopists owing to colonoscopy factors (i.e. bowel cleansing, cecal intubation, or mucosal exposure) and operator-related factors (i.e. workload, fatigue, or experience level) 12 13 . Such factors may contribute to a reduction in ADR performance and an increase in undetected colonic lesions 12 .
Artificial intelligence (AI) encompasses the fields of machine learning and deep learning 14 15 . Through neural network architectures and appropriate training, AI-based systems can perform tasks traditionally undertaken by humans 15 . Over the past decade, deep learning-based algorithms have been increasingly applied to medical imaging, where they have shown potential in assisting clinicians in enhancing detection and diagnostic performance 16 17 . In colonoscopy, both commercial and in-house computer-aided detection (CADe) systems have been evaluated in multiple randomized clinical trials (RCTs), demonstrating enhanced ADR and reduced adenoma miss rate in assisted compared with conventional colonoscopies 18 19 20 . Despite these findings, the incorporation of CADe into routine colonoscopy remains controversial. Endoscopist expert panels have issued weak, uncertain, or negative recommendations to its widespread adoption 21 22 . Moreover, evidence regarding the effectiveness of CADe in screening programs involving expert endoscopists with high baseline ADRs remains limited 18 23 .
PolyDeep (developed by the SING Group and the GIODO Group, Galicia, Spain, 2017) is a deep learning-based system integrating CADe and computer-aided optical diagnosis designed to support endoscopists during real-time colonoscopy procedures as previously described 24 25 26 27 28 . In our previous study, we evaluated CADe functionality in colonoscopy. Its detection network was built on YOLOv3 architecture and incorporates an object-tracking algorithm that maintains stable on-screen bounding boxes that highlight potentially suspicious lesions 24 26 29 . The system was evaluated in an observational study focused on polyp detection and in an RCT evaluating the adenoma miss rate 27 28 . The diagnostic performance of PolyDeep for polyp detection was comparable to that of expert endoscopists, while PolyDeep-assisted colonoscopy did not significantly reduce the adenoma miss rate compared with conventional colonoscopy 27 28 . Hence, previous clinical trials reported conflicting results about the benefit of PolyDeep in colonoscopy. In the present RCT, we aimed to determine whether PolyDeep-assisted colonoscopy increases the ADR compared with conventional colonoscopy.
Methods
Study design
PolyDeep Advance 3 was a multicenter RCT performed in three university hospitals of Galicia (Spain). The study was approved by the institutional review board (2022/067) and carried out according to good clinical practice and the principles of the Declaration of Helsinki. We report this RCT according to the CONSORT-AI guidelines.
Participants of the study
We prospectively enrolled individuals aged 40–79 years undergoing colonoscopy after a positive FIT (≥20 µg hemoglobin per gram of feces) within the Galician population-based CRC screening program, or as part of post-polypectomy surveillance after resection of advanced adenomas. We excluded patients with a personal history of CRC, previous colonic resection, or any hereditary syndrome or serrated polyposis. Patients with inadequate bowel cleansing according to the Boston Bowel Preparation Scale (BBPS; i.e. score ≥2 in each segment) or with no cecal intubation were excluded. All the participants included in the study signed a written informed consent form. Demographic information, colonoscopy details, and lesion characteristics were collected on an electronic case report form on the REDCap platform hosted by Galicia-Sur Health Research Institute. After trial completion, data from each center were subject to monitoring and verification procedures to ensure data quality and integrity.
Randomization process
We implemented a stratified block randomization design, with stratification according to indication (positive FIT or surveillance colonoscopy). Randomization was performed using blocks of two, comprising all possible group allocations, which were arranged in random order within a predefined randomization template. The template contained twice the required number of allocation slots (1708 slots for 854 participants) to ensure completeness and prevent allocation bias. After finalization, the randomization template was uploaded to the electronic case report form. Participants were subsequently randomized in a 1:1 ratio to undergo either conventional colonoscopy or PolyDeep-assisted colonoscopy. Endoscopists were aware of the assigned intervention, whereas patients were blinded to group allocation.
Clinical setting
Colonoscopies were performed by endoscopists in standard endoscopy units using high-definition colonoscopes (model EXERA III CV-190 or higher; Olympus, Tokyo, Japan). PolyDeep was connected to both the endoscopy processor and the monitor displaying the conventional colonoscopy view. When activated during the withdrawal phase, PolyDeep processed the video stream in real time and displayed bounding boxes over areas suspected of containing lesions and made a prediction of neoplasia ( Fig. 1 ). The endoscopists submitted all resected lesions for histopathological analysis. In cases where the lesion could not be resected, a biopsy was taken and sent for histological evaluation. The withdrawal time was calculated using the timer integrated in the colonoscope and displayed on the screen.
Fig. 1.
PolyDeep system for polyp detection and characterization. a Initial detection phase with a bounding box around a suspicious lesion. b Confirmed detection phase, highlighted by a red visual alert on the screen borders. c Optical diagnosis mode; the system provides a neoplasia probability score. A magnifying glass icon prompts the endoscopist for an optimized view, and an hourglass icon indicates an extended focus duration for a definitive diagnosis. d Validated optical diagnosis after focus stabilization, showing the final neoplasia probability.
A total of 26 expert endoscopists participated in the study within the Galician CRC screening program. All endoscopists had more than 1 year of experience, had performed over 300 colonoscopies, and underwent annual performance monitoring based on key quality indicators, including ADR. According to the internal 2024 quality reports of the Galician CRC screening program, the participating endoscopists achieved a median ADR of 61.4% (interquartile range 55.3%–66.2%) in colonoscopies performed after a positive FIT. Based on this program-level benchmark, endoscopists were classified as low detectors (≤61.4%) or high detectors (>61.4%).
During the colonoscopy, a study nurse or assistant collected data relating to the colonoscopy (i.e. withdrawal time) and the lesions identified (i.e. lesions characteristics and optical diagnosis) in a log sheet. Assistants were not involved in lesion detection and did not identify or point out polyps during the withdrawal phase. On the electronic case report form, investigators recorded patient baseline characteristics (sex, age, indication, and intervention group), colonoscopy-related variables (including BBPS score and cecal intubation), and lesion characteristics, together with optical diagnosis and histopathological findings when available.
End points
The primary end point of the trial was the ADR. Secondary end points included detection rates for serrated lesions (SLDR), polyps (PDR), advanced adenomas (AADR), advanced serrated lesions (ASLDR), and advanced polyps (APDR). We defined advanced polyps as adenomas ≥10 mm, or those with villous histology or high grade dysplasia, as well as advanced serrated lesions including sessile and traditional serrated adenoma ≥10 mm in size, or with dysplasia. Both primary and secondary end points were calculated as the proportion of colonoscopies in which at least one lesion of the corresponding type was detected, divided by the total number of colonoscopies performed. Withdrawal time was evaluated as an additional secondary end point to allow comparison between the two study groups.
Sample size
We calculated the sample size to detect a clinically relevant difference in ADR between conventional colonoscopy and PolyDeep-assisted colonoscopy. Based on prior evidence 27 28 , we assumed an ADR of 50% in the conventional colonoscopy group and 60% in the intervention group, corresponding to an absolute difference of 10%. The calculation was based on a comparison of two independent proportions under a two-sided superiority framework, with a type I error probability (α) of 0.05 and a type II error probability (β) of 0.20, corresponding to 80% power, assuming a 1:1 allocation ratio. Under these assumptions, 388 patients per group were required. Considering a dropout rate of 10%, we increased the total sample size to 854 patients (427 per arm). All sample size calculations were performed using the Epidat 3.1 (Xunta de Galicia, Servicio Galego de Saude, Galicia, Spain).
Statistical analysis
We defined a modified intention-to-treat population analysis to include all randomized colonoscopies with evaluable examinations. Categorical variables are presented as absolute and relative frequencies, and continuous variables are expressed as mean and SD. We performed a descriptive analysis for each study group. Comparisons were performed using the two-sided chi-squared test for categorical variables and the two-sided Student’s t test for independent continuous variables. Fisher’s exact test was applied when appropriate.
Associations between study groups and lesion detection outcomes were estimated using Poisson regression models with log link and robust variance, allowing the estimation of adjusted risk ratios (aRR) and their corresponding 95%CIs. All models were adjusted for age, sex, endoscopist, and group allocation. We reported the 95%CIs for primary and secondary end points.
Subgroup analyses were performed to evaluate PDR according to lesion location, lesion size, and colonoscopy indication. Bowel preparation quality was categorized using the BBPS total score as good (score 6–7) or excellent (score 8–9), and detection rates were calculated within each category. These analyses were considered exploratory and were further evaluated using a Poisson regression model to estimate aRR and 95%CIs. Detection rates were compared between low-and high-detector endoscopists (i.e. ≤61.4% and >61.4%), as defined by the annual ADR monitoring of the endoscopists participating in the population-based CRC screening program.
All statistical analyses were conducted using R software version 4.4.1 (R Foundation for Statistical Computing, Vienna, Austria).
Results
Study population and baseline characteristics
A total of 857 individuals were randomized, of whom 827 were included in the final analysis: 416 underwent conventional colonoscopy and 411 underwent PolyDeep-assisted colonoscopy ( Fig. 2 ). Overall, 30 participants were excluded after randomization because of inadequate bowel preparation or incomplete examination.
Fig. 2.
Flowchart of the study.
Baseline demographic and procedural characteristics were well balanced between groups ( Table 1 ). The mean age was 61.7 (SD 6.5) years, and 59.3% of participants were male. Most procedures (66.7%) were performed after a positive FIT, while the remaining were surveillance colonoscopies. No statistically significant differences were observed between groups with respect to age, sex, indication for colonoscopy, bowel preparation quality, withdrawal time, or the proportion of examinations performed by high- vs. low-performing endoscopists.
Table 1 Baseline characteristics of patients undergoing conventional and computer-assisted (PolyDeep) colonoscopy.
| Conventional (N = 416) | PolyDeep (N = 411) | P 1 | |
| Data are n (%) unless otherwise specified. ADR, adenoma detection rate; BBPS, Boston Bowel Preparation Scale; FIT, fecal immunochemical test. 1 Student's t test for independent continuous variables and chi-squared test for categorical variables, with significance set at P < 0.05. 2 Includes instrumentation time (e.g. polypectomy, biopsy, and water cleaning). 3 ADR >61.4%, based on reported performance in 2024. 4 Number of colonoscopies in which the endoscopists detect lesions on the initial diagnosis with no pathology assessment. | |||
| Age, mean (SD), years | 62.0 (6.5) | 61.4 (6.6) | 0.19 |
| Male sex | 246 (59.1) | 244 (59.4) | >0.99 |
| FIT indication | 272 (65.4) | 280 (68.1) | 0.44 |
| BBPS total score 8–9 | 243 (58.4) | 211 (51.3) | 0.05 |
| Withdrawal time, 2 mean (SD), minutes: seconds | 16:37 (10:03) | 17:08 (09:49) | 0.45 |
| High-performing colonoscopist 3 (ADR >61.4%) | 171 (41.1) | 166 (40.4) | 0.88 |
| Detection of lesions 4 | 306 (73.6) | 305 (74.2) | 0.83 |
| Colorectal cancer | 14 (3.4) | 8 (1.9) | 0.32 |
Across the entire cohort, 23 CRCs were detected, 14 in the conventional colonoscopy group and 9 in the PolyDeep-assisted group. At least one adenoma was identified in 59.3% of colonoscopies, serrated lesions in 8.0%, and any polyp in 69.2%. The mean number of polyps per colonoscopy was 2.9 (SD 3.5), including 1.4 (SD 1.9) adenomas and 0.6 (SD 1.7) serrated lesions. Most lesions were diminutive, with a mean size of 5.05 (SD 4.8) mm, and were more frequently located in the distal colon.
Primary and secondary outcomes
The primary end point, ADR, did not differ significantly between the PolyDeep-assisted and conventional colonoscopy groups (60.8% vs. 57.7%; aRR 1.05, 95%CI 0.94–1.17) ( Table 2 ). Similarly, no significant differences were observed in SLDR (8.8% vs. 7.2%; aRR 1.04, 95%CI 0.65–1.64) or PDR (69.6% vs. 68.8%; aRR 1.00, 95%CI 0.91–1.10). Detection rates for advanced lesions were also comparable between groups. AADR was 23.6% in the PolyDeep group and 24.5% in the conventional group (aRR 0.90, 95%CI 0.71–1.14), ASLDR was 4.9% vs. 4.8% (aRR 0.93, 95%CI 0.51–1.68), and APDR was 27.3% vs. 28.1% (aRR 0.91, 95%CI 0.73–1.13).
Table 2 Comparison of detection outcomes between conventional and computer-assisted (PolyDeep) colonoscopy.
| Conventional (N = 416) | PolyDeep (N = 411) | aRR (95%CI) 1 | |
| aRR, adjusted risk ratio. 1 Poisson regression models with log link and robust variance. 2 Includes adenomas and serrated lesions. 3 Includes lesions >10 mm, or with tubulovillous or villous histology, or with high grade dysplasia. 4 Includes serrated lesions >10 mm or with high grade dysplasia. 5 Includes both advanced adenoma and advanced serrated lesion. 6 Includes cecum, ascending colon, hepatic flexure, transverse colon, splenic flexure. 7 Includes descending colon, sigmoid colon, rectum. 8 Includes polyps <5 mm. 9 Include polyps with slightly elevated, flat and depressed morphology. | |||
| Detection rates, n (%) | |||
|
240 (57.7) | 250 (60.8) | 1.05 (0.94–1.17) |
|
30 (7.2) | 36 (8.8) | 1.04 (0.65–1.64) |
|
286 (68.8) | 286 (69.6) | 1.00 (0.91–1.10) |
|
102 (24.5) | 97 (23.6) | 0.90 (0.71–1.14) |
|
20 (4.8) | 20 (4.9) | 0.93 (0.51–1.68) |
|
117 (28.1) | 112 (27.3) | 0.91 (0.73–1.13) |
|
201 (48.3) | 203 (49.4) | 1.02 (0.89–1.17) |
|
220 (52.9) | 207 (50.4) | 0.94 (0.82–1.07) |
|
224 (53.8) | 234 (56.9) | 1.07 (0.95–1.21) |
|
258 (62.0) | 270 (65.7) | 1.06 (0.96–1.17) |
|
182 (43.8) | 175 (42.6) | 0.95 (0.81–1.10) |
|
132 (31.7) | 121 (29.4) | 0.95 (0.78–1.15) |
| Number of lesions, mean (SD) | |||
|
1.5 (2.0) | 1.4 (1.8) | – |
|
0.6 (1.4) | 0.6 (1.9) | – |
|
2.9 (3.0) | 2.8 (3.9) | – |
Analysis of lesion location demonstrated similar detection rates in both groups for polyps located in the proximal colon (49.4% with PolyDeep vs. 48.3% with conventional colonoscopy; aRR 1.02, 95%CI 0.89–1.17) and in the distal colon (50.4% vs. 52.9%; aRR 0.94, 95%CI 0.82–1.07). Detection performance across polyp size categories was also equivalent. No significant differences were observed for polyps <5 mm, <10 mm, or ≥5 mm. In particular, detection of diminutive lesions (<5 mm) occurred in 56.9% of PolyDeep-assisted examinations and 53.8% of conventional examinations (aRR 1.07, 95%CI 0.95–1.21), indicating consistent performance for small and clinically relevant lesions.
Subgroup analysis by colonoscopy indication
Among the 552 participants undergoing colonoscopy after a positive FIT, ADR was 65.4% in the PolyDeep-assisted group and 61.8% in the conventional group, without a statistically significant difference (aRR 1.02, 95%CI 0.90–1.15). SLDR (9.6% vs. 7.0%; aRR 1.15, 95%CI 0.66–1.98) and PDR (72.1% in both groups; aRR 0.97, 95%CI 0.87–1.08) were also similar ( Table 3 ). Detection of advanced lesions did not differ between groups: AADR 32.1% vs. 34.6% (aRR 0.87, 95%CI 0.69–1.10), ASLDR 6.4% vs. 5.5% (aRR 1.09, 95%CI 0.57–2.09), and APDR 36.8% vs. 37.5% (aRR 0.92, 95%CI 0.74–1.14). No differences were observed according to lesion location or size.
Table 3 Subgroup analyses of detection outcomes between conventional and computer-assisted (PolyDeep) colonoscopy by indication.
| Positive FIT (N = 552) | aRR 1 (95%CI) | Surveillance (N = 275) | aRR 1 (95%CI) | |||
| Conventional (N = 272) | PolyDeep (N = 280) | Conventional (N = 144) | PolyDeep (N = 131) | |||
| ADR, adenoma detection rate; aRR, adjusted risk ratio; BBPS, Boston Bowel Preparation Scale; FIT, fecal immunochemical test; PDR, polyp detection rate; SLDR, serrated lesion detection rate. 1 Poisson regression models with log link and robust variance. | ||||||
| Detection rates, n (%) | ||||||
| ADR | 168 (61.8) | 183 (65.4) | 1.02 (0.90–1.15) | 72 (50.0) | 67 (51.1) | 1.02 (0.80–1.29) |
| SLDR | 19 (7.0) | 27 (9.6) | 1.15 (0.66–1.98) | 11 (7.6) | 9 (6.9) | 0.72 (0.30–1.71) |
| PDR | 196 (72.1) | 202 (72.1) | 0.97 (0.87–1.08) | 90 (62.5) | 84 (64.1) | 1.00 (0.84–1.21) |
In the surveillance subgroup (n = 275), detection rates were lower overall but remained comparable between groups. ADR was 51.1% in the PolyDeep group and 50.0% in the conventional group (aRR 1.02, 95%CI 0.80–1.29), SLDR was 6.9% vs. 7.6% (aRR 0.72, 95%CI 0.30–1.71), and PDR was 64.1% vs. 62.5% (aRR 1.00, 95%CI 0.84–1.21). Advanced lesion detection and all additional stratified analyses by size and location were similarly balanced between groups.
Subgroup analysis by endoscopist performance
When stratified by endoscopist baseline performance, no statistically significant differences were observed between PolyDeep-assisted and conventional colonoscopy in either subgroup. Among low detectors (≤61.4% baseline ADR), ADR was 55.5% with PolyDeep and 52.2% with conventional colonoscopy (aRR 1.03, 95%CI 0.87–1.21). Corresponding SLDR values were 10.6% vs. 9.4% (aRR 0.94, 95%CI 0.55–1.60), and PDR was 66.1% vs. 63.3% (aRR 1.01, 95%CI 0.89–1.16).
Among high detectors (>61.4%), ADR was 68.7% in the PolyDeep group and 65.5% in the conventional group (aRR 1.07, 95%CI 0.91–1.24), SLDR was 6.0% vs. 4.1% (aRR 1.39, 95%CI 0.58–3.32), and PDR was 74.7% vs. 76.6% (aRR 0.98, 95%CI 0.86–1.10). These findings indicate consistent performance across both endoscopist strata ( Table 4 ).
Table 4 Subgroup analyses of detection outcomes between conventional and computer-assisted (PolyDeep) colonoscopy by endoscopists baseline adenoma detection rate.
| Low detectors 1 (N = 490) | aRR 2 (95%CI) | High detectors 1 (N = 337) | aRR 2 (95%CI) | |||
| Conventional (N = 245) | PolyDeep (N = 245) | Conventional (N = 171) | PolyDeep (N = 166) | |||
| ADR, adenoma detection rate; aRR, adjusted risk ratio; BBPS, Boston Bowel Preparation Scale; FIT, fecal immunochemical test; PDR, polyp detection rate; SLDR, serrated lesion detection rate. 1 Low detectors = ADR ≤61.4%; high detectors = ADR >61.4% 2 Poisson regression models with log link and robust variance. | ||||||
| Detection rates, n (%) | ||||||
| ADR | 128 (52.2) | 136 (55.5) | 1.03 (0.87–1.21) | 112 (65.5) | 114 (68.7) | 1.07 (0.91–1.24) |
| SLDR | 23 (9.4) | 26 (10.6) | 0.94 (0.55–1.60) | 7 (4.1) | 10 (6.0) | 1.39 (0.58–3.32) |
| PDR | 155 (63.3) | 162 (66.1) | 1.01 (0.89–1.16) | 131 (76.6) | 124 (74.7) | 0.98 (0.86–1.10) |
Subgroup analysis by bowel preparation quality
Bowel preparation quality did not modify the effect of PolyDeep. In colonoscopies with good preparation (BBPS total score 6–7), ADR was 62.0% in the PolyDeep group and 59.5% in the conventional group (aRR 1.06, 95%CI 0.89–1.23), SLDR was 10.5% vs. 9.2% (aRR 1.05, 95%CI 0.57–1.94), and PDR was 71.5% vs. 71.7% (aRR 1.00, 95%CI 0.89–1.14). Similarly, in examinations with excellent preparation (BBPS total score 8–9), ADR was 59.7% with PolyDeep and 56.4% with conventional colonoscopy (aRR 1.07, 95%CI 0.25–1.21), SLDR was 7.1% vs. 5.8% (aRR 1.21, 95%CI 0.62–2.35), and PDR was 67.8% vs. 66.7% (aRR 1.02, 95%CI 0.90–1.16) ( Table 5 ).
Table 5 Subgroup analyses of detection outcomes between conventional and computer-assisted (PolyDeep) colonoscopy by bowel cleansing.
| Good cleansing 1 (N = 373) | aRR 2 (95%CI) | Excellent cleansing 1 (N = 454) | aRR 2 (95%CI) | |||
| Conventional (N = 173) | PolyDeep (N = 200) | Conventional (N = 243) | PolyDeep (N = 211) | |||
| ADR, adenoma detection rate; aRR, adjusted risk ratio; BBPS, Boston Bowel Preparation Scale; FIT, fecal immunochemical test; PDR, polyp detection rate; SLDR, serrated lesion detection rate. 1 Good cleansing = BBPS total score 6–7; excellent cleansing = BBPS total score 8–9. 2 Poisson regression models with log link and robust variance. | ||||||
| Detection rates, n (%) | ||||||
| ADR | 103 (59.5) | 124 (62.0) | 1.06 (0.89–1.23) | 137 (56.4) | 126 (59.7) | 1.07 (0.25–1.21) |
| SLDR | 16 (9.2) | 21 (10.5) | 1.05 (0.57–1.94) | 14 (5.8) | 15 (7.1) | 1.21 (0.62–2.35) |
| PDR | 124 (71.7) | 143 (71.5) | 1.00 (0.89–1.14) | 162 (66.7) | 143 (67.8) | 1.02 (0.90–1.16) |
Discussion
In this multicenter RCT, we evaluated whether PolyDeep-assisted colonoscopy improves ADR compared with conventional colonoscopy in the Galician population-based CRC screening program. In a setting characterized by high baseline colonoscopy quality and procedures performed by expert endoscopists, PolyDeep did not confer a statistically significant improvement in ADR or in any secondary end points. These findings indicate that, in high-performance screening environments, CADe may not provide a meaningful incremental benefit over high-quality conventional colonoscopy.
Extensive evidence from randomized trials and meta-analyses has demonstrated that CADe systems generally improve ADR across different clinical settings and study designs 30 31 32 33 34 35 36 . A large meta-analysis reported a relative ADR increase of approximately 20% with CADe assistance, which decreased to 19% when analyses were restricted to expert endoscopists 31 . Similarly, a methodological review found significant ADR improvements in 72% of parallel-design studies compared with 33% of tandem-design studies 30 , supporting the design choice of the present trial. However, the magnitude of benefit consistently appears smaller for highly experienced endoscopists, which aligns closely with our results, where no statistically significant improvement in ADR was observed.
Regarding serrated lesion detection, we found no significant differences between study groups, in agreement with previous pooled analyses showing only modest or nonsignificant effects of CADe on SLDR, except for specific platforms such as ENDOAID 31 . Similarly, we observed only a negligible absolute increase in polyp detection (0.8%), contrasting with larger effects reported in some trials. For example, the ACCENDO-Colo study reported a 9.7% absolute increase in ADR and a 13% increase in PDR with CADe 37 . However, both the conventional and CADe arms in that study exhibited substantially lower baseline ADRs than in our cohort, supporting the hypothesis that CADe provides greater benefit in settings with lower baseline detection performance. In addition, recent evidence suggests that the effect of CADe on adenoma detection may be maintained across different levels of bowel preparation quality. In a recent study, it was reported that CADe-assisted colonoscopy improved ADR regardless of whether bowel preparation was rated as good (i.e. 6–7) or excellent (i.e. 8–9) 38 . In our study, a higher ADR was also observed in the CADe group compared with conventional colonoscopy, both in cases with good bowel preparation (62.0% vs. 59.5%) and excellent bowel preparation (59.7% vs. 56.4%), but with no statistically significant differences between groups. These findings support the notion that the potential effect of CADe may be preserved across different levels of bowel cleanliness.
All endoscopists in the study were experienced participants in a population-based CRC screening program and consistently achieved high baseline ADRs, with a median exceeding 60%. To further explore the impact of baseline performance, endoscopists were dichotomized into high and low detectors using a threshold of 61.4%. While we recognize that this threshold is numerically specific to our cohort, it was derived from the median ADR of the participants. This approach is consistent with prior literature using median-based splits to evaluate whether CADe benefit varies according to an operator’s underlying detection rate 20 . This high baseline performance may partly explain the discrepancy between our findings and those of previous studies, in which larger effects of CADe have often been reported, particularly in single-center settings or among less experienced endoscopists. Our original sample size assumptions (i.e. 50% ADR in the conventional group and 60% in the CADe group) were therefore not met. As the observed baseline ADR was already close to the expected CADe performance, the study may have been underpowered to detect small incremental improvements. Although the difference in ADR between PolyDeep-assisted and conventional colonoscopy did not reach statistical significance (60.8% vs. 57.7%; aRR 1.05, 95%CI 0.94–1.17), a 3.1% absolute increase was observed in the CADe group. The study was powered to detect a 10% absolute difference based on prior evidence 39 40 ; however, the higher-than-expected ADR in the control group may have limited the ability to detect smaller, yet potentially clinically relevant differences. This effect size was selected a priori based on previously reported assumptions in ADR with CADe systems, which accounted for a 10% difference in the CADe-DB trial performed with the EndoScreener system 39 . It is important to note that recent clinical trials of commercial CADe systems set their assumption ranging between an 8% and 15% difference. In this context, the assumed effect size may have been optimistic, particularly in a high-performance screening setting with experienced endoscopists and high baseline ADRs, where the true effect of CADe is likely to be smaller. Therefore, the absence of statistically significant differences should not be interpreted as definitive evidence of no benefit. In addition, prior exposure of participating endoscopists to the PolyDeep system in earlier trials 28 may have influenced baseline detection performance. In this context, a reduced novelty effect and familiarity with the system’s detection patterns may have contributed to improved baseline detection skills and have further attenuated the observable incremental benefit of CADe. It is important to note that our findings are specific to the PolyDeep system and should not be generalized to other CADe platforms, as differences in algorithm design and performance characteristics may influence detection outcomes.
Although current evidence supports the use of CADe systems to improve adenoma detection, uncertainty remains regarding their impact on clinically meaningful outcomes. Current guidelines from the American Gastroenterological Association do not recommend either for or against the routine use of CADe, while the European Society of Gastrointestinal Endoscopy provides only a weak recommendation, largely due to the lack of robust evidence on hard clinical end points 21 22 .
It is important to note that, in addition to prior familiarity with the CADe system, endoscopists were not blinded to group allocation. This is inherent to CADe-assisted procedures and may represent a source of performance bias. Although all procedures were performed by experienced endoscopists with consistently high baseline detection rates, the potential influence of allocation awareness on endoscopic performance cannot be excluded. However, other studies including lower-performing endoscopists may observe more potential benefit 28 . Most lesions detected in the current study were diminutive. Although PolyDeep was trained on lesions of varying sizes and morphologies, it was not specifically optimized for the detection of diminutive lesions. This may have further limited its potential to enhance detection beyond expert visual assessment, particularly in a cohort where small lesions predominated.
Beyond lesion detection alone, emerging computer-aided quality (CAQ) systems assess procedural quality metrics such as withdrawal speed, mucosal exposure, and endoscope motion 33 . PolyDeep was not evaluated in combination with CAQ technology in this trial. Evidence from randomized trials suggests that combining CADe with mucosal exposure devices or quality feedback systems can further enhance ADR 41 . Whether such combined strategies could improve outcomes in high-performance screening settings warrants future investigation.
PolyDeep was based on the YOLOv3 architecture, which represented state-of-the-art detection technology at the time of study design. Although newer iterations (e.g. YOLOv8) are now available, their performance has not been validated in RCTs during real-time colonoscopy procedures. Further clinical trials are needed to evaluate the clinical performance of these new technologies in newer systems. False-positive prompts were not systematically recorded in this study and therefore could not be analyzed, limiting the assessment of the overall clinical impact of PolyDeep. However, information regarding false positives was recorded in a previous clinical trial 27 .
Our clinical trial has several important strengths. First, it was designed as an RCT involving expert endoscopists with consistently high baseline ADRs, all of whom were active participants in a population-based CRC screening program, ensuring a highly relevant and rigorous clinical context. Second, its multicenter design enhances the external validity of the findings by incorporating diverse clinical environments and a heterogeneous group of endoscopists. Third, the study was conducted specifically within an organized screening population, a setting in which CADe systems have been comparatively less evaluated, thereby addressing an important gap in the current literature. Fourth, the use of ADR as the primary end point (arguably the most clinically meaningful quality indicator in colonoscopy, given its direct association with CRC incidence and interval cancer risk) provides robust and clinically relevant evidence regarding the real-world value of CADe assistance. Fifth, the large sample size and strict methodological framework allowed us to draw reliable conclusions about the true incremental benefit of CADe in this high-performance setting.
In conclusion, PolyDeep-assisted colonoscopy did not improve ADR in the context of screening colonoscopies performed by expert endoscopists. Subgroup analyses consistently demonstrated no significant advantage of CADe across different clinical contexts, endoscopist performance levels, bowel preparation quality, or lesion characteristics. While CADe remains a promising tool for enhancing colonoscopy quality in lower-performance or training settings 42 , its incremental benefit appears limited in mature, high-quality screening programs.
Footnotes
Conflict of Interest The authors declare that they have no conflict of interest.
Contributorsʼ Statement Pedro Davila-Piñón: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Astrid I Díez-Martín: Conceptualization, Data curation, Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Alba Nogueira-Rodríguez: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing - original draft, Writing - review & editing. Florentino Fdez-Riverola: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing. Daniel Glez-Peña: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing. Miguel Reboiro-Jato: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing. Luisa De Castro: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Daniel Fernández-De Castro: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Pablo Vega: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Miguel Telmo Galovart- Araguas: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Santiago Soto: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Sara Alonso-Lorenzo: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Alfonso Martinez-Turnes: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Noel Pin: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Sara Zarraquiños: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. David Remedios: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Cristina Sánchez-Gomez: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Raquel Souto-Rodríguez: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Franco Baiocchi: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. María José Iglesias-Varela: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Alejandro Ledo: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Coral Tejido-Sandoval: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Laura Rivas: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Natalia García-Morales: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Antonio Rodríguez-De Jesus: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Manuel Puga: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. María Belén Castiñeira-Domínguez: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Nereida Fernández-Fernández: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Arantza Germade-Martínez: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Enrrique Gonzalez-De La Ballina: Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Indhira Pérez-Medrano : Investigation, Methodology, Resources, Validation, Visualization, Writing - original draft, Writing - review & editing. Hugo López-Fernández: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing. Joaquin Cubiella: Conceptualization, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing.
References
- 1.Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin. 2023;73:17–48. doi: 10.3322/caac.21763. [DOI] [PubMed] [Google Scholar]
- 2.Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74:229–263. doi: 10.3322/caac.21834. [DOI] [PubMed] [Google Scholar]
- 3.Leslie A, Carey FA, Pratt NR, et al. The colorectal adenoma–carcinoma sequence. Br J Surg. 2002;89:845–860. doi: 10.1046/j.1365-2168.2002.02120.x. [DOI] [PubMed] [Google Scholar]
- 4.Bretthauer M, Løberg M, Wieszczy P, et al. Effect of colonoscopy screening on risks of colorectal cancer and related death. N Engl J Med. 2022;387:1547–1556. doi: 10.1056/NEJMoa2208375. [DOI] [PubMed] [Google Scholar]
- 5.Castells A, Quintero E, Bujanda L, et al. Effect of invitation to colonoscopy versus faecal immunochemical test screening on colorectal cancer mortality (COLONPREV): a pragmatic, randomised, controlled, non-inferiority trial. Lancet. 2025;405:1231–1239. doi: 10.1016/S0140-6736(25)00145-X. [DOI] [PubMed] [Google Scholar]
- 6.Nasr E, Al-Hamid Z, Younan MH, et al. Modernizing colorectal cancer care with artificial intelligence: real-time detection, radiomics, and digital pathology. Cureus. 2025 doi: 10.7759/cureus.95178. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Rasmussen SL, Pedersen L, Torp-Pedersen C, et al. Post-colonoscopy colorectal cancer and the association with endoscopic findings in the Danish colorectal cancer screening programme. BMJ Open Gastroenterol. 2025;12 doi: 10.1136/bmjgast-2024-001692.e001692 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Tiankanon K, Aniwan S. What are the priority quality indicators for colonoscopy in real-world clinical practice? Dig Endosc. 2024;36:30–39. doi: 10.1111/den.14635. [DOI] [PubMed] [Google Scholar]
- 9.Rex DK, Anderson JC, Butterly LF, et al. Quality indicators for colonoscopy. Am J Gastroenterol. 2024;119:1754–1780. doi: 10.1016/j.gie.2024.04.2905. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Kaminski MF, Wieszczy P, Rupinski M, et al. Increased rate of adenoma detection associates with reduced risk of colorectal cancer and death. Gastroenterology. 2017;153:98–105. doi: 10.1053/j.gastro.2017.04.006. [DOI] [PubMed] [Google Scholar]
- 11.Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med. 2014;370:1298–1306. doi: 10.1056/NEJMoa1309086. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zhao S, Wang S, Pan P, et al. Magnitude, risk factors, and factors associated with adenoma miss rate of tandem colonoscopy: a systematic review and meta-analysis. Gastroenterology. 2019;156:1661–1.674E14. doi: 10.1053/j.gastro.2019.01.260. [DOI] [PubMed] [Google Scholar]
- 13.Marcondes FO, Gourevitch RA, Schoen RE, et al. Adenoma detection rate falls at the end of the day in a large multi-site sample. Dig Dis Sci. 2018;63:856–859. doi: 10.1007/s10620-018-4947-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Dimopoulou K, Spinou M, Ioannou A, et al. Artificial intelligence in colonoscopy: enhancing quality indicators for optimal patient outcomes. World J Gastroenterol. 2025;31 doi: 10.3748/wjg.v31.i40.111499.111499 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Cristea DM, Onita D, Iantovics LB. Biomimetic transfer learning-based complex gastrointestinal polyp classification. Biomimetics (Basel) 2025;10:699. doi: 10.3390/biomimetics10100699. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Nam JG, Hwang EJ, Kim J, et al. AI improves nodule detection on chest radiographs in a health screening population: a randomized controlled trial. Radiology. 2023;307 doi: 10.1148/radiol.221894.e221894 [DOI] [PubMed] [Google Scholar]
- 17.Kunar MA, Watson DG. Framing the fallibility of computer-aided detection aids cancer detection. Cogn Res Princ Implic. 2023;8:30. doi: 10.1186/s41235-023-00485-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Maas MHJ, Rath T, Spada C, et al. A computer-aided detection system in the everyday setting of diagnostic, screening and surveillance colonoscopy: an international, randomized trial. Endoscopy. 2024;56:843–850. doi: 10.1055/a-2328-2844. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Seager A, Sharp L, Neilson LJ, et al. Polyp detection with colonoscopy assisted by the GI Genius artificial intelligence endoscopy module compared with standard colonoscopy in routine colonoscopy practice (COLO-DETECT): a multicentre, open-label, parallel-arm, pragmatic randomised controlled trial. Lancet Gastroenterol Hepatol. 2024;9:911–923. doi: 10.1016/S2468-1253(24)00161-4. [DOI] [PubMed] [Google Scholar]
- 20.Mangas-Sanjuan C, de-Castro L, Cubiella J, et al. Role of artificial intelligence in colonoscopy detection of advanced neoplasias: a randomized trial. Ann Intern Med. 2023;176:1145–1152. doi: 10.7326/M22-2619. [DOI] [PubMed] [Google Scholar]
- 21.Sultan S, Shung DL, Kolb JM, et al. AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy. Gastroenterology. 2025;168:691–700. doi: 10.1053/j.gastro.2025.01.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Bretthauer M, Ahmed J, Antonelli G, et al. Use of computer-assisted detection (CADe) colonoscopy in colorectal cancer screening and surveillance: European Society of Gastrointestinal Endoscopy (ESGE) Position Statement. Endoscopy. 2025;57:667–673. doi: 10.1055/a-2543-0370. [DOI] [PubMed] [Google Scholar]
- 23.Ahmad A, Wilson A, Haycock A, et al. Evaluation of a real-time computer-aided polyp detection system during screening colonoscopy: AI-DETECT study. Endoscopy. 2022;55:313–319. doi: 10.1055/a-1966-0661. [DOI] [PubMed] [Google Scholar]
- 24.Nogueira Rodríguez A, Daniel González D, Hugo López Fernández PD. Deep learning techniques for computer-aided diagnosis in colorectal cancer. Vigo, Pontevedra, Spain: University of Vigo; 2022. https://www.investigo.biblioteca.uvigo.es/xmlui/handle/11093/3769 https://www.investigo.biblioteca.uvigo.es/xmlui/handle/11093/3769
- 25.Nogueira-Rodríguez A, Reboiro-Jato M, Glez-Peña D, et al. Performance of convolutional neural networks for polyp localization on public colonoscopy image datasets. Diagnostics (Basel) 2022;12:898. doi: 10.3390/diagnostics12040898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Nogueira-Rodríguez A, Domínguez-Carbajales R, Campos-Tato F, et al. Real-time polyp detection model using convolutional neural networks. Neural Comput Appl. 2022;34:10375–10396. [Google Scholar]
- 27.Davila-Piñón P, Pedrido T, Díez-Martín AI, et al. PolyDeep Advance 1: clinical validation of a computer-aided detection system for colorectal polyp detection with a second observer design. Diagnostics (Basel) 2025;15:458. doi: 10.3390/diagnostics15040458. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Davila-Piñón P, Díez Martín AI, Nogueira-Rodríguez A, et al. Clinical evaluation of PolyDeep, a computer-aided detection system: a multicenter randomized tandem colonoscopy trial. Diagnostics (Basel) 2025;15:2751. doi: 10.3390/diagnostics15212751. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Nogueira-Rodríguez A, Glez-Peña D, Reboiro-Jato M, et al. Negative samples for improving object detection – a case study in AI-assisted colonoscopy for polyp detection. Diagnostics (Basel) 2023;13:966. doi: 10.3390/diagnostics13050966. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Lee MCM, Parker CH, Liu LWC, et al. Impact of study design on adenoma detection in the evaluation of artificial intelligence-aided colonoscopy: a systematic review and meta-analysis. Gastrointest Endosc. 2024;99:676–6.87E18. doi: 10.1016/j.gie.2024.01.021. [DOI] [PubMed] [Google Scholar]
- 31.Makar J, Abdelmalak J, Con D, et al. Use of artificial intelligence improves colonoscopy performance in adenoma detection: a systematic review and meta-analysis. Gastrointest Endosc. 2025;101:68–8.1E9. doi: 10.1016/j.gie.2024.08.033. [DOI] [PubMed] [Google Scholar]
- 32.Fernandes C, Estevinho M, Marques Cruz M, et al. Adenoma detection rate by colonoscopy in real-world population-based studies: a systematic review and meta-analysis. Endoscopy. 2025;57:49–61. doi: 10.1055/a-2382-5795. [DOI] [PubMed] [Google Scholar]
- 33.Cold KM, Vamadevan A, Vilmann AS, et al. Computer-aided quality assessment of endoscopist competence during colonoscopy: a systematic review. Gastrointest Endosc. 2024;100:167–1760. doi: 10.1016/j.gie.2024.04.004. [DOI] [PubMed] [Google Scholar]
- 34.Kumar S, Maheshwari M, Aleem S, et al. Novel artificial intelligence systems in detecting adenomas in colonoscopy: a systemic review and network meta-analysis. Clin Transl Gastroenterol. 2025;16 doi: 10.14309/ctg.0000000000000904.e00904 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Spadaccini M, Hassan C, Mori Y, et al. Artificial intelligence and colorectal neoplasia detection performances in patients with positive fecal immunochemical test: meta-analysis and systematic review. Dig Endosc. 2025;37:815–823. doi: 10.1111/den.15034. [DOI] [PubMed] [Google Scholar]
- 36.Rondonotti E, Di Paolo D, Rizzotto ER, et al. Efficacy of a computer-aided detection system in a fecal immunochemical test-based organized colorectal cancer screening program: a randomized controlled trial (AIFIT study) Endoscopy. 2022;54:1171–1179. doi: 10.1055/a-1849-6878. [DOI] [PubMed] [Google Scholar]
- 37.Spada C, Salvi D, Ferrari C, et al. A comprehensive RCT in screening, surveillance, and diagnostic AI-assisted colonoscopies (ACCENDO-Colo study) Dig Liver Dis. 2025;57:762–769. doi: 10.1016/j.dld.2024.12.023. [DOI] [PubMed] [Google Scholar]
- 38.Muccichini E, Rondonotti E, Spadaccini M, et al. Bowel preparation quality and computer-aided detection of adenomas during colonoscopy. Dig Liver Dis. 2025;57:2174–2179. doi: 10.1016/j.dld.2025.08.072. [DOI] [PubMed] [Google Scholar]
- 39.Wang P, Liu X, Berzin TM, et al. Effect of a deep-learning computer-aided detection system on adenoma detection during colonoscopy (CADe-DB trial): a double-blind randomised study. Lancet Gastroenterol Hepatol. 2020;5:343–351. doi: 10.1016/S2468-1253(19)30411-X. [DOI] [PubMed] [Google Scholar]
- 40.Mangas-Sanjuan C, Santana E, Cubiella J, et al. Variation in colonoscopy performance measures according to procedure indication. Clin Gastroenterol Hepatol. 2020;18:1216–122300. doi: 10.1016/j.cgh.2019.08.035. [DOI] [PubMed] [Google Scholar]
- 41.Beran A, Nayfeh T, Ramai D, et al. Artificial intelligence-assisted colonoscopy with or without mucosal exposure device for detection of colorectal adenomas: a meta-analysis. Endosc Int Open. 2025;13 doi: 10.1055/a-2676-4144.a26764144 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Okumura T, Kudo S, Ide Y, et al. Long-term impact of computer-aided adenoma detection: a prospective observational study. Endoscopy. 2026;58:121–129. doi: 10.1055/a-2661-2624. [DOI] [PubMed] [Google Scholar]