This study aimed to evaluate the effectiveness of AI-assisted colonoscopy in a real-world setting. AI-assisted colonoscopy resulted in significantly higher ADRs and APCs compared to standard colonoscopy (ADR, 35.9% vs. 26.4%; APC, 0.69 vs. 0.43). However, the detection rates of advanced adenomas and SSLs did not significantly differ between the two groups.
Multiple RCTs have evaluated the effectiveness of AI-assisted colonoscopy, and a recent meta-analysis of 21 RCTs confirmed that AI-assisted colonoscopy increases the ADR [18]. However, two main issues limit the applicability of these findings in real-world clinical practice. First, the lack of blinding may have introduced a bias favoring the AI group. Although one double-blind RCT using a sham control showed increased ADR [9], there remains a lack of sufficient blinded studies. Second, these RCTs were conducted under strict conditions that may not reflect real-world factors affecting colonoscopy quality, such as examiner fatigue or time pressure. To address this, non-RCT observational studies have been conducted in real-world clinical practice settings. A meta-analysis by Patel et al. found no significant improvement in polyp detection using AI in real-world settings [13], while another systematic review reported a small but statistically significant improvement in ADR (36.3% vs. 35.8%, p = 0.04) [14]. Considering the inconsistent outcomes of previous non-RCTs, this study provides substantial evidence supporting the role of AI in clinical practice, demonstrating an increase in ADR in the AI-assisted colonoscopy group. Several other studies using AI-assisted colonoscopy support our findings. For instance, a single-center retrospective study showed a significant improvement in ADR with AI assistance compared to conventional methods (47.9% vs. 38.5%, p = 0.03) [19]. Similarly, a study from Korea found a significantly higher ADR in the AI-assisted group compared to the standard colonoscopy group (45.1% vs. 38.8%, p = 0.010) [20]. An additional finding in our study was that the overall SSLDR increased from 3.0 to 5.5%, although the difference was not statistically significant (p = 0.076), likely due to the limited sample size. However, considering that the SSLDR in studies without CADe systems typically ranges from 2 to 2.5% [21, 22], our finding is noteworthy. Similar to our results, a recent study using the CADe system reported an increase in the SSLDR from 2.5 to 5.7% (p = 0.001) [20], although another study did not show significant changes [19], indicating ongoing controversy. Unlike ADR, there are no universally accepted benchmarks for SSLDR, despite SSLs accounting for 15–30% of CRC cases [23]. Due to their flat morphology, SSLs are particularly challenging to detect, highlighting the need for further investigation into the role of AI in improving SSLDR.
In contrast, other studies have reported that AI-assisted colonoscopy does not significantly improve ADR in real-world settings [10,11,12]. Although our study showed a positive effect of AI-assisted colonoscopy on ADR, this result may have been influenced by the fact that the endoscopists were aware that their colonoscopy performance was being monitored, potentially leading to changes in their behavior. Moreover, more than two-thirds of the detected lesions were < 5 mm in size. Therefore, the increased ADR associated with AI-assisted colonoscopy may result from the increased detection of diminutive adenomas. Considering that a long-term prospective study showed that most diminutive polyps exhibit slow growth and typically follow a benign course [24], this raises questions regarding their clinical relevance. This finding emphasizes the need for a long-term follow-up study to clarify the effect of AI-assisted colonoscopy on the prevention of CRC. Moreover, there was no significant difference in advanced adenoma detection between the AI-assisted colonoscopy and standard colonoscopy groups, likely because these lesions are easier to detect without AI assistance. Another limitation of CADe is the frequent occurrence of false positives due to mucosal folds or residuals, which can lead to unnecessary resections, longer procedure times, and increased costs.
Although CADe is a valuable tool in colonoscopy, the role of endoscopists remains crucial, as polyp detection depends on their ability to ensure adequate mucosal exposure and identify abnormalities [25]. Subgroup analysis in this study showed no significant difference in ADRs between AI-assisted colonoscopy and standard colonoscopy in the low-detector group (20% vs. 19.3%, p = 0.945). In contrast, the high-detector group exhibited a numerically higher ADR (45.8% vs. 39.2%, p = 0.162) and a statistically significant increase in SSLDR (8.0% vs. 3.3%, p = 0.043) with AI-assisted colonoscopy. This suggests that the additional benefits of CADe with increased ADRs may be limited if the endoscopist does not perform a meticulous examination. However, the observed differences in ADR according to endoscopist performance in our study were based on only two endoscopists per group, providing insufficient evidence to support the generalization of this finding and thus should be interpreted with caution. Further large-scale studies involving more endoscopists are necessary to confirm and validate this observation. Regarding endoscopy experience, CADe was associated with increased ADR in the trainee group (51.5% vs. 27.2%, p = 0.023) in the subgroup analysis, although the small sample size may have affected the reliability of the results. A previous study indicated that CADe helps trainees improve their ADR, thereby reducing the performance gap between trainees and experts [26]. This suggests that AI can reduce the learning curve of trainees and lead to better screening outcomes, as it can assist trainees in lesion recognition by pointing out areas that might otherwise be overlooked by trainees during the procedure. Further studies are needed to assess the impact of AI-assisted colonoscopy across endoscopists with varying levels of experience to better understand its effects on detection and screening outcomes.
A limitation of AI-assisted colonoscopy is its inability to detect lesions in areas that are not fully exposed. Various mucosal exposure devices, such as transparent caps and Endocuff Vision, have been extensively studied for their efficacy. The Endocuff, a device designed to flatten colon folds and improve mucosal visibility, has demonstrated efficacy in several RCTs, significantly improving ADR compared to standard colonoscopy [27]. A study comparing colonoscopy with the combination of CADe and Endocuff to CADe alone found that the ADR was significantly higher in the Endocuff combination group [28]. Additionally, a recent three-arm RCT demonstrated that combining Endocuff with CADe significantly improved the ADR (58.7%) compared with CADe alone (53.8%) or standard colonoscopy (46.3%) [29]. However, the effectiveness of transparent caps in improving ADR remains controversial, with some studies showing benefits and others showing no significant effect [30]. In our study, the use of transparent caps was based on the endoscopists’ preferences, making it challenging to assess the specific effects of their combination with CADe. Further research is needed to determine the optimal combination of mucosal exposure devices and CADe to enhance the polyp detection outcomes.
This study’s non-randomized observational design enhances the applicability of its findings to real-world clinical practice. Additionally, propensity score matching reduced group bias, improving the reliability of our findings. Baseline characteristics such as age, sex, and BMI showed no significant differences, and these were further minimized in the matched cohort, indicating an effective bias control. Using a concurrent control group was another strength of this study. While historical controls are susceptible to various biases, concurrent control groups minimize differences caused by time-related changes and external factors, thereby providing greater validity.
This study had some limitations. First, it was conducted at a single center, limiting the generalizability of the results and emphasizing the need for multicenter studies to improve external validity. In addition, since this was a non-blinded study, there is a possibility of selection bias in patient allocation. However, we minimized this potential bias by applying propensity score matching to adjust for baseline characteristics between the two groups. Second, the relatively small number of endoscopists, with two of the four endoscopists having an ADR < 25%, may have influenced the findings. Including endoscopists with low baseline ADRs may have contributed to the overall increase in the ADR observed in this study. Future studies adjusting for endoscopists’ baseline ADRs will provide a clearer assessment of the true impact of AI. Third, the endoscopists’ awareness of AI usage may have prompted more careful evaluations, potentially influencing the results. Implementing a blinded design may yield more relevant findings in future studies. Furthermore, the study excluded patients with complicated colonic conditions, such as inflammatory bowel disease or polyposis syndrome, limiting the applicability of the findings to these populations. Therefore, future studies should include a broader range of patients. Moreover, the study did not establish a connection between improved ADRs and long-term patient benefits, such as reduced CRC incidence, highlighting the need for further research to evaluate these long-term outcomes. Finally, we did not perform a separate analysis of detection rates for laterally spreading tumors (LST), particularly the non-granular (LST-NG) subtype. These lesions are known to be easily missed and more difficult to achieve complete resection, making them clinically significant [31]. Therefore, the detection rate of LSTs should be considered an important component in evaluating the performance of CADe systems, underscoring the need for future large-scale studies to address this issue.
In conclusion, AI-assisted colonoscopy significantly improved ADR in real-world settings, demonstrating its potential to enhance screening quality and facilitate early polyp detection. The synergy between CADe and the endoscopist’s careful observation offers promising advancements in CRC prevention.