THE IMPROVEMENT ON ADENOMA DETECTION RATE AND OTHER SECONDARY INDICATORS OF THE TWO REAL-TIME ARTIFICIAL INTELLIGENCES IN HIGH ADENOMA DETECTORS: A RANDOMIZED MUTI-CENTER TRIAL

Introduction: Adenoma per colonoscopy (APC) has recently been proposed as a quality measure for colonoscopy. We evaluated the impact of a novel AI system, compared to standard HD colonoscopy, for APC measurement.
Methods: This was a U.S. based, multi-center, prospective randomized trial (NCT04979962) investigating a novel AI detection system - EW10-EC02 that enables a real-time colorectal polyp detection enabled with the colonoscope (CAD-EYE™); Figure 1. Eligible average risk subjects (45 and older) undergoing screening or surveillance colonoscopy were randomized to undergo either computer-assisted colonoscopy (CAC) or conventional colonoscopy (CC). Primary outcomes were APC and positive predictive value (PPV, total number of adenomas divided by total polyps removed). Secondary outcomes were withdrawal time, ADR, sessile serrated lesion detection rate, polyp detection rate and polyp per colonoscopy.
Results: Of 1033 subjects (age: 59.1+/-9.8; 49.9% male) randomized, 510 underwent CAC vs. 523 underwent CC with no significant differences in age, gender, ethnicity, or colonoscopy indication between the 2 groups. For the primary aim, CAC led to a significantly higher APC compared to CC: 0.99± 1.6 vs. 0.85±1.5, p=0.02, Incident Rate Ratio 1.17 (1.03-1.33, p=0.02) with no significant difference in the withdrawal time: 11.28±4.59 min vs. 10.8±4.81 min; p=0.11 between the 2 groups. For the co-primary end point, the positive predictive value of a polyp being adenoma (or non-adenoma) was not inferior (<10%). There were no significant differences in ADR (46.9% vs. 42.8%), advanced adenoma (6.5% vs. 6.3%), sessile serrated lesion detection rate (12.9% vs. 10.1%) and polyp detection rate (63.9% vs 59.3%) between the 2 groups. There was a higher polyp per colonoscopy with CAC compared to CC: 1.67 ± 2.1 vs. 1.33 ± 1.8 (incidence rate ratio 1.27; 1.15-1.4; p<0.01).
Conclusion: Use of a novel AI detection system leads to a significantly higher number of adenomas per colonoscopy compared to conventional HD colonoscopy without any increase in colonoscopy withdrawal time, thus supporting use of AI-assisted colonoscopy to improve colonoscopy quality.

Figure 1. (a and b) Polyp is detected by the CADEYE system with rectangular blue box

Introduction: Randomized controlled trials have shown that computer-assisted detection (CAD) can improve adenoma detection rate (ADR). Up to 20% of lesions were still missed by CAD due to “non-visibility” on screen. This study aims to determine whether the combined use of device that increase mucosal exposure (Endocuff) and CAD could further increase the ADR,as compared to CAD alone and white light (WLI) colonoscopy. This is an on-going three-arm prospective randomized trial comparing the use of CAD (with or without endocuff) and conventional WLI colonsocopy on colorectal adenoma detection.

Method: Consecutive patients aged 40 or above undergoing elective colonoscopy were recruited. Patients with history of colonic resection or cancer, familial colorectal cancer syndrome, bleeding tendency, severe comorbid illness, pregnancy women or unable to provide written informed consent were excluded. Eligible patients were randomized in a 1:1:1 ratio to receive CAD (OIP-1, Olympus) with Endocuff (CAD-Endocuff; both by Olympus), CAD alone (CAD) or WLI under high definition colonoscope. Procedures were performed by either experienced endoscopist or trainees with more than 2 years of colonoscopy experience as well as hands-on training on CAD. Analysis was based on per-protocol analysis with ADR as primary outcome.

Result: This is the planned interim analysis of the first 300 patients with complete colonsocopy (Table 1). The mean age of patients was 64.7 (SD 10.3) years and there was 169 (56.3%) male. 58 (19.3%) patients underwent colonoscopy for screening. Among the three groups, the CAD-Endocuff group had the highest ADR (65.7%), PDR (80.8%), SDR (41.4%) and AADR (14.1%). The WLI group had the lowest ADR (46.8%), PDR (59.6%), SDR (23.4%) and AADR (10.5%), with the CAD group in the intermediate (Table 1 and Figure 1). The mean number of adenoma, polyp and advanced adenoma detected per patient followed the same trend and was highest in the CAD-Endocuff group and lowest in the WLI group (Figure 1). On this interim analysis, there was already a significant difference in the ADR (ANOVA; P=0.03), PDR (P<0.01) and SDR (P=0.02) among the three groups. Specficially, CAD-Endocuff group was significantly better than than WLI group in ADR (P=0.01; after Bonferrini correction), PDR (P<0.01), SDR (P=0.01) and the mean number of adenoma (P=0.01) or polyp (P<0.01) detected per patient. CAD alone, however, was only better than WLI in PDR (P=0.02) and mean number of polyp detected (P=0.01).

Conclusion: Interim analysis of this ongoing randomized trial already showed a promising effect of the combined use of CAD and endocuff on enhancing colonic lesions detection, including adenoma and serrated lesions. Further patient enrollment are ongoing to delineate the the potential incremental benefits of adding mucosal exposure device to CAD on colorectal lesion detection.

Background: The subjective diagnosis of malignant biliary strictures during cholangioscopy through visual interpretation of images is often inaccurate and shows high interobserver variation. Aims: Our objective was to develop and validate machine learning (ML) algorithms using both cholangioscopy image features and clinical data to diagnose malignant biliary strictures for real-time decision support. Methods: We used data from a multicenter cohort of patients who underwent cholangioscopy for evaluation of indeterminate biliary strictures. The final diagnosis was based on pathological examination for malignant strictures and a minimum of 12 months follow-up without detection of cancer for benign strictures. We excluded patients who underwent cholangioscopy for biliary stone disease, or those that underwent cholangioscopy using the first-generation cholangioscope or those with no established final diagnosis. We trained a convolutional neural network (CNN) to detect the presence of a stricture in cholangioscopy images and generate summary image features. We combined these image features with clinical variables of importance identified through logistic regression to identify the presence of malignancy. Data was split by centers into training, development, and testing sets to test external validity (Figure 1). We computed the estimates and 95% confidence intervals of the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive and negative predictive values (PPV and NPV). Results: We analyzed data, including 667,538 images, from 502 patients (mean age 64yr, F 45%, 55.8% malignant strictures), from 27 centers (15 North America, 7 Asia, 2 Europe, 2 Australia, 1 South America) (Figure 1). Patients with malignant strictures were more likely to be males (60% vs 49%), of older age (60.1 yr vs 67.6 yr), have jaundice (67% vs 33%), weight loss (45% vs 15%) and higher AST (107 vs 73) and ALT (147 vs 131) as compared to those with benign strictures (p<0.05). The AUC for the CNN to detect presence of stricture in cholangioscopy images was 0.92. The AUC for the CNN for predicting malignancy using a combination of image and clinical features was 0.82 as compared to 0.62 when using image-derived features alone (Table 1). Clinical features improved the logistic regression model on all measures of discrimination when compared to image-derived features alone. Discussion: Our findings show external validity of ML models to diagnose malignant biliary strictures for decision support during cholangioscopy. Our models combine clinical variables and image analysis to inform the endoscopist of the probability that a stricture is malignant. Thus, they support decision-making during cholangioscopy to allow timely management of malignant strictures and to prevent unwarranted surgery.

Background and Aims
Endoscopic scores for Ulcerative Colitis (UC) such as the Mayo Endoscopic Subscore (MES) and the Ulcerative Colitis Endoscopic Index of Severity (UCEIS) are objectively categorize the severity of the disease based on presence or absence of endoscopic findings. Therefore, they may not reflect the range of clinical severity within each category. But, Inflammatory bowel disease expert endoscopists do not only categorize the severity, but rather diagnose the overall impression of degree of inflammation. Therefore, the purpose of this study was to develop an AI that can accurately represent the complexed assessment of endoscopic severity of UC by expert endoscopists, similarly to Visual Analogue Scale (VAS).
Methods
To enable the AI to perform continuous evaluations of inflammation in line with the strategy used by IBD expert endoscopists, we did not utilize scores determined from images using MES or UCEIS for physician data. Rather, we incorporated data for the relationships identified by IBD expert endoscopists who compared the severity of paired images. This study was conducted using a method that incorporates data on relationships identified by comparing the severity of paired images created from 59595 endoscopic images by an IBD expert endoscopist into a Ranking- Convolutional Neural Network, and then the severity was then expressed on a scale called UC Endoscopic Gradation Scale (UCEGS) rather than a score. Using 4,000 images for which the MES had been assessed beforehand by an IBD expert endoscopist, correlation coefficients were calculated to ensure that there were no inconsistencies in assessments of severity made using results of this novel AI diagnosed UCEGS and the current MES score. The correlation coefficients of the means of the UCEGS results for the 50 test images evaluated by the five IBD expert endoscopists and the novel AI were also calculated. Finally, To accurately determine the severity and area of colorectal inflammation associated with UC, we developed a user interface in which the UCEGS results diagnosed by this novel AI was incorporated into the schema for the colon.
Results
Spearman's correlation coefficient between MES and AI-diagnosed UCEGS was approximately 0.89, indicating a strong positive correlation for the order of severity between the AI-diagnosed UCEGS and MES. Correlation coefficient between IBD expert endoscopists and the AI of the evaluation results were all higher than 0.95 (P<0.01).
Conclusion
In this study, we developed a novel AI that does not rely on conventional scoring methods but instead aims to leverage the intelligence of IBD expert endoscopists when evaluating the disease status of UC. This AI quantifies inflammation as a gradient, allowing for an automated visualization of the expert endoscopist's assessment of mucosal inflammation.

Introduction: Artificial intelligence-based polyp detection system (CADe) is known to improve polyp detection, especially when used by endoscopists with low adenoma detection rate (ADR). However, few clinical studies have evaluated the usefulness of CADe in high adenoma detectors (ADR>35%). We aimed to assess the benefit of our developed CADe (Deep-GI) and the commercially available CADEYE^® in high adenoma detectors by using white-light colonoscopy (WLE) as a control.
Methods: We recruited average-CRC-risk individuals aged 50-75 who underwent screening colonoscopy at our three referral centers from Febuary 2022 to October 2022. All subjects were randomly assigned to perform colonoscopy under Deep-GI, CADEYE, or WLE with a 1:1:1 ratio. All trainees and staff with a recorded ADR> 35% participated in the study. Both Deep-GI and CADEYE systems have real-time notifications with bounding boxes and voice alarms projected on the endoscopy monitor. CADe was activated in Deep-GI and CADEYE groups before colonoscope insertion. The primary outcome was the ADR. Secondary outcomes were the proximal adenoma detection rate (pADR), advanced adenoma detection rate (AADR), and the number of adenomas/proximal adenomas/advanced adenomas per colonoscopy (APC, pAPC, and AAPC, respectively). The differences in measured outcomes were compared to controls using a linear regression model. Bonferroni’s method was used to adjust for multiple comparisons.
Results: A total of 956 participants (38% male) with a mean age of 64±9.1 were randomized to Deep-GI (n=318), CADEYE (n=319), and control (n=319). Patients and procedural characteristics were comparable among the 3 groups (Table 1). The ADR in the control, CADEYE, and Deep-GI groups were 40.4%, 47.6%, and 55%, respectively (p=0.001). The pADR were 23.8%, 30.4%, and 38.7%, respectively (p<0.001). AADR were 5.3%, 9.4%, and 8.8%, respectively (p=0.12) (Figure 1A). After multiple comparison adjustments, ADR and pADR in Deep-GI group were significantly higher than those of the controls (p<0.05 in all comparisons). The comparisons for ADR and pADR between CADEYE vs. controls, and Deep-GI vs. CADEYE showed no statistically significant differences. The APC in the control, CADEYE, and Deep-GI groups were 0.67, 0.98, and 1.13, respectively (p<0.001). The pAPC were 0.35, 0.49, and 0.64, respectively (p<0.001). Deep-GI and CADEYE showed significantly higher APC and pAPC but not AAPC when compared with the controls (p<0.05 in all comparisons) (Figure 1B). In comparison, Deep-GI vs. CADEYE showed no statistically significant differences in APC, pAPC, and AAPC.
Conclusion: In high adenoma detectors (ADR>35%), Deep-GI system improved their performance by significantly increasing ADR and pADR over WLE. The number of APC and pAPC also significantly increased when both CADes were activated, although many of them were non-advanced adenomas.