Artificial intelligence-based mucosa touch rate: A novel real-time quality control indicator for colonoscopy

Abstract

BACKGROUND

The effectiveness of colonoscopy is highly operator dependent, but existing quality indicators are retrospective and cannot guide performance in real time.

AIM

To develop an artificial intelligence (AI)-based mucosa touch rate (MTR) as a novel real-time quality metric for mucosal contact-induced > 2/3 visual field loss frame proportion.

METHODS

An EfficientNet-B2 model was trained on 11793 images to identify mucosa touch. The correlation between MTR and polyp detection rate (PDR) was analyzed using 162 procedural videos. A prospective, single-blind trial enrolled 2866 patients: 1366 with AI-assisted real-time MTR feedback and 1500 controls.

RESULTS

The AI model showed high accuracy for mucosa touch (100%) and ileocecal recognition (99.03%), with an area under the curve of 0.97. MTR was strongly negatively correlated with PDR (r = -0.948, P < 0.01). The AI-assisted group had a significantly higher PDR than controls (51.17% vs 40.07%, P < 0.001). Low-seniority endoscopists benefited most, with a 19.90% PDR increase (36.27% to 56.17%, P < 0.001).

CONCLUSION

The AI-based MTR is accurate and reliable, its real-time feedback improves colonoscopy quality for less experienced endoscopists and enables coaching and skill standardization.

Key Words: Colonoscopy; Artificial intelligence; Mucosa touch rate; Polyp detection rate; Adenoma detection rate

Core Tip: This study introduces an artificial intelligence-based mucosa touch rate (MTR) as a novel real-time quality indicator for colonoscopy. Using a deep learning model, MTR objectively quantifies mucosal contact during withdrawal. The results demonstrate a strong negative correlation between MTR and polyp detection rate (PDR). Prospective validation shows that real-time MTR feedback significantly improves PDR, particularly among less experienced endoscopists, highlighting its potential for real-time skill assessment and standardized training in colonoscopy quality control.

INTRODUCTION

Colorectal cancer (CRC) ranks third in terms of global incidence and second in cancer mortality, with more than 1 million new cases yearly and an increasing incidence in those under 50 years of age[1,2]. In China, it is the second most common malignancy and fourth in mortality, with increasing incidence and mortality from 2000 to 2016[3].

Colonoscopy is the main method for lower digestive tract examination[4], but the polyp miss rate (6%-27%) is strongly influenced by endoscopists’ skill[5-10], making colonoscopy quality control essential. Although the polyp detection rate (PDR) or adenoma detection rate (ADR) is widely used to evaluate colonoscopy quality[11-13], it is cumbersome to calculate and cannot be used for real-time quality assessment[14], and more critically, there is no objective real-time indicator to evaluate endoscopists’ withdrawal technique a key factor affecting mucosal visualization[15].

Recent artificial intelligence (AI) advances have extended to enhancing imaging quality and diagnostic precision through techniques such as hyperspectral and spectral imaging, as demonstrated in esophageal cancer diagnostics[16,17]. Wang et al[18] used a deep learning-based automatic polyp detection system to significantly improve endoscopists’ ADR. While most AI applications in colonoscopy focus on polyp detection or withdrawal time monitoring, few address the direct assessment of endoscopic technique[19,20]. A recent systematic review by Cold et al[21] highlighted the scarcity of AI systems capable of providing real-time feedback on withdrawal technique.

To address this issue, we introduced a new index called the mucosa touch rate (MTR). Withdrawal is the key phase for lesion detection in colonoscopy: Insufficient endoscope control during withdrawal leads to unstable endoscope movement and the endoscope tip being too close to the intestinal mucosa, resulting in endoscopic field defects (only red areas are visible). When red areas exceed 2/3 of the image, it is defined as a “mucosa touch image” (Figure 1). Excessive such images can cause incomplete visualization and small lesions to be missed. Previously, endoscopists could not quantify mucosa touch due to limited image analysis capabilities, but AI now enables analysis of procedure videos from ileocecal arrival to anal exit. The MTR is defined as the proportion of images with the endoscope tip too close to the intestinal wall during withdrawal.

Figure 1 Image classification. Annotation: In the pre-experiment, labeled images were annotated by multiple experts and then converted into a dataset. Model test precision/recall confirmed images with ≥ 2/3 mucosa touch meeting the mucosa touch image criteria.

The threshold of > 2/3 visual field loss for defining a “mucosa touch image” was empirically derived from pilot observations and aligned with previous studies assessing endoscopic visualization quality[22]. This threshold was established based on a consensus among three senior expert endoscopists (each with > 10 years of experience) during the model development phase, who determined that beyond this point, the field of view is substantially compromised, significantly increasing the risk of missing subtle lesions. It is acknowledged that such contact may be intentional during certain maneuvers (e.g., lens cleaning, lesion resection). Therefore, for the purpose of this study, MTR was calculated specifically during the withdrawal phase, excluding time periods dedicated to polypectomy or other therapeutic interventions, to better capture unintentional contact that reflects suboptimal inspection technique.

We constructed an AI-based MTR model using deep learning to quantify mucosal touch events in real time. This model was designed to classify “mucosa touch” and “ileocecal region” images (the latter being a key landmark for complete colonoscopy) with high accuracy. We subsequently conducted a prospective clinical study to validate the MTR model. This two-phase design (model construction followed by clinical validation) aimed to: (1) Confirm the accuracy and robustness of the AI-MTR model in identifying mucosa touch and calculating MTR; (2) Assess whether MTR monitoring improves colonoscopy quality; (3) Explore its differential effects on endoscopists of varying seniority; and (4) Identify subgroups of patients who may benefit most from MTR-based quality control.

In this study, AI deep learning technology was combined with colonoscopy data to construct auxiliary diagnostic models for the intelligent recognition of colonoscopy. For clinical validation, complete withdrawal phases were captured under standardized conditions [standard withdrawal speed, minimum 6-minute withdrawal time, adequate bowel preparation confirmed by a Boston bowel preparation scale (BBPS) score ≥ 6].

MATERIALS AND METHODS

Study design and phases

This study was conducted in two consecutive phases at the Endoscopy Center of the First Affiliated Hospital of Army Medical University. Phase one (model construction, preliminary study) consisted of retrospective image collection (January 2017-December 2021), AI model development, and a preliminary correlation study. In phase two (prospective clinical validation) a prospective, single-blind, self-controlled clinical trial was conducted (December 2024-June 2025). The overall study design is illustrated in Supplementary Figure 1.

Phase one: Model construction and correlation study

Participants and datasets: A retrospective set of 11793 colonoscopy images from 1073 patients (2017-2021) was acquired using Olympus EVIS LUCERA systems. Expert gastroenterologists assigned these images to six categories (Figure 1) and divided them into training (7472 images/672 patients), tuning (2312/209), and testing (2009/192) sets. To minimize bias in dataset partitioning, images were stratified by patient and randomly allocated to training, tuning, and testing sets using a computer-generated random sequence, ensuring no patient overlap across sets. To correlate MTR with PDR, 162 full procedure videos from 6 endoscopists were randomly selected. The validation videos were obtained using both Olympus EVIS LUCERA CV260 (SL)/CV290 (SL) and Fuji 4400/4450 HD processors.

Development of automated MTR detection software

The EfficientNet-B2 architecture was selected as the backbone and modified by uniformly scaling three dimensions (network width, depth, and resolution) using fixed coefficients to adapt to medical image classification. The model was trained to: (1) Recognize the ileocecal region (to confirm complete colonoscopy); and (2) Identify mucosa touch images (frames with > 2/3 visual field loss due to excessive proximity). Real-time monitoring functions were integrated; the model highlighted mucosa touch areas in red on a high-resolution monitor and issued voice alerts (latency < 0.5 seconds) at 60 fps, with ≥ 90% sensitivity for abnormality detection. Details in attached Video 1.

To minimize labeling bias, all images were independently annotated by two expert endoscopists, with discrepancies resolved by a third senior reviewer. Borderline cases (e.g., 60%-70% visual field loss) were discussed in consensus meetings to ensure consistent labeling. Data augmentation (rotation, brightness/contrast variation) was applied during training to improve robustness.

Model performance testing

Three types of tests were conducted: (1) Basic performance test: Two subsets (A/B, 52 positive/52 negative samples each) were extracted from the testing set. Subset A evaluated ileocecal region recognition, and subset B evaluated mucosa touch recognition. Metrics included sensitivity, precision, and accuracy; (2) Generalization test: The testing set was expanded threefold (156 positive/156 negative samples per test) to assess the performance on unseen data; and (3) Environmental factors test: The image brightness (0.2-2.0), contrast (0.2-2.0), and rotation angle (-330° to -30°) were adjusted to simulate clinical variations. Accuracy was measured within clinically relevant ranges (brightness 0-1.2, contrast 0.2-0.8, rotation all angles).

Correlation analysis between the MTR and the PDR

The previously developed model was used to analyze 162 unaltered and complete colonoscopy videos from 6 endoscopists at the Endoscopy Center of the First Affiliated Hospital of Army Medical University from September to December 2023 to explore the correlation between the MTR and the PDR. The endoscopists were divided into a senior group (with ≥ 5 years of practice, ≥ 1000 cases per year; n = 3) and a junior group (with < 5 years of practice, < 500 cases per year; n = 3). Each physician randomly selected 27 cases to calculate the average MTR, after which the annual PDR of each physician was statistically analyzed. It should be noted that this correlation analysis included only six endoscopists and should therefore be interpreted as an exploratory finding; validation in larger, multi-operator cohorts is needed to confirm the robustness of this relationship.

Phase two: Prospective clinical validation

Study design and workflow: This prospective, single-blind, self-controlled trial (December 2024-June 2025) deployed the AI-MTR system in two endoscopy rooms, processing video at 60 fps with < 0.5 seconds latency. Patients were naturally randomized to AI-assisted (real-time MTR feedback) or conventional groups based on room availability. Six endoscopists, stratified by seniority (high: ≥ 5 years, ≥ 1000 procedures/year; low: < 5 years, < 500 procedures/year), performed all examinations using standardized withdrawal (≥ 6 minutes, BBPS ≥ 6). The system provided real-time visual and voice alerts for mucosa touch. All procedure videos and MTR values were automatically recorded, encrypted, and uploaded for analysis. An independent, blinded team adjudicated all lesions and pathology to determine PDR and ADR.

To minimize selection bias, endoscopy rooms were assigned daily in a fixed rotation, and all endoscopists worked in both AI-equipped and non-AI rooms across different days. Case complexity and indications were evenly distributed across rooms via the hospital’s centralized booking system. Tables’ results presents baseline characteristics after propensity score matching, confirming balanced groups.

Study population

Patients with American Society of Anesthesiologists physical status classification ≤ III and who were able to provide informed consent were included. The exclusion criteria included absolute contraindications to colonoscopy, prior colon surgery, and a history of CRC/inflammatory bowel disease/familial polyposis, pregnancy, or anesthetic allergy. Patients were excluded after enrollment if their bowel preparation was inadequate (BBPS score < 6 or any segment < 2), if the colonoscopy did not reach the cecum, or if complications required early termination.

Grouping

Patients were assigned to the AI-assisted or control group based on the daily assignment of endoscopy rooms (two equipped with AI system, two without), a process termed “natural randomization”. Although not truly random, this approach mimics real-world conditions and was accounted for in statistical analyses using propensity score matching. Six endoscopists were divided into the following groups: High-seniority (H): ≥ 5 years of experience, ≥ 1000 colonoscopies/year (n = 3) and low-seniority (L): < 5 years of experience, < 500 colonoscopies/year (n = 3). All procedures followed standardized protocols.

Sample size estimation

Sample size was determined on the basis of the PDR (primary outcome): Using preliminary data [low-seniority PDR: 27% (control) vs 38% (experimental)], 80% power, and α = 0.025 (one-sided), 313 cases/group were needed (accounting for 10% attrition). Additionally, according to the American Gastroenterological Association, reliable ADR analysis requires ≥ 250 cases/endoscopist, totaling 3000 cases (1500/group). The initial target enrollment was set at 3000 cases. However, only 2866 cases were ultimately included, which was attributed to the job transfer of one low-seniority endoscopist, making it impossible to collect complete data as planned.

Statistical analysis

Model construction phase: Continuous variables are expressed as the mean ± SD and were compared using the independent samples t-test if normally distributed; otherwise, the Mann-Whitney U test was applied. Categorical data are presented as n (%) and were analyzed using the χ² test. The correlation between MTR and PDR was assessed via Pearson’s correlation analysis. A two-sided P value < 0.05 was considered to indicate statistical significance. Analyses were performed using IBM SPSS Statistics (V26.0). Prospective clinical validation phase: The primary outcomes (PDR and ADR) were compared between groups using the χ² test. To adjust for potential confounders, propensity score matching was applied in the main analysis, and generalized linear mixed-effects models were used for sensitivity analysis, with endoscopists included as random effects. Statistical significance was defined as a two-sided P value < 0.05. All analyses were conducted using R software.

RESULTS

Phase one: Model performance

Basic performance testing: Two subsets (A and B) were randomly selected from the basic performance test set of the 2009 colonoscopy images, each with 52 positive and 52 negative samples. Subset A used ileocecal section images as positive samples and non-ileocecal section images as negative samples; subset B took mucosa touch images as positive samples and non-mucosa touch images as negative samples. The test results are shown in Table 1, and the receiver operating characteristic (ROC) curves are shown in Figure 2. The area under the ROC curve was 0.97 (Figure 2), indicating strong discriminative ability.

Figure 2 Receiver operating characteristic curve of the artificial intelligence-mucosa touch rate model. ROC: Receiver operating characteristic.

Table 1 Basic performance test.

Targets	Test subset A (95%CI)	Test subset B (95%CI)
Sensitivity (%)	98.07 (97.81, 98.33)	100.00 (100.00, 100.00)
Precision (%)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
Accuracy (%)	99.03 (99.00, 99.06)	100.00 (100.00, 100.00)

CI: Confidence interval.

Generalization performance testing

The generalization of the deep learning model to unfamiliar samples was tested by expanding the test subset three times, with 156 positive and 156 negative samples randomly selected each time. The results are shown in Table 2. The model exhibited strong generalizability and excellent stability.

Table 2 Model generalization performance tests.

Test content	Recall % (95%CI)	Precision % (95%CI)	Accuracy % (95%CI)
Ileocecal recognition generalization
The first time	96.15 (96.05, 96.25)	99.34 (99.13, 99.55)	97.76 (97.27, 98.25)
The second time	96.79 (96.34, 97.24)	98.69 (98.33, 99.05)	97.75 (97.52, 97.98)
The third time	96.15 (95.89, 96.41)	99.34 (98.93, 99.75)	97.76 (97.11, 98.41)
Mucosa touch recognition generalization
The first time	98.72 (98.39, 99.05)	100.00 (100.00, 100.00)	99.36 (99.03, 99.69)
The second time	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
The third time	98.72 (98.43, 99.01)	100.00 (100.00, 100.00)	99.36 (99.00, 99.72)

CI: Confidence interval.

Factors that impact MTR detection

In practical colonoscopy, factors such as brightness, contrast, and rotation angle can distort captured images and cause model recognition errors. To test these impacts, algorithm performance tests were conducted for the three factors. The results are shown in Table 3. For brightness, tests used a variation factor of 0.2-2.0 (step 0.2) on the base performance set. The images were severely distorted (not clinically applicable) at factors ≥ 1.4, but the model maintained > 90% recall, precision, and accuracy at 0-1.2. For contrast, a variation factor of 0.2-2.0 (step 0.2) was applied. The images were manually indiscriminable at factors ≥ 1.0, whereas the model remained unaffected at 0.2-0.8. For rotation, angles of -30° to -330° (increment 30°) were tested, with minimal impact on the classification network.

Table 3 Factors that impact mucosa touch rate detection.

Change factor	Ileocecal recognition			Mucosa touch recognition
	Recall % (95%CI)	Precision % (95%CI)	Accuracy % (95%CI)	Recall % (95%CI)	Precision % (95%CI)	Accuracy % (95%CI)
Calculated test metrics for the test set of factors affecting the brightness
0.2	96.15 (95.87, 96.43)	100.00 (100.00, 100.00)	98.08 (97.90, 98.26)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
0.4	96.15 (96.00, 96.03)	98.04 (97.89, 98.19)	97.12 (96.85, 97.39)	100.00 (100.00, 100.00)	98.11 (98.01, 98.21)	99.04 (98.76, 99.32)
0.6	94.23 (94.14, 94.32)	98.00 (97.95, 98.05)	96.15 (95.88, 96.42)	98.07 (97.88, 98.26)	98.07 (97.47, 98.67)	98.07 (96.89, 99.25)
0.8	94.23 (93.17, 95.29)	96.08 (94.33, 97.83)	95.19 (92.83, 97.55)	96.15 (94.52, 97.78)	98.04 (96.14, 99.94)	97.12 (94.86, 99.38)
1.0	92.31 (90.21, 94.41)	96.00 (93.68, 98.32)	94.23 (92.35, 96.11)	94.23 (91.36, 97.10)	96.08 (92.89, 99.27)	95.19 (93.68, 96.70)
1.2	92.31 (89.87, 94.75)	96.00 (92.59, 99.41)	94.23 (91.44, 97.02)	94.23 (91.33, 97.13)	94.23 (92.35, 96.11)	94.23 (90.66, 97.80)
1.4	86.54 (83.77, 89.31)	88.23 (82.99, 93.47)	87.5 (85.44, 89.56)	92.31 (88.75, 95.87)	94.12 (90.33, 97.91)	93.27 (89.71, 96.83)
1.6	19.23 (14.69, 23.77)	58.82 (50.69, 66.95)	52.88 (48.99, 56.77)	42.31 (38.55, 46.07)	50.00 (45.06, 54.94)	50.00 (45.06, 54.94)
1.8	19.23 (14.69, 23.77)	58.82 (50.69, 66.95)	52.88 (48.99, 56.77)	42.31 (38.55, 46.07)	50.00 (45.06, 54.94)	50.00 (45.06, 54.94)
2.0	19.23 (14.69, 23.77)	58.82 (50.69, 66.95)	52.88 (48.99, 56.77)	42.31 (38.55, 46.07)	50.00 (45.06, 54.94)	50.00 (45.06, 54.94)
Calculated test metrics for the test set of contrast influencing factors
0.2	96.15 (93.77, 98.53)	98.04 (97.33, 98.75)	97.12 (96.55, 97.69)	98.08 (96.47, 99.69)	100.00 (100.00, 100.00)	99.04 (98.15, 99.93)
0.4	98.08 (96.81, 99.35)	98.08 (96.99, 99.17)	98.08 (97.00, 99.16)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
0.6	94.23 (92.65, 95.81)	96.08 (94.33, 97.83)	95.19 (92.69, 97.69)	96.15 (94.87, 97.43)	94.34 (91.58, 97.10)	95.19 (93.67, 96.71)
0.8	92.31 (90.77, 93.85)	94.12 (93.17, 95.07)	93.27 (91.22, 95.32)	94.23 (92.59, 95.87)	96.08 (94.33, 97.83)	95.19 (93.74, 96.64)
1.0	86.53 (81.79, 91.27)	86.53 (83.44, 89.62)	86.54 (84.65, 88.43)	92.31 (90.11, 94.51)	73.85 (70.33, 77.37)	79.81 (75.99, 83.63)
1.2	86.53 (83.77, 89.29)	78.95 (76.88, 81.02)	81.73 (78.54, 84.92)	92.31 (90.41, 94.21)	73.85 (70.41, 77.29)	79.81 (77.25, 82.37)
1.4	78.84 (76.38, 81.30)	70.69 (67.89, 73.49)	73.08 (71.21, 74.95)	92.31 (89.76, 94.86)	68.57 (65.71, 71.43)	75.00 (72.99, 77.01)
1.6	75.00 (73.68, 76.32)	67.24 (63.79, 70.69)	69.23 (65.33, 73.13)	88.46 (85.67, 91.25)	63.89 (60.55, 67.23)	69.23 (66.31, 72.15)
1.8	59.62 (57.22, 62.02)	58.49 (57.71, 59.27)	58.65 (54.99, 62.31)	84.62 (81.67, 87.57)	61.97 (59.43, 64.51)	66.34 (64.58, 68.10)
2.0	48.08 (46.71, 49.45)	42.37 (39.97, 44.77)	41.35 (38.53, 44.17)	76.92 (74.66, 79.18)	55.56 (53.78, 57.34)	57.69 (56.43, 58.95)
Calculated test metrics for the test set of rotational change influencing factors
-30°	98.07 (97.99, 98.15)	98.07 (97.99, 98.15)	98.07 (97.99, 98.15)	98.07 (97.99, 98.15)	100.00 (100.00, 100.00)	99.04 (98.08, 100.00)
-60°	96.15 (95.21, 97.09)	98.04 (97.32, 98.76)	97.12 (95.69, 98.55)	96.15 (94.87, 97.43)	100.00 (100.00, 100.00)	98.08 (96.87, 99.29)
-90°	98.07 (96.87, 99.27)	100.00 (100.00, 100.00)	99.03 (98.06,100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
-120°	98.07 (97.66, 98.48)	98.07 (97.99, 98.15)	98.07 (98.00, 98.14)	100.00 (100.00, 100.00)	98.11 (98.03, 98.19)	99.04 (98.08, 100.00)
-150°	94.23 (92.66, 95.80)	96.08 (94.77, 97.39)	95.19 (94.15, 96.23)	96.15 (95.21, 97.09)	100.00 (100.00, 100.00)	98.08 (97.99, 98.17)
-180°	98.07 (97.88, 98.26)	100.00 (100.00, 100.00)	99.03 (98.06, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)	100.00 (100.00, 100.00)
-210°	96.15 (94.77, 97.53)	98.04 (96.58, 99.50)	97.12 (94.99, 99.25)	96.15 (95.37, 96.93)	100.00 (100.00, 100.00)	98.08 (98.01, 98.15)
-240°	94.23 (92.63, 95.83)	96.08 (95.77, 96.39)	95.19 (94.15, 96.23)	98.07 (97.16, 98.98)	100.00 (100.00, 100.00)	99.04 (98.08, 100.00)
-270°	98.07 (97.19, 98.95)	100.00 (100.00, 100.00)	99.03 (98.06, 100.00)	100.00 (100.00, 100.00)	98.11 (98.03, 98.19)	99.04 (98.99, 99.09)
-300°	94.23 (92.41, 96.05)	96.08 (95.88, 96.28)	95.19 (93.77, 96.61)	98.07 (98.00, 98.14)	100.00 (100.00, 100.00)	99.04 (98.89, 99.19)
-330°	96.15 (94.58, 97.72)	98.04 (96.59, 99.49)	97.12 (95.88, 98.36)	96.15 (94.79, 97.51)	100.00 (100.00, 100.00)	98.08 (97.99, 98.17)

CI: Confidence interval.

Correlation analysis between the MTR and the PDR

The study revealed that there was a strong negative correlation between the MTR and PDR (r = -0.948, 95% confidence interval: -0.985 to -0.911; P < 0.01) (Figure 3). The MTR of junior endoscopists was greater (9.96% ± 0.65% vs 7.84% ± 0.78%, P < 0.05). PDR was chosen over ADR because it covers all polyps (vs ADR’s adenoma-only focus) to comprehensively assess endoscopists’ skills via MTR.

Figure 3 Scatter plot of mucosa touch rate vs polyp detection rate. MTR: Mucosa touch rate; PDR: Polyp detection rate.

This strong negative correlation suggests that MTR is a sensitive proxy for inspection quality, although the near-perfect correlation warrants caution regarding potential overfitting, which we mitigated through rigorous train-test separation and prospective validation.

Phase two: Prospective clinical validation

Between December 1, 2024, and June 30, 2025, a total of 3463 consecutive patients who underwent colonoscopy were screened. A total of 597 patients were excluded on the basis of the inclusion and exclusion criteria. Finally, 2866 eligible patients were included in the analysis. Baseline demographic data and procedure-related data are shown in Table 4. No significant differences were detected between the two groups (P > 0.05). In the control group, there were 700 males (46.67%) and 800 females (53.33%), with a mean age of 49.68 years. In the experimental group, there were 649 males (47.51%) and 717 females (52.49%), with a mean age of 50.66 years.

Table 4 Comparison of baseline demographic data characteristics between the control group and experimental group, mean ± SD/n (%).

	Control group (n = 1500)	Experimental group (n = 1366)	P value
Sex			> 0.05
Male	700 (46.67)	649 (47.51)
Female	800 (53.33)	717 (52.49)
Age, years	49.68 ± 13.8	50.66 ± 13.47	> 0.05
Examination purpose			> 0.05
Symptomatic	988 (65.87)	913 (66.84)
Health check-up	375 (23.80)	328 (24.01)
Disease monitoring	107 (53.33)	98 (53.33)
Abnormal test results	30 (7.13)	27 (1.98)

Primary outcomes: PDR and ADR

Overall comparison: Compared with the control group, the experimental group had a significantly greater PDR (51.17% vs 40.07%, χ² = 34.62, P < 0.001). The ADR was slightly greater in the experimental group, but the difference was not statistically significant (10.76% vs 9.40%, χ² = 1.89, P > 0.05) (Table 5).

Table 5 Overall polyp detection rate and adenoma detection rate in the clinical validation phase.

Group	Sample size	PDR		ADR		P value (PDR)	P value (ADR)
		n (%)	(95%CI)	n (%)	(95%CI)
Experimental	1366	699 (51.17)	(48.32, 54.02)	147 (10.76)	(7.35, 14.17)	< 0.001	> 0.05
Control	1500	601 (40.07)	(34.87, 45.27)	141 (9.40)	(6.89, 11.91)

PDR: Polyp detection rate; ADR: Adenoma detection rate; CI: Confidence interval.

Seniority-based comparison

Overall detection rates: High-seniority: The PDR increased from 43.87% to 49.60% (χ² = 5.23; P < 0.05). Low-seniority: The PDR (36.27% to 56.17%, χ² = 59.00, P < 0.001) increased significantly (Table 6).

Table 6 Polyp detection rate and adenoma detection rate by seniority and group.

Endoscopist group	Subgroup	Sample size	PDR		ADR		P value (PDR)	P value (ADR)
			n (%)	(95%CI)	n (%)	(95%CI)
High-seniority	Control	750	329 (43.87)	(39.75, 47.99)	84 (11.20)	(7.46, 14.96)	< 0.05	> 0.05
	Experimental	750	372 (49.60)	(46.25, 52.95)	92 (12.27)	(8.73, 15.81)
Low-seniority	Control	750	272 (36.27)	(33.82, 38.72)	57 (7.60)	(4.38, 10.82)	< 0.001	> 0.05
	Experimental	616	346 (56.17)	(50.36, 61.98)	62 (10.06)	(6.37, 13.75)

PDR: Polyp detection rate; ADR: Adenoma detection rate; CI: Confidence interval.

Subgroup analysis

Age subgroup: In the experimental group, the PDR and ADR increased significantly across all age subgroups, indicating that AI-assisted MTR monitoring was universal in colonoscopies across age groups (P < 0.05). People over 40 years old were the key group for polyp/adenoma detection (Table 7).

Table 7 Polyp detection rate and adenoma detection rate by age subgroup.

Age subgroup	Subgroup	PDR		ADR		P value (PDR)	P value (ADR)
		n (%)	(95%CI)	n (%)	(95%CI)
< 40 years	Control (n = 320)	68 (21.25)	(17.65, 24.85)	25 (7.81)	(5.20, 10.42)	< 0.01	< 0.05
	Experimental (n = 290)	89 (30.69)	(26.71, 34.67)	39 (13.45)	(10.24, 16.66)
40-60 years	Control (n = 790)	293 (37.09)	(33.85, 40.33)	118 (14.94)	(11.37, 18.51)	< 0.001	< 0.001
	Experimental (n = 720)	350 (48.61)	(45.70, 51.52)	168 (23.33)	(19.64, 27.02)
> 60 years	Control (n = 390)	162 (41.54)	(39.45, 43.63)	67 (17.18)	(16.40, 17.96)	< 0.001	< 0.001
	Experimental (n = 356)	183 (51.40)	(46.65, 56.15)	88 (24.72)	(20.98, 28.46)

PDR: Polyp detection rate; ADR: Adenoma detection rate; CI: Confidence interval.

DISCUSSION

Colonoscopy represents the primary method for CRC screening[4], yet its effectiveness is highly dependent on the operator’s technical proficiency. Traditional quality metrics, such as ADR and PDR[23-26], are retrospective and do not facilitate real-time improvement during the procedure[27]. This study introduced MTR as a novel, AI-derived, real-time quality measure that objectively quantifies endoscopic technique by calculating the proportion of frames with significant visual field loss due to mucosal contact. Our findings demonstrate that MTR is strongly correlated with PDR and that real-time MTR feedback significantly improves detection rates, particularly among less experienced endoscopists.

The primary innovation of this work lies in the translation of an often-overlooked technical event, “mucosa touch”, into an objective, quantifiable, and actionable metric. Unlike withdrawal time, which is susceptible to recording bias and does not directly reflect inspection quality[28], MTR is objectively computed via a robust deep learning model that analyzes real-time video sequences. Our AI system achieved high accuracy in identifying mucosa touch events (100% in the test set) and exhibited strong robustness under varying brightness, contrast, and rotation conditions, supporting its potential integration into daily clinical practice.

A key finding was a strong negative correlation between the MTR and the PDR (r = -0.948, P < 0.01), indicating that higher MTR values, reflecting poorer control of the endoscope, are associated with lower PDRs. This correlation underscores MTR’s validity as a proxy for technical skill. More importantly, prospective validation revealed that real-time MTR feedback significantly improved the PDR. The PDR of the AI-assisted group was 51.17%, whereas that of the conventional group was 40.07% (P < 0.001).

Notably, the benefits of MTR monitoring were most pronounced among low-seniority endoscopists, whose PDR increased from 36.27% to 56.17% (P < 0.001), a 19.90% absolute improvement. In contrast, high-seniority operators showed a more modest gain (5.73% increase). This suggests that the MTR can narrow the technical gap between endoscopists a critical need in China, where there is wide variation in colonoscopy quality across institutions[24]. The model provides real-time feedback (e.g., voice alerts for mucosa touch), helping low-seniority endoscopists adjust their technique (e.g., increasing endoscope-mucosa distance) and learn faster. Informal feedback from endoscopists indicated that while the visual alerts were generally helpful, the auditory alerts were occasionally perceived as distracting during therapeutic maneuvers or in complex anatomies. This suggests that customizable alert settings (e.g., adjustable sensitivity, optional mute) would improve usability and reduce alert fatigue in routine practice.

The selection of the PDR over the ADR as the primary end point was deliberate and methodologically justified. PDR reflects the endoscopist’s visual inspection proficiency more directly and is not contingent upon pathological confirmation, which introduces variability due to cost, compliance, and procedural factors[29,30]. This methodological choice is further supported by evidence demonstrating a strong correlation between the PDR and the ADR, establishing PDR as a robust surrogate marker for prospective quality assessment[31-33]. The main purpose of this study was to verify whether MTR can optimize the inspection process in real time. The significant improvement in the PDR perfectly proved this point. Although the ADR increased in the AI-assisted group (10.76% vs 9.40%), the difference was not statistically significant, likely because of the relatively low baseline ADR and sample size limitations specific to adenoma prevalence. However, the 1.36% absolute increase in ADR, while not statistically significant, is clinically meaningful given that each 1% increase in ADR is associated with a 3% reduction in interval CRC risk. Additionally, the significant improvement in PDR, particularly in high-yield subgroups such as middle-aged (40-60 years) and older (> 60 years) patients, supports the clinical utility of MTR monitoring[34,35]. MTR monitoring should be prioritized for this group to maximize CRC prevention.

The most compelling application of MTR lies in its ability to standardize quality and accelerate training. The dramatic 19.9% improvement in the PDR among low-seniority endoscopists suggests that MTR feedback acts as a real-time coach, enabling rapid skill acquisition. This addresses a critical challenge in gastrointestinal endoscopy, particularly in regions with uneven distributions of expertise. We envision MTR being integrated into daily practice in two key ways. First, as a real-time decision-support tool during colonoscopy, it provides instant auditory/visual cues to optimize the withdrawal technique; and second, as an objective benchmark for continuous quality improvement programs, it allows departments to track operator performance over time and identify areas for targeted training.

The significant improvement in PDR without a statistically significant increase in ADR warrants further discussion. This may reflect that MTR feedback primarily enhances visual inspection thoroughness, leading to increased detection of all polyp types, including non-neoplastic lesions. ADR, being dependent on pathological confirmation, can be influenced by factors such as polyp resection rates, specimen retrieval, and histopathological variability. Additionally, the relatively low baseline ADR in our cohort may have limited statistical power to detect a modest yet clinically meaningful difference. Nevertheless, the strong correlation between PDR and ADR established in literature suggests that improvements in PDR are likely to translate into meaningful adenoma detection gains over time.

The integration of MTR into routine quality assurance programs presents both opportunities and challenges. Practically, MTR monitoring could be embedded into existing endoscopy reporting systems, providing immediate feedback to endoscopists and aggregated data for departmental quality dashboards. Potential barriers include workflow adaptation, initial acceptance by experienced endoscopists, and the need for seamless integration with various endoscopy platforms. To facilitate adoption, we propose a phased implementation: Starting with training and auditing settings, followed by optional real-time assistance, and eventually routine use for all screening colonoscopies. User training and highlighting the educational value for less experienced operators may improve acceptance. Preliminary feedback from our center did not indicate significant “alert fatigue”, but formal usability studies in high-volume settings are warranted.

Compared with existing quality indicators, the MTR offers distinct advantages, as it operates in real time, is fully automated and objective, and focuses on a modifiable aspect of the endoscopic technique. While metrics such as the cecal intubation rate reflect procedural completeness and ADR/PDR serve as outcome audits[36,37], the MTR provides immediate, actionable feedback that empowers endoscopists to optimize inspection techniques during the procedure itself.

This study has several limitations that warrant consideration. First, its single-center design may restrict the generalizability of the findings; future multicenter studies utilizing diverse endoscopic platforms are needed to validate the results across broader clinical settings. Second, the absence of per-patient MTR data precluded the establishment of specific intervention thresholds, a gap that should be addressed in future investigations. Furthermore, incomplete subgroup data limited our ability to analyze the potential influence of comorbidities on MTR values. Methodologically, although propensity score matching was applied to mitigate potential bias, the use of “natural randomization” based on endoscopy room availability, while pragmatic, may still carry a risk of residual selection bias. Finally, the relatively short follow-up period precluded the assessment of long-term outcomes, such as the impact on interval CRC rates.

To facilitate the clinical translation of MTR, several key directions for future research are recommended: (1) Multicenter validation across varied endoscopic systems and patient populations; (2) Establishment of standardized, clinically actionable MTR thresholds for real-time feedback; (3) Evaluation of long-term outcomes, including interval cancer incidence and survival endpoints; and (4) Integration of MTR with other AI-based quality metrics, such as withdrawal time and fold examination completeness, to develop a composite real-time quality assessment tool.

This study establishes MTR as a novel, real-time colonoscopy quality indicator. AI-driven feedback during withdrawal boosted PDR, minimized operator variability, and elevated overall quality. This shift from retrospective audit to real-time guidance holds particular promise for standardizing practice across healthcare settings and for training novice endoscopists.

CONCLUSION

The AI-based MTR model reliably quantifies withdrawal technique. Prospective validation confirmed that MTR monitoring enhances colonoscopy PDR, narrows the technical gap between endoscopists of varying seniority, and identifies high-benefit patient subgroups. This feasible, real-time indicator addresses the limitations of traditional metrics, improving routine quality and supporting scalable CRC prevention. Future studies should pursue multicenter validation, MTR threshold determination, and long-term outcome assessment to unlock its full potential for global standardization.

ACKNOWLEDGEMENTS

We thank the nurses and technicians at the Endoscopy Center of Southwest Hospital for their assistance with data collection.