Introduction

Recently, artificial intelligence-based computer-aided detection (CADe) for colonoscopy is garnering significant attention due to its effectiveness in increasing adenoma detection rate (ADR). However, theoretically, CADe system cannot decrease exposure error which polyps obscured by an object. In this regard, computer-aided quality improvement (CAQ) which expects to induce the meticulous mucosal inspection and leading to a reduction of blind spot area. In this study, we developed and evaluated CAQ for colonoscopy.

Aims & Methods

This was a single center, retrospective study from period April 1, 2022 to February 25, 2023. We randomly extracted colonoscopy videos and created 18,922 training images. Quadrant images were made from one image and three endoscopists annotated the following scores; i) appropriateness for mucosal inspection, ii) appropriate distance, and iii) presence of blind area in the field of view (e.g., fold or flexure). We conducted supervised machine learning using annotated images. The CAQ deployed multi-head deep convolutional neural network model which can output multiclass prediction from one image. Consequently, the CAQ system outputs aforementioned score from whole colonoscopy video. To validate its performance, we retrospectively analyzed colonoscopy procedures performed by seven endoscopists using the CAQ. We compared calculated CAQ model scores and expertise of endoscopists.

Results

29 videos were included as validation samples (8 videos were performed by non-expert endoscopists) As a result, there was a significant difference in the proportion of appropriate images during the examination between experts and non-experts (median 56% vs. 49%, p=0.005, Mann–Whitney U test) and there was also a significant difference in the evaluation score of the distance between the colon mucosa and the endoscope (median 4.2 vs. 3.2, p=0.0004, Mann–Whitney U test). In addition, there was a significant difference in the scores evaluating the presence of blind areas in the field of view (median 1.6 vs. 1.2, p=0.0007, Mann-Whitney U test).

Conclusion

In this study, developed CAQ system could quantify the skill of the endoscopists. Such CAQ would be expected to standardize the observation technique of colonoscopy.

Introduction

Computer-aided detection (CADe) systems improve adenoma detection rate (ADR). However, there is some concern that the long-term use of CADe systems might lead to a decline in endoscopists’ independent detection skills, particularly their ability to detect lesions without assistance—a phenomenon referred to as "deskilling".

Aims & Methods

This study aims to evaluate the impact of CADe on the learning curve of endoscopists for ADR. This was a propensity score–matched (PSM), prospective, single-center, observational study conducted at a tertiary care hospital from January 2021 to December 2023. CADe systems were installed in half of the endoscopy units, and patients were equally distributed between the rooms. Patients in CADe-equipped rooms were examined using EndoBRAIN-EYE (Olympus Corp, Tokyo, Japan). Patients aged ≥ 20 years scheduled for colonoscopy were included, excluding those with polyposis, inflammatory bowel disease, known polyps, incomplete colonoscopy, emergency cases, previous colorectal surgery, and examination by novices. Endoscopists were classified as high-detectors (ADR ≥ 25%) or low-detectors (ADR < 25%) based on the ADR recorded before CADe implementation. The primary outcome was to evaluate the learning effect of CADe using cumulative summation (CUSUM) analysis based on ADR during both CADe and non-CADe procedures to assess skill acquisition and transfer. The CUSUM scores were calculated from the acceptable probability of failure and unacceptable probability of failure. Acceptable and unacceptable ADR were defined as 40% and 30% for high detectors, and 35% and 25% for low detectors, respectively. Performance was judged based on the boundary lines (h0 and h1): acceptable if the CUSUM plot fell below h0, unacceptable if it went above h1, and inconclusive requiring further observation if it stayed between these boundaries. The secondary outcomes included ADR, advanced adenoma detection rate (AADR), and sessile serrated lesion detection rate (SSLDR). The study received ethics approval (no.19H072) and complied with the Declaration of Helsinki.

Results

Out of 5,717 patients initially enrolled, 4,712 patients (2,356 in CADe group and 2,356 in non-CADe group) were analyzed after PSM for known factors (age, sex, indication, and bowel preparation). During CADe-assisted procedures, both high-detectors and low-detectors showed patterns below the acceptable performance line (h0), indicating improved adenoma detection performance with AI support. However, during non-CADe procedures, the patterns diverged significantly. For high-detectors, their CUSUM plots consistently fell below the acceptable performance line (h0) over time during non-CADe colonoscopies, indicating that they successfully acquired and maintained improved adenoma detection skills even without AI assistance. In contrast, the CUSUM plots of the low-detectors remained between the acceptable (h0) and unacceptable (h1) limits throughout the study period, showing no significant change in learning trajectory during non-CADe procedures. The CADe group had a significantly higher ADR (40.5% [95% CI: 38.5-42.5%]) compared to non-CADe group (32.4% [95% CI: 30.5-34.4%]; p < 0.001). SSLDR was also significantly higher in the CADe group (6.9% [95% CI: 5.9-8.0%] vs. 4.5% [95% CI: 3.7-5.4%]; p = 0.001), while AADR showed no significant difference (4.1% [95% CI: 3.4-5.0%] vs. 4.7% [95% CI: 3.9-5.6%]; p = 0.39).

Conclusion

The implementation of the CADe system did not result in deskilling regarding adenoma detection.

Disclosure

The conflict of interests were following: Shin-ei Kudo and Masashi Misawa received speaking honoraria from Olympus Corporation and have an ownership interest of the products of Cybernet Systems. Kensaku Mori received research grants from Cybernet Systems, Morita Mfg, Atom Medical Inc, and Asahi Intecc. Additionally, Kensaku Mori serves as a Technical Advisor for Sanyu Industry. Tetsuo Nemoto received a research grant from Olympus Corporation for other studies. Other authors have nothing to declare.

Introduction

AI (artificial intelligence)-driven computer-aided detection (CADe) systems are typically trained using high-quality, standardized colonoscopy images/videos, assuming its use under optimal bowel preparation. However, it remains uncertain how varying levels of bowel cleanliness affect CADe performance.

Aims & Methods

To measure the impact of bowel preparation quality on the adenoma detection capability of CADe systems.
This is a cumulative post-hoc analysis of individual data from three randomized controlled studies with similar design comparing CADe-assisted with standard colonoscopy (SC). Only patients with adequate bowel preparation (Boston Bowel Preparation Scale [BBPS] >2 in each colonic segment) were included. The primary outcome measure was the difference in adenoma detection rate (ADR) and adenoma per colonoscopy (APC) between the CADe-assisted and SC groups, according to the quality level of bowel preparation, classified as good (BBPS 6–7) or excellent (BBPS 8–9).

Results

A total of 2103 patients (1047 CADe, 1056 SC; mean age 61.5 ± 9.2 years; 50.4% male) were included. Bowel preparation was rated as good in 906 patients (43.1%) and excellent in 1197 patients (56.9%). CADe increased ADR and APC compared to SC in both good (52.2% vs. 43.8%; 1.21 vs. 0.93) and excellent (54.5% vs. 43.9%; 1.09 vs. 0.89) bowel preparation groups. The absolute increase in ADR with CADe was +8.4% (95% CI: 1.9–14.9) and +10.6% (95% CI: 5.0–16.2) in good and excellent preparation groups, respectively, with no significant interaction between preparation quality and CADe effect (Δ = +2.2%, 95% CI: -6.4–10.8; interaction p = 0.612). Similarly, APC increases were +0.28 (95% CI: 0.06–0.50) and +0.20 (95% CI: 0.03–0.36) for the good and excellent preparation groups, respectively, without significant interaction observed (Δ = -0.08, 95% CI: –0.22 to +0.06, interaction p = 0.261). Similar trends were observed when the analysis was restricted to the right colon or to diminutive adenomas.

Conclusion

In patients with adequate bowel preparation, CADe improves adenoma detection regardless of cleansing quality grades.

References

1. Repici, A, Badalament M, Maselli R et al. Efficacy of Real-Time Computer-Aided Detection of Colorectal Neoplasia in a Randomized Trial. Gastroenterology 2020; 159: 512–520
2. Repici A, Spadaccini M, Antonelli G, et al. Artificial intelligence and colonoscopy experience: lessons from two randomised trials. Gut 2022; 71: 757-765.
3. 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.

Introduction

Recently, artificial intelligence-based computer-aided detection (CADe) for colonoscopy has garnered attention due to its effectiveness in increasing the adenoma detection rate (ADR). However, theoretically, CADe systems cannot reduce exposure errors for polyps obscured by objects. In this regard, computer-aided quality improvement (CAQ) which expects to induce the meticulous mucosal inspection and leading to a reduction of blind spot area. In this study, we developed and evaluated CAQ for colonoscopy.

Aims & Methods

This was a single-center, retrospective study conducted from April 1, 2022 to October 31, 2023. We randomly extracted videos of the entire colonoscopies and created 18,922 training images. Quadrant images were made from one image and three endoscopists annotated the following scores; i) (overall) quality of image(not evaluable 0 or evaluable 1), ii) appropriate distance between the colon mucosa and the endoscope(Scored on a scale of 0 to 8), iii) the amount of residue in the colon,(Scored on a scale of 0 to 8) and ⅳ) presence of blind area in the field of view (e.g., fold or flexure) (Scored on a scale of 0 to 4). We conducted supervised machine learning using annotated images. The CAQ deployed multi-head deep convolutional neural network model which can output multiclass prediction from one image. Consequently, the CAQ system outputs aforementioned score from colonoscopy video. To validate its performance, we randomly selected videos from the cecum to the hepatic flexure performed by 19 endoscopists and calculated CAQ model scores. We compared CAQ scores between endoscopists with low and high ADR (ADR cutoff is 35%) and with and without the history of detecting early depressed-type colorectal cancer by endoscopists.

Results

62 videos were included as validation samples. Among the 19 endoscopists, 11 had a low ADR and 8 had a high ADR. The results showed a significant difference in the rate of evaluable images between endoscopists with low and high ADR (median 70% vs. 77%, p = 0.003, t-test). There was also a significant difference in the evaluation score of the distance (median 3.8 vs. 4.2, p = 0.003, t-test). A significant difference was also obtained in the score for the amount of residue in the colon(median 5.3 vs. 5.7, p = 0.01, t-test). However, there was no significant difference in the scores evaluating the presence of blind areas(median 1.01 vs. 1.10, p = 0.13, t-test)
Among the 19 endoscopists, 5 endoscopists had a history of detecting early depressed-type colorectal cancer, while 14 did not. The results showed a significant difference was observed in the rate of evaluable images between endoscopists with and without a history of detecting early depressed-type colorectal cancer(median 80% vs. 71%, p < 0.001, t-test) . Additionally, there was a difference in the evaluation scores for the distance (median 4.5 vs. 3.8, p < 0.001, t-test). There was also a difference in the scores assessing the presence or absence of blind spots (median 1.2 vs. 1.0, p = 0.006, t-test). Moreover, a significant difference was found in the evaluation scores for the amount of residue in the colon(median 5.9 vs. 5.3, p < 0.001, t-test) .

Conclusion

The CAQ system developed in this study was able to quantify the skill of endoscopists. Such CAQ is expected to standardize the observation technique of colonoscopy and may lead to elucidation of detection techniques for significant lesions such as early depressed-type colorectal cancer.