Background & Aims: One fourth of colorectal neoplasias are missed during screening colonoscopies—these can develop into colorectal cancer (CRC). Deep learning systems allow for real-time computer-aided detection (CADe) of polyps with high-accuracy. We performed a multicenter, randomized trial to assess the safety and efficacy of a CADe system in detection of colorectal neoplasias during real-time colonoscopy.
Methods: We analyzed data from 685 subjects (61.32±10.2 years old; 337 women) undergoing screening colonoscopies for CRC, post-polypectomy surveillance, or work up due to positive results from a fecal immunochemical test or signs or symptoms of CRC, at three centers in Italy from September through November 2019. Patients were randomly assigned (1:1) to groups who underwent high-definition colonoscopies with the CADe system or without (controls). The CADe system included an artificial intelligence-based medical device (GI Genius, Medtronic) trained to process colonoscopy images and superimpose them, in real time, on the endoscopy display a green box over suspected lesions. A minimum withdrawal time of 6 min was required. Lesions were collected and histopathology findings were used as the reference standard. The primary outcome was adenoma detection rate (ADR, the percentage of patients with at least 1 histologically proven adenoma or carcinoma). Secondary outcomes were adenomas detected per colonoscopy, nonneoplastic resection rate, and withdrawal time.
Results: The ADR was significantly higher in the CADe group (54.8%) than in the control group (40.4%) (relative risk [RR], 1.30; 95% CI, 1.14–1.45). Adenomas detected per colonoscopy were significantly higher in the CADe group (mean, 1.07±1.54) than in the control group (mean 0.71±1.20) (incidence rate ratio, 1.46; 95% CI, 1.15–1.86). Adenomas 5 mm or smaller were detected in a significantly higher proportion of subjects in the CADe group (33.7%) than in the control group (26.5%; RR, 1.26; 95% CI, 1.01–1.52), as were adenomas of 6–9 mm (detected in 10.6% of subjects in the CADe group vs 5.8% in the control group; RR, 1.78; 95% CI, 1.09–2.86), regardless of morphology or location. There was no significant difference between groups in withdrawal time (417±101 sec for the CADe group vs 435±149 for controls; P=.1) or proportion of subjects with resection of non-neoplastic lesions (26.0% in the CADe group vs 28.7% of controls; RR, 1.00; 95% CI, 0.90–1.12).
Conclusions: In a multicenter, randomized trial, we found that including CADe in real-time colonoscopy significantly increases ADR and adenomas detected per colonoscopy without increasing withdrawal time. ClinicalTrials.gov no: 04079478
KEY WORDS: Artificial Intelligence, Adenoma per colonoscopy, comparison, early detection
One fourth of colorectal neoplasia is missed at screening colonoscopy , representing the main
cause of interval Colorectal Cancer (CRC) [2,3], and resulting in an unacceptable variability in the key
quality indicator, namely Adenoma detection rate (ADR), among endoscopists [4,5].
Failure in polyp recognition is a major determinant for this miss rate of colorectal neoplasia [6,7].
Each colonoscopy is made of approximately 50,000 frames, corresponding to approximately 25-30 frames
per second, and one polyp may be recognizable only in a few frames, explaining how failure in polyp
recognition is likely to occur, irrespectively of the endoscopy setting.
The theoretical and technological advances in Deep Learning led to the development of Computer
Aided Polyp Detection (CADe) systems. These systems showed a high-accuracy in polyp recognition when
retrospectively applied to endoscopy videos or images both in terms of true- and false-positive results [8–
The impact of CADe on neoplasia detection may be considered as a good proxy for the efficacy of
these systems in reducing the miss rate, as gradients in ADR and APC are inversely related with the miss rate
after normalizing for disease-prevalence . However, polyp detection does not only depend on polyp
recognition – the main factor currently addressed by CADe – as it is also affected by the degree of exposure
of the mucosa that in turn is related with the speed of withdrawal time, endoscopist skill, level of cleansing
and other factors . Thus, the actual effect of CADe on detection of colorectal neoplasia is still unclear. The
only available clinical study adopted an experimental setting with two monitors (one with and one without
CADe) due to the unfeasibility of a simultaneous CADe diagnosis . A new CADe system (GI-Genius,
Medtronic) avoids the need of a second display to show the AI detection box on the endoscopy image, and as
such it is fully integrated in the endoscopy workflow, enabling real-time video processing at the same frame
rate as the standard procedure, without requiring any artificial modifications of the usual colonoscopy
In addition to its efficacy, it is relevant to qualify CADe safety indicators, to better understand how it
is going to impact CRC prevention (e.g. possibile unnecessary resections, withdrawal time, etc.).
Aim of the AID (Artificial Intelligence for Colorectal Adenoma Detection) study was to assess the
safety and efficacy of a CADe system for the detection of colorectal neoplasia.
This parallel, randomised, multicentre trial was performed in three Italian endoscopy centres
participating in the organised population CRC screening programme (IRB: ICH2363/2019). The study was
reported according to the CONSORT guidelines for RCTs and was registered on the ClinicalTrial.gov (NCT:
04079478). This was a no-profit study, and no funding was received or solicited, except the loan of the
equipment by Medtronic. All authors had access to the study data and reviewed and approved the final
The target population included 40- to 80-years-old subjects undergoing colonoscopy for primary
CRC screening or post-polypectomy surveillance, as well as for work-up following FIT positivity (cut
off=20 µg Hb/g faeces) or for symptoms/signs. Patients were excluded in case of personal history of CRC, or
IBD, previous colonic resection, antithrombotic therapy precluding polyp resection, and lack of informed
Artificial Intelligence (CADe)
A Convolutional Neural Network (GI-Genius, Medtronic) was trained and validated (99.7% per
lesion sensitivity, 0.9% of false-positive frames ) using a series of videos of 2,684 histologically
confirmed polyps from 840 patients enrolled in a high-quality randomized trial . This CADe receives as
input the digital image from the endoscopy processor and outputs the coordinates of a bounding box only
when an instance of the target polyp is recognized in the image. The output appears on the same endoscopy
screen as the system latency in outputting the video processor images with the detections is not perceivable
by the user (1.52 ± 0.08 µs), thus allowing real-time assessment. (Figure 1).
Before colonoscopy initiation, eligible subjects were randomised in a 1:1 ratio by the endoscopists to
receive colonoscopy with or without CADe in both insertion and withdrawal phases of the procedure.
Randomization was based on a a list of random numbers generated for each centre by the coordinating
centre. Randomization was stratified by gender, age, and personal history of adenomas. The operator was not
blinded to the study arm assigned to the patient.
Colonoscopies were performed by 6 experienced endoscopists (2 for each centre; >2000 screening
colonoscopies) in three centres participating in the organized screening program. All procedures were
performed with a high-definition ELUXEO 700 series (EC-760R, EC-760ZP, FUJIFILM Co., Tokyo) or
EXERA III (Olympus CV-190; Olympus Co., Tokyo). For the purpose of the study, use of magnification,
chromoendoscopy or light-modification technologies was restricted only for polyp characterization at
endoscopist’s discretion. Bowel preparation (Appendix, Table A9) was evaluated and graded by the
endoscopist performing the exam, using the Boston Bowel Preparation (BBPS) scale . Subjects with 0 or
1 in any one of the three segments were excluded from the primary analysis. The endoscopist and facility
staff were allowed to adopt their standard procedures for subject management and monitoring, including use
of conscious sedation. Caecal intubation was assessed by the endoscopist by the identification of the
ileocecal valve and the appendix orifice via photo documentation.
Intubation time and inspection time during withdrawal were measured using a stopwatch, pausing during
therapeutic interventions and washing. Endoscopists were required to comply with a minimum of 6 minutes
of inspection (i.e., clean withdrawal time) [15-16]. All polyps were classified according to their location,
size, and morphology according to Paris classification . Location was considered proximal if proximal to
the splenic flexure. All polyps were removed (biopsy for non-resectable lesions), irrespective of size, color
or subjective interpretation, with the exception of diminutive hyperplastic-appearing polyps located in the
rectum and—according to the judgment of the endoscopists—not clinically significant.
All resected or biopsy specimens were fixed in 10% buffered formalin solution in separate jars. They
were processed and stained for histopathology using standard methods and evaluated by expert pathologists
participating in the organized screening program (one in each centre), who were blinded to the assigned
examination mode. All lesions were classified according to Vienna classification . An advanced
adenoma was defined as an adenoma ≥10 mm and/or with villous component >20%, and/or high-grade
The primary outcome was the adenoma detection rate (ADR) according to intervention arm. ADR
was defined as the proportion of patients with at least one histologically proven adenoma or carcinoma .
Sessile Serrated Lesions (SSLs) were not computed in ADR calculation (ED-1; R2-3/4); Secondary
outcomes were proximal ADR, total number of polyps detected, SSL detection rate, mean number of
adenomas per colonoscopy (APC), cecal intubation rate, withdrawal time. The proximal ADR was defined as
the prevalence of patients with at least one adenoma detected proximal to the splenic flexure (including
cecum, ascending, and transverse colon). APC was defined as the total number of adenomas divided by the
number of colonoscopies performed. APPC was defined as the total number of adenomas divided by the total
of colonoscopies where at least one adenoma was discovered. We also defined non-neoplastic resection rate
as the proportion of patients with no adenoma or SSL within any excised lesions who had undergone at least
one excision with histopathological examination .
The sample size was calculated based on the evaluation of primary outcome that is the per-patient
adenoma detection rate (ADR). A sample size of 322 patients per arm was required, based on the expected
ADR of 35% for both arms, a non-inferiority margin of 10%, power of 90% and an alpha level of 2.5% (one
sided). Non-inferiority was met for the primary endpoint if the lower 2-sided 95% confidence interval (CI)
excluded a 10% or greater difference in favor of the control group. The 10% non-inferiority margin reflects a
typical maximum clinically acceptable difference for comparative studies of this type. If non-inferiority was
demonstrated for the primary endpoint, the endpoint was assessed for superiority (1-sided p value <.025)
using the Fisher exact test.
The primary outcome analysis was the comparison of ADR between the two study arms. Intention
to-treat (and per protocol) analysis was conducted. Differences were expressed as relative risk (RR) with
95% CIs. Categorical variables were described by frequency counts and percentages. Quantitative variables
were described by mean and standard deviations. Chi-square and t-tests were used to compare categorical
and continuous variables between the two groups, respectively. Multivariate estimations of prevalence ratios
were obtained using log-binomial regression; adjustments were made for age, gender, colonoscopy
indication. We also estimated the prevalence of adenomas by colonic location (distal, including the
descending-sigmoid colon, and rectum).vs proximal colon (including, cecum, ascending, and transverse
colon), and by morphology (Paris Classification: polypoid vs. non-polypoid lesions) in the two arms. The
overall APC was calculated in addition to APC stratified by age, gender and colonoscopy indication. Using
Poisson regression, we calculated incident rate ratios (IRR) to assess the relationship between study arm,
age, gender, and colonoscopy indication. Lastly, a per-polyp analysis to assess differences in adenoma
location, size, and morphology was also performed using mixed effects logistic regressions to control for
multiple lesions per patient. Data were reported as odds ratios. Mixed effects models were fit with a random
intercept and treatment as a fixed effect. We did not control for clustering within endoscopy center. A
p < 0.05 was considered statistically significant. All statistical analyses were performed using R software
version 3.5.1 (2018-07-02).
A total of 700 subjects were considered eligible for the study between September and November
2019. After the exclusion of 15 patients (Figure 2), the study cohort was represented by 685 (males: 49.2%,
mean age 61.32+10.2 years) randomized patients. Of these, 341 were allocated in the CADe arm, and 344 in
the control; no difference in clinical indication between the two arms was found (Table 1), and it was overall
primary CRC screening or post-polypectomy surveillance in 46.3% (317/685), work-up of FIT+ in 30.2%
(207/685), and GI-symptoms in 23.5% (161/685). No difference between CADe and control was observed in
term of adequate (BBPS >2 in all colonic segments) cleansing (339/341, 99.4% vs. 342/344, 99.4%; p= 1.0)
and caecal intubation rate (326/341 subjects, 95.6% vs. 339/344, 98.5%; p=0.7).
Per-patient analysis (Figure 3)
In the CADe group, 187/341 patients were diagnosed with at least one adenoma or CRC at
colonoscopy as compared with 139/344 patients in the control group, corresponding to an ADR of 54.8%
and 40.4%, respectively. After adjusting for age, gender and indication, ADR was significantly higher in the
CADe as compared with control group (RR, 1.30; 95% CI: 1.14-1.45; Table 2; Appendix Table A1). The per
protocol analysis produced similar results (Appendix Table A2).
Regarding morphology (Table 2), the rate of patients with nonpolypoid lesions was higher in the
CADe than in control group (90/338 [26.6%] vs. 63/343 [18.4%]; RR, 1.42 [1.09, 1.79]), as well as that with
polypoid lesions (37.3% vs. 26.5%; RR, 1.36 [1.12, 1.62]). Regarding polyp size (Table 2), the proportion of
patients with <10 mm adenomas was higher in the CADe group (151/341 [44.3%]) than in the control group
(111/344, 32.3%; RR, 1.33 [1.13, 1.53]), whilst no difference for those with >10 mm as largest lesion was
observed. As shown in Table 2, the difference between the two arms was significant for both <5 mm (RR,
1.26; 95% CI: 1.01-1.52) and 6-9 mm adenomas (RR, 1.78 [1.09-2.86]). Regarding location (Table 2), the
proportion of patients with proximal adenomas was higher in the CADe group (123/338 [36.4%]) than in the
control group (97/341, 28.4%; RR, 1.28 [1.03, 1.59]), as well as that with distal adenomas (109/338 [37.0%]
vs. 72/341 [23.6%], 37%; RR, 1.53 [1.18, 1.98]). Regarding multiplicity, 131 (19.1%) patients had >2
adenomas: percentages of patients with multiple adenoma in CADe and control group were 23.2% (79) and
15.1% (52), respectively (RR, 1.50; 95% CI:1.19-1.95).
Regarding histology, 45 patients were diagnosed with advanced neoplasia in the CADe group,
compared with 36 control group patients, corresponding to a detection rate of advanced neoplasia of 13.2%
and 10.5%, respectively (adjusted RR: 1.22; 95% CI: 0.80-1.93, Table 2; Appendix Table A1). No
difference in the proportion of patients with at least one SSL was found between the two groups (CADe,
7.0% vs control, 5.2%; p=0.326).
Individual detection rates of at per centre and endoscopist levels are provided in the Appendix (Table
A5-A8), as well as ADR according to indications.
Non-neoplastic resection rate
Overall, 460/685 (67.1%) patients had polyp resections. Of these, 120/460 (27.1%) did not have
histologically proven adenomas, SSLs, or CRCs. The non-neoplastic resection rate were 68/262 (26.0%) and
57/198 (28.8%) in CADe and control group, respectively (RR:1.00; 95% CI:0.90-1.12; P=0.940).
Per polyp analysis
In the 262 and 198 patients with polyp resection in the CADe and control groups, 353 and 243
adenomatous polyps were detected, respectively. Characteristics of detected polyps and cancers according to
intervention arm are summarized in Appendix Tables A3 and A4.
The APC overall was 0.87±1.39 (Table 3), and it was significantly higher in the CADe than in the
control group (1.07±1.54 versus 0.71±1.20), IRR, 1.46; 95% CI:1.15-1.86). The association between the
APC and study arm remained significant after adjusting for gender, indication, and withdrawal time in a
random effect model (OR, 1.80; 95% CI:1.14-2.81) (Appendix Table A4).
Differences in APC between CADe and control group were also analysed according to polyp
characteristics, as detailed in (Table 3). A statistically significant increase in APC between CADe and
control group was found for both polypoid and nonpolypoid lesions, as well as for both proximal and distal
locations. Regarding polyp size the difference was significant only for <10 mm polyps.
The addition of real-time CADe to colonoscopy resulted in a 30% and 46% relative increase in ADR
and APC, demonstrating its efficacy in improving the detection of colorectal neoplasia at screening and
diagnostic colonoscopy. Safety of CADe was demonstrated by the lack of increase of both useless resections
and withdrawal time, as well as by the exclusion of any underskilling in the study period. CADe efficacy
appeared to be independent of morphology and location of neoplasia, and it was mainly explained by the
additional detection of <5 mm and 6-9 mm polyps.
CADe was already shown to be highly accurate for polyp recognition in retrospective assessment of
videos with already diagnosed polyps [8–10]. However, there was uncertainty on its impact on polyp
detection that also depends on the degree of mucosa exposure. The 14% absolute increase in ADR obtained
by CADe in our study indicates that failure in polyp recognition is a clinically relevant cause of miss rate. Of
note, the efficacy of CADe in reversing such miss rate also indicates that the same operator that missed the
lesion in the first place was able to correctly diagnose it when the lesion was presented by the CADe. This
underlines that the main cognitive challenge in polyp recognition is the discrimination between the candidate
lesion and the surrounding healthy mucosa, while its correct characterization as neoplastic tissue – that
occurs after CADe detection – is apparently a much easier task. This is well in line with previous evidence
that the use of dye-spray or electronic chromoendoscopy was able to increase the ADR of the operator by
emphasizing the contrast between the neoplastic and healthy tissue [13,19]. The additional advantage of
CADe is that it completely automatizes – by presenting a well delimited box around the putative lesion
(Figure 1) – the detection phase, while chromoendoscopy still requires the endoscopist to identify the lesion
in the first place. Future studies should address whether the addition of CADe over chromoendoscopy further
improves neoplasia detection over CADe alone.
Our study setting allowed us to assess the safety of CADe colonoscopy for experienced endoscopists.
Differently from previous studies based on experimental setting , we fully integrated CADe in the
endoscopy system, completely mimicking the usual routine of the operators by over-imposing the CADe box
over the same endoscopic screen. Our study excludes any under-skilling effect – i.e. an ADR reduction – on
the operators, marginalizing any possibility for the CADe to distract or reduce the level of alertness of the
operator. Secondly, we excluded any clinically relevant effect of non-neoplastic resections due to CADe.
First, the actual number of patients with useless resection of non-neoplastic polyps did not increase.
Secondly, there was not increase in the withdrawal time. Both of these outcomes indicate that the
endoscopist is fast and accurate in lesion characterization, discarding non-neoplastic lesions detected by
CADe with the same competence as when detected without it. Despite we did not assess the actual number of
false positive activations by the system, as this would have altered the routine setting of our study, we
already showed such number to be less than 1% of the whole colonoscopy video .
The contribution of small lesions to the gradient between the two study arms is in line with pooling
of tandem studies showing such lesions to be associated with miss rate at colonoscopy . When
considering the additional detection as a proxy for a reduction of miss rate, CADe was able to increase the
ADR by targeting the miss rate of the most subtle lesions in the colonoscopy field. The correspondence
between the increase in ADR and APC underlines the efficacy of CADe not only in classifying qualitatively
one patient as an adenoma-bearing or negative one, but also in correctly diagnosing all the burden of
neoplasia that occur in each single patient. Such synergic effect should anticipate a more profound result in
terms of post-colonoscopy CRC prevention. Of note, the proportion of patients with multiplicity was higher
in the CADe as compared with control arm.
The independence of the CADe effect from location and morphology is also in line with previous
tandem studies that excluded a role for proximal location as well as for sessile versus flat morphology in the
miss rate of neoplasia at colonoscopy . Contrarily, the association between increased yield and <10 mm
size was not unexpected as miss rate for these lesions appeared to be substantially superior to those for >10
mm lesions in the pooling of tandem studies . Due to the high prevalence and subtle appearance,
diminutive and small lesions may be considered as a proxy for the technical competence of the endoscopist.
Such competence in turn has been strictly associated with the degree of post-colonoscopy CRC risk.
There are limitations to our analysis. First, we could not exclude a psychological bias as the operator
was aware of the randomized intervention. However, the ADR in the control arm was actually superior (41%
vs. 35%, Appendix Table A5) to what it was previously recorded in the clinical setting of one of the three
centres and used to estimate the sample size. The ADR in the control arm was approximately 2-fold superior
to the previous CADe study with an experimental setting , being more representative of the western-like
endoscopy setting. In addition, the equivalence in withdrawal time exclude a somewhat reduced degree of
mucosal exposure in the control arm. When selecting only the FIT+ patients, the ADR in the control arm was
also in line with previous studies from our group (Appendix Table 5). Finally, the increase in ADR was
consistent in all the three centres involved (Appendix Table A5), marginalizing the possibility of operator
related bias. Third, we did not include low-detectors, inexperienced or non-gastroenterologist endoscopists in
our study. Thus, there is uncertainty on whether CADe would be equally beneficial in these categories for the
following reasons. First, low-detectors have been shown to be suboptimal in the technical exposure of
colorectal mucosa. Thus, the additional contribution of an improved polyp recognition with CADe is
uncertain. Secondly, we cannot exclude that less experienced endoscopists would remove an excess of non
neoplastic lesions triggered by false positive results, differently from what reported in our analysis.
Alternatively, endoscopists with less experience could require more time to assess the false positives by
CADe, affecting the efficiency of the colonoscopy procedure, i.e. prolonging the withdrawal time.
In conclusion, we showed the safety and efficacy of integrating CADe in real-time colonoscopy. The
substantial improvement of ADR and APC without increasing the removal of non-neoplastic lesions is likely
to improve the quality of colonoscopy without affecting its efficiency.
REFERENCE 1 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 Published Online First: 6 February 2019. doi:10.1053/j.gastro.2019.01.260
2 Pohl H, Robertson DJ. Colorectal Cancers Detected After Colonoscopy Frequently Result From Missed Lesions. Clinical Gastroenterology and Hepatology 2010;8:858–64. doi:10.1016/j.cgh.2010.06.028
3 Rex DK, Eid E. Considerations regarding the present and future roles of colonoscopy in colorectal cancer prevention. Clin Gastroenterol Hepatol 2008;6:506–14. doi:10.1016/j.cgh.2008.02.025
4 Rex DK. Colonoscopic withdrawal technique is associated with adenoma miss rates. Gastrointest Endosc 2000;51:33–6. doi:10.1016/s0016-5107(00)70383-x
5 Corley DA, Jensen CD, Marks AR, et al. Adenoma detection rate and risk of colorectal cancer and death. N Engl J Med 2014;370:1298–306. doi:10.1056/NEJMoa1309086
6 Lee CK, Park DI, Lee S-H, et al. Participation by experienced endoscopy nurses increases the detection rate of colon polyps during a screening colonoscopy: a multicenter, prospective, randomized study. Gastrointest Endosc 2011;74:1094–102. doi:10.1016/j.gie.2011.06.033
7 Buchner AM, Shahid MW, Heckman MG, et al. Trainee participation is associated with increased small adenoma detection. Gastrointest Endosc 2011;73:1223–31. doi:10.1016/j.gie.2011.01.060
8 Misawa M, Kudo S-E, Mori Y, et al. Artificial Intelligence-Assisted Polyp Detection for Colonoscopy: Initial Experience. Gastroenterology 2018;154:2027-2029.e3. doi:10.1053/j.gastro.2018.04.003
9 Urban G, Tripathi P, Alkayali T, et al. Deep Learning Localizes and Identifies Polyps in Real Time With 96% Accuracy in Screening Colonoscopy. Gastroenterology 2018;155:1069-1078.e8. doi:10.1053/j.gastro.2018.06.037
10 Hassan C, Wallace MB, Sharma P, et al. New artificial intelligence system: first validation study versus experienced endoscopists for colorectal polyp detection. Gut Published Online First: 15 October 2019. doi:10.1136/gutjnl-2019-319914
11 Yamada M, Saito Y, Imaoka H, et al. Development of a real-time endoscopic image diagnosis support system using deep learning technology in colonoscopy. Sci Rep 2019;9:14465. doi:10.1038/s41598-019-50567-5
12 Wang P, Berzin TM, Glissen Brown JR, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut 2019;68:1813–9. doi:10.1136/gutjnl-2018-317500
13 Repici A, Wallace MB, East JE, et al. Efficacy of Per-oral Methylene Blue Formulation for Screening Colonoscopy. Gastroenterology 2019;:S0016508519303622. doi:10.1053/j.gastro.2019.02.001
14 Lai EJ, Calderwood AH, Doros G, et al. The Boston bowel preparation scale: a valid and reliable instrument for colonoscopy-oriented research. Gastrointest Endosc 2009;69:620–5. doi:10.1016/j.gie.2008.05.057
15 Barclay RL, Vicari JJ, et al. Colonoscopic withdrawal times and adenoma detection during screening collonoscopy. N Egl J Med 2006;355(24):2533-41.
16 Rex DK, Colonoscopic withdrawal technique is associated with adenoma miss rates. Gastrointest Endosc 2000; 51(1):33-6
17 The Paris endoscopic classification of superficial neoplastic lesions: esophagus, stomach, and colon: November 30 to December 1, 2002. Gastrointest Endosc 2003;58:S3-43. doi:10.1016/s0016-5107(03)02159-x
18 Schlemper RJ, Riddell RH, Kato Y, et al. The Vienna classification of gastrointestinal epithelial neoplasia. Gut 2000;47:251–5. doi:10.1136/gut.47.2.251
19 Atkinson NSS, Ket S, Bassett P, et al. Narrow-Band Imaging for Detection of Neoplasia at Colonoscopy: A Meta-analysis of Data From Individual Patients in Randomized Controlled Trials. Gastroenterology 2019;157:462–71. doi:10.1053/j.gastro.2019.04.014
20 Zimmermann-Fraedrich K, Sehner S, Rex DK, et al. Right-Sided Location Not Associated With Missed Colorectal Adenomas in an Individual-Level Reanalysis of Tandem Colonoscopy Studies. Gastroenterology 2019;157:660-671.e2. doi:10.1053/j.gastro.2019.05.011
Figure 1 – CADe is able to identify and localize the adenomatous lesion in real-time colonoscopy. The output appears on the same screen of the endoscopy system without affecting the routinely technique of the operator.
Figure 2 – Study flow-chart including clinical outcomes. *Relative Risk 1.34 (1.19-1.48) § Incidence Risk Ratio 1.53 (1.13-2.07)
Figure 3 – (Per-patient) Adenoma detection rate as well as (Per-patient) ADR by adenomas features. A) Adenoma Size Category, large (≥10mm) vs small (<10 mm); B) Adenoma Morphology, polypoid: [pedunculated (0-1p), sessile (0-1s) or mixed (0-1sp)] vs. non-polypoid lesions: [superficial slightly elevated (IIa), flat (IIb), superficial depressed (IIc), and excavated (III) types]. C) Sessile serrated Lesions, (SSL). D) Colon location, proximal colon (Cecum, ascending, and transverse) vs. distal colon (Descending, sigmoid, and rectum). The asterisk indicates statistically significant difference between the two arms.
Table 1. Patients’ characteristics by according to intervention arm.
CADe (341 patients)
Control (344 patients) P-Value
Mean Age (SD), years 61.5 (9.7) 61.1 (10.6) 0.442
Gender, n (%) 0.541
Female 169 (49.6) 179 (52.0)
Male 172 (50.4) 165 (49.6)
Indication for colonoscopy, n (%) 0.818
FIT+ 102 (29.9) 105 (30.5)
Primary CRC Screening 77 (22.6) 76 (22.1)
Surveillance 86 (25.2) 78 (22.7)
GI Symptoms 76 (22.3) 85 (24.7)
Mean BBPS score (SD)
Right Colon 2.4 (0.50) 2.4 (0.52) 0.748
Transverse 2.5 (0.50) 2.5 (0.52) 0.943
Left Colon 2.5 (0.51) 2.5 (0.50) 0.645 Adequate preparation (BBPS >2 in all segments) 339 (99.4) 342 (99.4) 0.999
Mean Insertion Time (IQR), minutes 9.0 (5-11) 8.1 (2-10) 0.056
Mean Inspection Time (IQR), sec 7.3 (6-8) 7.0 (6-8) 0.100 # of polyps of any histology per patient (range) 1.89 (0-13) 1.24 (0-9) <0.001 SD, Standard deviation; FIT, fecal immunochemical test; BBPS, Boston Bowel Preparation Scale, IQR, interquartile range. * The information was missing in 7 and 13 cases in CADe e Control arm, respectively.
Table 2. Detection rate according to intervention arm, as well as per patient distribution of adenomatous polyps according to morphology, size, and location.
Per patient analysis^
Total Number (N=685)
RR [95% CI] P-value
All adenomas and CRCs 187 (54.8) 139 (40.4) 326 (47.6)
1.30 (1.14-1.45)§ <0.001
Non-advanced adenomas 142 (41.6) 103 (29.9) 245 (35.8)
1.35 (1.13-1.57) 0.001
Advanced adenomas† 35 (10.3) 33 (7.3) 68 (9.9)
1.07 (0.68-1.66) 0.769
adenocarcinoma (CRC) 10 (2.9) 3 (0.9) 13 (1.9)
3.36 (0.93-12.11) 0.067
Sessile serrated Lesion 24 (7.0) ç 18 (5.2) 42 (6.1) ç
1.34 (0.75-2.37) 0.326
Non-neoplastic polyp* 68 (19.9) 57 (16.6) 125 (18.2)
1.20 (0.88-1.65) 0.254
Size category‡ <5 mm 115 (33.7) 91 (26.5) 206 (30.1)
1.26 (1.01-1.52) 0.038
6-9 mm 36 (10.6) 20 (5.8) 56 (8.2)
1.78 (1.09-2.86) 0.025
≥10 mm 36 (10.6) 28 (9.1) 64 (9.3)
1.29 (0.81-2.02) 0.278
Morphology†‡ Polypoid* 126 (37.3) 91 (26.5)
217 (31.9) 1.36 (1.12-1.62) 0.003
Non-polypoid**†† 90 (26.6) 63 (18.4)
153 (22.5) 1.42 (1.09-1.79) 0.010
Location‡ Proximal Colon°‡‡ 123 (36.1) 97 (28.2)
1.28 (1.03-1.59) 0.028
Distal Colon°° 109 (32.0) 72 (20.9)
1.53 (1.18-1.98) 0.001
^ It refers to proportion of patients with at least one adenoma or CRC, unless otherwise specified. §After adjustement for age, gender, and colonoscopy indication; crude RR for adenoma detection rate: 1.36 (1.16-1.59); crude RR for advanced neoplasia detection rate: 1.30 (0.86-1.99). *Normal, hyperplastic, inflammatory and others. ‡According to the size of the largest neoplastic lesion. †‡There were 4 cases with missing data: 3 in the CADe and 1 in the control group. *[elevated more than 2.5 mm above the mucosal layer: pedunculated (0-1p), sessile (0-1s) or mixed (0-1sp)]. **non-polypoid lesions: superficial slightly elevated (IIa), flat (IIb), superficial depressed (IIc), and excavated (III) types. †Advanced adenoma was defined as an adenoma of 10 mm or more, or as an adenoma (irrespective of size) with
at least 25% villous histology or with high-grade dysplasia. ç2 SSL with cytological dysplasia, corresponding to a cumulative SSL with dysplasia/all SSL ratio of 4.8% ††Including 29 (8.6%) CADe and 14 (4.1%) control group cases who had synchronous polypoid adenomas. ‡‡Including 45 (13.2%) CAD-e and 30 (8.7%) control cases who had synchronous adenomas in the distal colon. °Cecum, ascending, and transverse. °°Descending, sigmoid, and rectum
Table 3 – Per polyp analysis: mean number of adenomas per colonoscopy (APC) and Poisson regression analysis by polyp characteristics among study participants (n = 685). IRR, incidence risk ratio as adjusted for patient age, gender, and indication.
Per polyp analysis
IRR [95% CI]
APC (SD) APC (SD) Morphology†‡
Polypoid* 0.61 (1.20) 0.42 (0.92) 1.44 (1.05-1.96) Non-polypoid**†† 0.42 (0.94) 0.28 (0.83) 1.47 (1.00-2.15)
Size‡ <10 mm 0.92 (1.40) 0.62 (1.08) 1.50 (1.17-1.91) ≥10 mm 0.11 (0.35) 0.09 (0.31) 1.07 (0.66-1.74)
Location‡ Proximal Colon°‡‡ 0.60 (1.10) 0.45 (0.92) 1.35 (1.00-1.81) Distal Colon°° 0.43 (0.79) 0.26 (0.80 1.60 (1.14-2.07)
‡There were missing data in 3 CADe cases and 3 control case.*[elevated more than 2.5 mm above the mucosal layer: pedunculated (0-1p), sessile (0-1s) or mixed (0-1sp)]. **non-polypoid lesions: superficial slightly elevated (IIa), flat (IIb), superficial depressed (IIc), and excavated (III) types. ††Including 29 (8.5%) CADe and 14 (4.1%) control group cases who had synchronous polypoid adenomas. ‡‡Including 45 (13.2%) CAD-e and 30 (8.7%) control cases who had synchronous adenomas in the distal colon.°Cecum, ascending, and transverse. °°Descending, sigmoid, and rectum
What you need to know: Background and Context: Deep learning systems allow for real-time computer-aided detection (CADe) of polyps with high-accuracy, but these systems have not been tested in randomized trials.
New Findings: In randomized trial, inclusion of CADe in colonoscopy significantly increased adenoma detection rates and adenomas detected per colonoscopy, without increasing withdrawal time. Higher proportions of adenomas smaller than 5 mm and 6–9 mm were detected with CADe, regardless of morphology or location.
Limitations: This study was performed in an expert setting; studies are needed for inexperienced endoscopists.
Impact: Including CADe in colonoscopy examinations increases detection of adenomas without affecting safety.
Lay Summary: Including real-time, computer-aided detection of polyps in colonoscopy examination increases rates of adenoma detection by 30%.