Abbreviations
ADR:
adenoma detection rate
AFI:
autofluorescence imaging endoscopy
AI:
artificial intelligence
ASGE:
American Society for Gastrointestinal Endoscopy
BLI:
blue light imaging
CE:
chromoendoscopy
CI:
confidence interval
CRC:
colorectal cancer
EMR:
endoscopic mucosal resection
ESD:
endoscopic submucosal dissection
ETMI:
endoscopic trimodal imaging
FACILE:
Frankfurt Advanced Chromoendoscopic IBD LEsions
FAP:
familial adenomatous polyposis
FICE:
flexible spectral imaging color enhancement
FTRD:
full thickness resection device
GRADE:
Grading of Recommendations Assessment, Development and Evaluation
HD-WLE:
high definition white-light endoscopy
HGD:
high grade dysplasia
I-SCAN:
i-SCAN digital contrast
JNET:
Japan NBI Expert Team
LCI:
linked color imaging
LGD:
low grade dysplasia
LST:
laterally spreading tumor
MB-MMX:
methylene blue formulation
NBI:
narrow band imaging
NICE:
NBI International Colorectal Endoscopic
NG:
nongranular
PDR:
polyp detection rate
PICO:
patient, intervention, comparator, outcome
RCT:
randomized controlled trial
SD-WLE:
standard definition white-light endoscopy
SPS:
serrated polyposis syndrome
SSL:
sessile serrated lesion
UC:
ulcerative colitis
WASP:
Workgroup serrAted polypS and Polyposis
WLE:
white-light endoscopy
This Guideline is an official statement of the European Society of Gastrointestinal
Endoscopy (ESGE). It is a revision of the previously published 2014 Guideline addressing
the role of advanced endoscopic imaging for detection and differentiation of colorectal
neoplasia.
Introduction
Colonoscopy is the key examination technique in colorectal cancer (CRC) screening
programs for detection and treatment of early precursor lesions and timely diagnosis
of colorectal cancer [1 ]
[2 ]. The quality of colonoscopy, which depends on both bowel preparation and examination
technique, is the main determining factor that drives the protective effect of this
invasive examination in decreasing the societal disease burden [3 ]
[4 ]
[5 ].
Over the last 15 years, several new techniques to improve polyp detection and characterization
have been developed and studied [6 ]. For all these techniques, the possible financial burden, learning curve, and additional
cost need to be balanced against the potential benefit. In general, there is a potential
bias in the available literature given that it is impossible to blind the endoscopist
to the technique that is being studied. Therefore, even the setting of a fully randomized
trial, there is always a potential bias in favor of any technique that may affect
the performance of the endoscopists, even subconsciously.
This update of the previously published Guideline [7 ] aims to put into perspective the new evidence that has become available over the
last 5 years, and to provide statements on the possible role of advanced techniques
in polyp detection or characterization in the average risk and high risk populations.
The potential role of artificial intelligence (AI) in the detection and characterization
of colorectal lesions, including possible hazards of its implementation, has been
addressed for the first time.
With regard to training, in optical diagnosis of diminutive polyps, detection of colitis-associated
neoplasia, and prediction of invasion with larger polyps, we refer to the standardized
ESGE training curriculum. Although this is a work in progress, we anticipate that
the curriculum will be available in 2020 and want to include this defined standard
in the Guideline as a prerequisite for obtaining cognitive chromoendoscopy (CE) skills
for lesion characterization and detection.
Methods
The ESGE commissioned this Guideline (Guideline Committee chair, J.v.H.) and appointed
a guideline leader (R.B.), who invited the listed authors to participate in the project
development. The key questions were prepared by the coordinating team (R.B., E.D.,
J.E.E., M.P., M.K., C.H., H.N.) and were then approved by the other members. The coordinating
team established task force subgroups, based on the statements of the previous 2014
Guideline [7 ], each with its own leader, and divided the key topics among those task forces (Appendix 1 s ; see online-only Supplementary Material) with a specific focus on the update of literature
and revision of the statements.
The Guideline was developed during September 2018 and June 2019. The work included
telephone conferences, a face-to-face meeting, and online discussions, and additional
Delphi voting if necessary. In addition to the five task forces of the previous Guideline,
we included a sixth task force to address the role of artificial intelligence (AI)
in the detection and characterization of colorectal polyps. The task forces conducted
a literature search related to the following techniques: high definition endoscopy,
chromoendoscopy or dye-based endoscopy, virtual chromoendoscopy (narrow band imaging
[NBI], i-SCAN digital contrast [I-SCAN], flexible spectral imaging color enhancement
[FICE], and blue light imaging [BLI]), autofluorescence imaging (AFI) endoscopy, and
add-on devices. Techniques that have been under development or without clear clinical
implementation since the publication of the previous Guideline were not included (i. e.,
confocal endomicroscopy, endocytoscopy, optical coherence tomography). Key questions
were formulated using patient, intervention, comparator, outcome (PICO) methodology
[8 ].
The literature search was conducted through Medline (via Pubmed) and the Cochrane
Central Register of Controlled Trials up to June 2019. New evidence on each key question
was summarized in tables, using the Grading of Recommendations Assessment, Development
and Evaluation (GRADE) system [9 ]. Grading depends on the balance between the benefits and risk or burden of any health
intervention (Appendix 2 s ). Further details on guideline development have been previously reported [10 ].
The results of the search were presented to all group members during a meeting in
Prague on April 1st 2019. Subsequently, drafts were made by each task force chair
and distributed between the task force members for revision and online discussion.
Statements were created by consensus, or by Delphi voting of two rounds for task force
2.
In July 2019, a draft prepared by R.B. and all the task force chairs was sent to all
group members. After agreement of all members had been obtained, the manuscript was
reviewed by two external reviewers, Prof. Brian Saunders and Dr. David Tate. It was
then sent for further comments to the ESGE national societies and individual members.
It was then submitted to the journal Endoscopy for publication. The final revised manuscript was agreed upon by all the authors.
This Guideline was issued in 2019 and will be considered for update in 2024. Any interim
updates will be noted on the ESGE website: http://www.esge.com/esge-guidelines.html .
Evidence and statements
Evidence statements are compared to those of the previous 2014 Guideline [7 ]. The 2014 statements are shown in italic. The statements are grouped according to
the different task force topics.
Detection of colorectal neoplasia in the average risk population
Detection of colorectal neoplasia in the average risk population
2014 statements:
ESGE suggests the routine use of high definition white-light endoscopy systems for
detecting colorectal neoplasia in average risk populations (weak recommendation, moderate
quality evidence).
ESGE does not recommend routine use of virtual pancolonic chromoendoscopy, AFI, or
add-on devices for detecting colorectal neoplasia in average risk populations (strong
recommendation, high quality evidence).
2019 statement:
ESGE suggests that high definition endoscopy, and dye or virtual chromoendoscopy,
as well as add-on devices, can be used in average risk patients to increase the endoscopist’s
adenoma detection rate. However, their routine use must be balanced against costs
and practical considerations.
Weak recommendation, high quality evidence.
The term “average risk population” refers to patients undergoing screening colonoscopy
outside the setting of colitis or hereditary syndromes. Colorectal cancer screening
is performed on a large scale in Europe, and therefore a small increase in adenoma
detection may have a significant effect on the health care outcome of colorectal cancer
[11 ]. Nonetheless, because of the widespread use of colonoscopy for colorectal cancer
screening, the cost and practicality of advanced imaging techniques or add-on devices
must be taken into consideration to avoid excessive financial or organizational burdens.
High definition endoscopy
A 2011 meta-analysis of five studies including 4422 average risk patients showed a
3.5 % (95 % confidence interval [CI] 0.9 % – 6.1 %) incremental yield from high definition
white-light endoscopy (HD-WLE) over standard definition white-light endoscopy (SD-WLE)
for the detection of patients with at least one adenoma [12 ]. There were no differences between HD-WLE and SD-WLE for high risk adenomas. We
postulate that the difference in the fields of view of the endoscopes that were used
is unlikely to account for the increased yield observed with HD-WLE, because three
randomized controlled trials (RCTs) from two centers found no significant difference
in polyp detection rates between SD-WLE endoscopies with 140° and 170° fields of view
[13 ]
[14 ]
[15 ].
Also in a two-center RCT [16 ] published after the meta-analysis, the proportion of participants in whom adenomas
were detected with HD-WLE was higher as compared with SD-WLE (45.7 % vs. 38.6 %, P = 0.166). The difference was significant for patients with flat adenomas (9.5 % vs.
2.4 %, P = 0.003) and right-sided adenomas (34.0 % vs. 19.0 %, P = 0.001).
A recent RCT [17 ] comparing HD-WLE with SD-WLE in 1855 patients has shown a significant increase in
detection of sessile serrated lesions, also precursors for CRC (8.2 % vs. 3.8 %),
as well as adenocarcinomas (2.6 % vs. 0.5 %). However, in this study no difference
in adenoma detection rate (ADR) or polyp detection rate (PDR) was seen.
Two recent multicenter RCTs [18 ]
[19 ] have postulated that two generations of improvements in colonoscopes are necessary
to significantly increase ADR. The two RCTs compared the latest generation HD-WLE
colonoscopes from one company (Olympus 190C) against standard definition next-to-last
generation colonoscopes (Olympus 160C) in both a hospital [18 ] and in a private practice [19 ] setting. Results from the two trials were not fully concordant. In the hospital
setting, a significant decrease in adenoma miss rates was found with high definition
colonoscopes (16.6 %, 95 %CI 13.0 % – 20.1 % vs. 30.2 %, 95 %CI 25.9 % – 34.6 %; P < 0.001) as well as a significant increase in ADR (43.8 % vs. 36.5 %, P = 0.03) [18 ]. In the private practice setting [19 ] however, the ADR difference in favor of the latest-generation colonoscope did not
reach statistical significance (32 % vs. 28 %, P = 0.10). The detection of diminutive polyps (< 5 mm) was significantly increased
(22.5 % vs. 15.6 %, P < 0.001) for HD-WLE, as well as the adenoma per patient rate (all adenomas/all patients:
0.57 vs. 0.47, P < 0.001). Details of these RCTs are available in Table 1 s (see Appendix 3 s ; online-only Supplementary Material).
The cost – effectiveness of using HD-WLE in routine practice was not studied. High
definition colonoscopes are available from all major manufacturers.
Based on the above results with moderate-to-high quality evidence, we can conclude
that high definition systems may be of benefit to improve polyp and adenoma detection,
although trial results are not entirely consistent.
Virtual chromoendoscopy
Narrow band imaging (NBI)
Four meta-analyses and one Cochrane systematic review of RCTs compared detection of
colorectal lesions in average risk populations using WLE and NBI [20 ]
[21 ]
[22 ]
[23 ]
[24 ]. When considering HD-WLE versus HD-NBI, none of these showed a significant difference
in adenoma detection rate between the two technologies. HD-NBI showed a small increase
in detection rate when compared to SD-WLE only.
A very recent meta-analysis [25 ] comprised data of 4491 individual patients from 11 RCTs. In this study, high definition
NBI (HD-NBI) showed a significant increase in unadjusted odds ratio for adenoma detection
compared to HD-WLE (OR 1.14, 95 %CI 1.01 – 1.29, P = 0.04; ADRs, HD-WLE 42.3 % vs. HD-NBI 45.2 %). When subanalyses were performed,
NBI showed an increased detection only when preparation was best (compared to average).
Moreover, it was only second-generation NBI, with a brighter light, that significantly
increased ADR, and not the first-generation (second-generation NBI OR 1.28, 95 %CI
1.05 – 1.56, P = 0.02).
We can therefore conclude that the additional value of NBI in polyp detection is rather
marginal, taking into consideration the marginal significance in the meta-analysis.
The introduction of better imaging quality with HD systems has probably a more important
role.
i-SCAN digital contrast (I-SCAN), flexible spectral imaging color enhancement (FICE),
blue light imaging (BLI), and linked color imaging (LCI)
One meta-analysis, published in 2014 and including 5 studies with 3032 patients [23 ], compared HD-FICE and HD-i-SCAN versus HD-WLE in the detection of adenomas and found
no additional detection with these advanced techniques (RR 1.09, 95 %CI 0.97 – 1.23).
An RCT [26 ] published after the meta-analysis showed a favorable result for I-SCAN technology,
with a significantly higher ADR in the I-SCAN group compared to the HD-WLE colonoscopy
group (47.2 % vs. 37.7 %, P = 0.01). This result, however, was mainly due to an increased detection rate of diminutive,
flat, and right-sided adenomas.
Data on BLI and LCI for the detection of colorectal lesions are preliminary. Recent
RCTs on LCI showed an increased per-patient ADR compared to HD-WLE (37 % vs. 28 %)
[27 ], as well as a reduction in the miss rate in the right colon [28 ]. The single recent RCT on BLI [29 ] showed an increased mean adenoma per patient rate (mean ± standard deviation [SD]
1.27 ± 1.73 vs. 1.01 ± 1.36, P = 0.008), but no increase in ADR or PDR compared to HD-WLE.
Details of the most important studies are available in Table 2 s .
In conclusion, data on advanced imaging with these techniques is scarce and the beneficial
effect in terms of incremental polyp detection seems to be clinically marginal.
Autofluorescence imaging (AFI) endoscopy
One meta-analysis published in 2015 [30 ], including six RCTs and 1199 patients, evaluated AFI for the detection of colorectal
neoplasia in average risk patients, and showed no significant difference between AFI
and WLE in ADR or PDR (ADR, OR 1.01, 95 %CI 0.74 – 1.37, P = 0.96; PDR, OR 0.86, 95 %CI 0.57 – 1.30, P = 0.71), with no significant heterogeneity among the studies (P = 0.67, I
2 = 0).
One recently published RCT [31 ] focused on the role of updated AFI in the detection of flat lesions and showed a
significant increase in the detection of right-sided flat lesions (adenomas and carcinoma,
not sessile serrated polyps) (0.87, 95 %CI 0.78 – 0.97 vs 0.53, 95 %CI 0.46 – 0.61),
but no increase in overall ADR or PDR.
Details of these two studies are available in Table 3 s .
Based on the findings of the meta-analysis there seems to be no major additional value
of AFI for polyp detection in the average risk population. In addition, the system
is not commercially available.
Add-on devices
In 2018, two network meta-analyses investigating the efficacy of add-on devices to
improve ADR (cap, Endocuff, Endorings) were published [32 ]
[33 ] (Table 4 s ).
One network meta-analysis, including 25 RCTs and 16 103 patients [32 ], showed an overall slight increase in ADR for add-on devices compared to standard
colonoscopy (39.3 % vs. 35.1 %; relative risk [RR] 1.13, 95 %CI 1.03 – 1.23; P = 0.007). When individual devices were considered, both Endocuff versus HD-WLE and
Endorings versus standard colonoscopy showed a small but significant improvement in
ADR; these however would be of benefit mostly for already high-performing endoscopists.
The use of a short transparent cap at the tip of the endoscope resulted in a statistically
insignificant increase in ADR compared to HD-WLE (37 % vs. 34.3 %; RR 1.07, 95 %CI
0.96 – 1.19; P = 0.19). However, the considerable heterogeneity (I
2 = 89 %) should lead to cautious interpretation of these results. Subgroup analysis
revealed a substantial increase of ADR and PDR of lesions ≤ 5 mm (RR 1.53, 95 %CI
1.13 – 1.71, RR 1.38, 95 %CI 1.10 – 1.43, respectively).
The second network meta-analysis [33 ] included 10 studies reporting on 6047 patients and showed, in contrast to the first,
an overall increase in ADR for Endocuff in comparison to HD-WLE (OR 1.36, 95 %CI 1.12 – 1.60;
P = 0.001), but when a subgroup analysis was performed this was only significant in
low-performing endoscopists (for ADR < 25 %: OR 1.85, 95 %CI 1.35 – 2.53, P = 0.0001).
Most RCTs do not report cost–effectiveness data and this aspect has not yet been evaluated
systematically.
Based on the available data, the evidence for general use of add-on devices is rather
weak and cost–effectiveness has never been well assessed. It might however have a
role in helping low-performers to reach the important ADR threshold of 25 % [4 ].
Dye-based chromoendoscopy (CE)
A recently updated Cochrane systematic review of 2016 [34 ] analyzed 7 RCTs (total 2727 patients) that assessed the role of dye-based CE in
detecting colorectal lesions outside the setting of polyposis or colitis. Pancolonic
CE significantly increased the number of patients with at least one polyp detected
(OR 1.87, 95 %CI 1.51 – 2.30) and of those with at least one neoplastic polyp (adenoma
or carcinoma) detected (OR 1.53, 95 %CI 1.31 – 1.79). Limitations of the systematic
review were the lack of blinding in the RCTs, and the significant heterogeneity observed
between the studies. Indeed, quality of evidence was graded as low in this review.
Since the publication of that Cochrane systematic review, two large multicenter RCTs
have been published. The first [35 ], including 1065 patients, showed an increase in the mean adenoma per patient rate
(0.79 vs. 0.64, P = 0.005), but not in ADR (40.4 % vs. 37.5 %; OR 1.13, 95 %CI 0.87 – 1.48; P = 0.35) or sessile serrated lesion detection rate, using routine pancolonic CE compared
to HD-WLE.
A recent phase 3 multicenter RCT [36 ] has evaluated the role of a novel pH- and time-dependent peroral methylene blue
formulation (MB-MMX) that is delivered in pills taken during the bowel preparation
phase. This RCT enrolled 1205 patients undergoing screening or surveillance colonoscopy
and found an increased overall ADR in the MB-MMX group compared to the placebo group
(56.29 vs 47.81 %; OR 1.46, 95 %CI 1.09 – 1.96). The MB-MMX group showed a higher
number of patients with adenomas ≤ 5 mm (37.11 % vs. 30.90 %; OR 1.36, 95 %CI 1.01 – 1.83).
Details of the abovementioned studies are available in Table 5 s .
We can conclude that chromoendoscopy increases ADR and PDR; however its systematic
implementation may be hampered in daily practice because of practical considerations
and additional costs. The use of MB-MMX may help to overcome these.
Detection of colorectal neoplasia in high risk populations with hereditary syndromes
Detection of colorectal neoplasia in high risk populations with hereditary syndromes
2014 statements:
ESGE recommends the routine use of high definition pancolonic chromoendoscopy in patients
with known or suspected Lynch syndrome (conventional chromoendoscopy, NBI, i-SCAN)
or serrated polyposis syndrome (conventional chromoendoscopy, NBI) (strong recommendation,
low quality evidence).
ESGE does not make any recommendation for the use of advanced endoscopic imaging in
patients with suspected or known familial adenomatous polyposis (FAP) including attenuated
and MUTYH-associated polyposis (insufficient evidence to make a recommendation).
2019 statements:
ESGE recommends the routine use of high definition systems in individuals with Lynch
syndrome.
Strong recommendation, high quality evidence.
ESGE suggests that the use of virtual chromoendoscopy may be of benefit in individuals
with Lynch syndrome undergoing colonoscopy; however its routine use must be balanced
against costs, training, and other practical considerations.
Weak recommendation, moderate quality evidence.
ESGE suggests the use of high definition systems and dye-based chromoendoscopy in
the diagnosis and surveillance of individuals with serrated polyposis syndrome; however
routine use must be balanced against costs, training, and practical considerations.
Weak recommendation, moderate quality evidence.
ESGE does not recommend the systematic use of dye-based nor virtual chromoendoscopy
for familial adenomatous polyposis (FAP), MUTYH -associated polyposis, or hamartomatous polyposis.
Strong recommendation, moderate quality evidence.
Lynch syndrome
Lynch syndrome is the most common cause of hereditary colorectal cancer (CRC). It
is an autosomal dominant disorder caused by germline mutations in the DNA mismatch
repair (MMR) genes (i. e., MLH1 , MSH2 , MSH6 , PMS2 , and EpCAM ). An accelerated progression from adenoma to CRC has been described, and often the
adenomas display advanced histological features (i. e., high grade dysplasia or a
villous component), are frequently flat in morphology, and located in the proximal
colon, compared with sporadic adenomas. An intensive surveillance strategy with annual
or biennial colonoscopy starting at early ages has reduced both the incidence and
mortality associated with CRC. A high detection rate for these aggressive adenomas
is especially important to minimize the risk of interval CRC.
In total, seven studies comparing indigo carmine CE with WLE in patients with Lynch
syndrome have been published [37 ]
[38 ]
[39 ]
[40 ]
[41 ]
[42 ]
[43 ] (Table 6 s ). Three single-center studies with a small number of patients in a back-to-back design
showed that CE was superior to SD-WLE, with an adenoma miss rate ranging from 61 %
to 74 % [37 ]
[38 ]
[41 ]. A recent back-to-back multicenter study, where the second pass was performed by
a different endoscopist in order to minimize the second inspection bias, again showed
superiority of SD-CE over SD-WLE (ADRs of 41 % and 23 %, respectively; adenoma miss
rate 52 %). Nevertheless, the study had no comparator arm, was slightly underpowered
(β-risk of 26 %) and the withdrawal time during CE was twice that of WLE [39 ]. All these results are methodologically flawed by the back-to-back design that may
lead to an overestimation of the effect of CE over WLE.
There are three trials with a control arm. A study by Stoffel et al. included 54 patients
in four centers [40 ]. After the first pass with SD-WLE, 28 patients were randomly allocated to a second
pass with CE and 27 to a second pass with an intensive 20-minute inspection; no significant
difference in adenoma miss rate was shown.
Very recently, two well-powered randomized, multicenter, controlled studies with a
comparator arm were published. Haanstra et al. showed no differences in neoplasia
detection rate between CE and WLE in 246 Lynch patients, either at baseline (27 %
vs. 30 %, respectively; P = 0.56) or in the 2-year follow-up colonoscopy (26 % vs. 28 %, respectively; P = 0.81) [42 ]. This study is limited by the fact that CE was applied only proximal to the splenic
flexure and that the study extended over a very long recruitment period (10 years)
which may entail important variability in procedure performances and ability for detecting
colorectal lesions. Rivero-Sánchez et al. performed a study with only HD endoscopes
and high-detector endoscopists in 256 Lynch patients in 14 different hospitals, and
showed that ADR was statistically not different between HD-CE and HD-WLE (34.4 % [95 %CI
26.4 % – 43.3 %] vs. 28.1 % [95 %CI 21.1 % – 36.4 %], P = 0.28) [43 ]. In both trials, CE was more time-consuming and detected more clinically irrelevant
lesions.
In Lynch patients, three single-center back-to-back studies were performed with high
definition virtual CE, which appeared to be superior to HD-WLE for polyp detection
[44 ]
[45 ]. East et al. showed in a nonrandomized back-to-back study in 62 Lynch patients that
during a second inspection, with NBI, additional adenomas were detected in 17 /62
patients (27 %). In this study, ADR increased to 26 /62 (42 %) after both WLE and
NBI: 9/62 patients had at least one adenoma detected that was missed during the first
inspection with WLE [44 ]. Bisschops et al. showed in a randomized crossover study in 61 Lynch patients that
the adenoma miss rate was significantly lower when I-SCAN was used, in comparison
to HD-WLE (12 % vs. 62 %) [45 ]. Both studies were conducted by a single expert endoscopist and in the second study,
the ADR was relatively low for HD-WLE inspection (19 %).
On the other hand, virtual CE appears to be inferior to dye-based CE in two back-to-back
studies. In a German cohort study the incremental yield of CE versus SD-WLE (n = 47)
and NBI (n = 62) was assessed, showing a higher detection with CE during second inspection
[41 ]. Very recently, a study comparing NBI to CE in a back-to-back design has been published
as an abstract. This multicenter French study, in 138 Lynch patients, showed an adenoma
miss rate of 48 % for the third generation of HD-NBI devices (Exera III, 190 series)
when followed by a second pass with dye-based CE by the same endoscopist. The authors
concluded that although NBI colonoscopy is less time-consuming, it cannot be recommended
to replace dye-based CE in Lynch syndrome patients [46 ].
Finally, one study in 75 patients compared AFI, with the Xillix system (XillixTechnologies
Corporation, Richmond, British Columbia, Canada), to WLE in a crossover trial, showing
a better detection of adenomas for AFI (92 % vs. 68 % for WLE) [47 ].
Details for the most important studies are available in Table 7 s .
In conclusion, evidence suggests a benefit of dye-based CE in Lynch syndrome patients
at the expense of longer procedure times. However, most of the studies were performed
with standard definition endoscopes, had a small and heterogeneous sample, and a nonrandomized
back-to-back design that may have led to a bias in favor of dye-based CE. Recent evidence
from two well-powered multicenter trials with a parallel design have shown no differences
in ADR between WLE and dye-based CE [42 ]
[43 ]. This possibly implies that a thorough inspection by high detector endoscopists
and using high definition endoscopes might decrease the advantageous effect of dye-based
CE in Lynch patients. These two RCTs are the reason for a slight discrepancy between
the recommendations in this Guideline and the recently published Guideline on the
management of polyposis syndromes [48 ], that also included dye-based CE as a suggestion. However the new evidence was not
available at the time of development of that Guideline. On the other hand, two studies
have reported superiority of virtual CE (NBI and I-SCAN) over WLE. Conversely, two
other studies have shown that dye-based CE was superior to virtual CE. Most of these
studies have methodological limitations such as back-to-back design, the second pass
being performed by the same expert endoscopist, or there being a low ADR in the first
pass.
Taking this into consideration, ESGE recommends at least the use of HD endoscopes
in Lynch patients and suggests in addition that, in view of the evidence, advanced
imaging techniques such as virtual chromoendoscopy can be useful.
Serrated polyposis syndrome (SPS)
Serrated polyposis syndrome (SPS) has emerged as the most frequent colorectal polyposis
syndrome. This entity is associated with an increased risk of CRC and is often grouped
with the hereditary polyposis syndromes although no underlying gene defect has been
identified yet.
Although recent studies show an increase in SPS prevalence [49 ]
[50 ]
[51 ], attributed to major clinical and pathological awareness and better endoscopic diagnostic
accuracy [52 ]
[53 ], SPS remains an underdiagnosed entity [54 ]. SPS diagnosis depends directly on the capacity for detecting serrated lesions (SLs),
which are often easily overlooked due to their imperceptibility [51 ]. In a fecal immunochemical test (FIT)-based CRC screening program, a reassessment
colonoscopy within 1 year after a screening colonoscopy tripled the number of patients
diagnosed with SPS. Use of CE, either dye-based or virtual, at reassessment colonoscopy
was associated with a higher detection rate of serrated lesions, but not of adenomas
[55 ].
Recently, a multicenter randomized back-to-back study evaluated the usefulness of
dye-based CE with indigo carmine for the detection of colonic polyps in SPS patients
under surveillance [56 ]. Patients were randomly assigned to a group: one received two HD-WLE examinations
(n = 43) and the other received HD-WLE followed by 0.4 % indigo carmine CE (n = 43).
This study demonstrated a significantly higher additional polyp detection rate in
the HD-CE group (0.39, 95 %CI 0.35 – 0.44) than in the HD-WLE group (0.22, 95 %CI
0.18 – 0.27, P < 0.001). HD-CE detected more serrated lesions than HD-WLE (40 % vs. 24 %, P = 0.001), more serrated lesions proximal to the sigmoid (40 % vs. 21 %, P = 0.001), and more > 5-mm serrated lesions proximal to the sigmoid (37 % vs. 18 %;
P = 0.013). Over 70 % of additional serrated lesions detected by CE were hyperplastic
polyps and at least two-thirds of them were located proximal to the sigmoid colon.
Detection of adenomas and serrated lesions > 10 mm in size did not differ significantly
between groups. The additional detection rate for SSP was higher in the HD-CE group
(0.29 in HD-CE vs. 0.13 in HD-WL, P = 0.059) but not statistically significant. In a multivariate logistic regression
analysis, only use of HD-CE was independently associated with an increase in polyp
detection throughout the colon.
The role of virtual CE (i. e., NBI) in SPS has been evaluated in two randomized crossover
studies. A first single-center study including 22 patients showed that NBI had a lower
polyp miss rate than high resolution WLE (10 % vs. 36 %); however this was not confirmed
in a second multicenter study including 52 SPS patients (20 % vs. 29 %; P = 0.065) [57 ]
[58 ]. The authors explained this contradictory result by the fact that the pilot study
was performed by a single endoscopist, at a single institution and with older equipment.
A recent multicenter prospective randomized controlled trial evaluated the usefulness
of Endocuff-assisted colonoscopy in SPS surveillance, showing no increase in detection
of sessile serrated lesions, adenomas, or polyps overall [59 ].
Details of the abovementioned studies are available in Table 8 s .
Thus, based on the abovementioned single RCT [56 ], the use of dye-based CE improves polyp detection and could be considered in the
surveillance of SPS patients. However, its routine use must be balanced against practical
considerations.
Detection and differentiation of colorectal neoplasia in inflammatory bowel disease
(IBD)
Detection and differentiation of colorectal neoplasia in inflammatory bowel disease
(IBD)
Patients with long-standing or extensive ulcerative colitis (UC) or Crohn’s disease
are at an increased risk of developing CRC compared to the average risk population.
Accordingly, regular and extensive surveillance colonoscopies are recommended [60 ]
[61 ]. In this context, advanced endoscopic imaging may be of benefit by (i) increasing
the neoplasia detection rate; (ii) improving the differentiation of lesions (colitis-associated
neoplasia, sporadic neoplasia, and non-neoplastic lesions); and (iii) reducing the
number of unnecessary biopsies.
2014 statements:
ESGE recommends the routine use of 0.1 % methylene blue or 0.1 % – 0.5 % indigo carmine
pancolonic chromoendoscopy with targeted biopsies for neoplasia surveillance in patients
with long-standing colitis. In appropriately trained hands, in the situation of quiescent
disease activity and adequate bowel preparation, nontargeted four-quadrant biopsies
can be abandoned (strong recommendation, high-quality evidence).
ESGE found insufficient evidence to recommend for or against the use of virtual chromoendoscopy
or autofluorescence imaging (AFI) for the detection of colorectal neoplasia in inflammatory
bowel disease (insufficient evidence to make a recommendation).
2019 statements:
ESGE recommends the routine use of dye-based pancolonic chromoendoscopy or virtual
chromoendoscopy with targeted biopsies for neoplasia surveillance in patients with
long-standing colitis, in the situation of quiescent disease activity and adequate
bowel preparation.
Strong recommendation, moderate quality evidence.
ESGE recommends that after proper training in colonoscopy has been obtained, as defined
in the ESGE curriculum, in the situation of quiescent disease activity and adequate
bowel preparation, nontargeted four-quadrant biopsies can be abandoned.
Strong recommendation, high quality evidence.
ESGE suggests that in the case of high risk patients with a personal history of colonic
neoplasia, tubular-appearing colon, strictures, or primary sclerosing cholangitis,
chromoendoscopy-targeted biopsies can be combined with four-quadrant nontargeted biopsies
every 10 cm in the colon.
Weak recommendation, low quality evidence.
In general, surveillance of long-standing colitis can only be accurately performed
in the absence of disease activity and with an adequate bowel preparation. Indeed,
all the imaging studies mentioned below only apply to patients with long-standing
colitis undergoing surveillance in the setting of quiescent disease activity and adequate
bowel preparation. The use of dye-based or virtual CE is technically cumbersome in
the presence of active colitis, multiple inflammatory or post-inflammatory polyps,
or poor bowel preparation.
SD-WLE or HD-WLE versus dye-based CE
Overall, in eight prospective studies comparing dye-based CE with SD-WLE, the former
consistently increased the proportion of patients found with dysplasia by a factor
of 2.08 – 3.26 [62 ]
[63 ]
[64 ]
[65 ]
[66 ]. A meta-analysis showed a pooled incremental yield of CE with random biopsies over
SD-WLE with random biopsies for the detection of patients with neoplasia of 7 % (95 %CI
3.2 % – 11.3 %). Moreover, the difference in proportion of lesions detected by targeted
biopsies only was 44 % (95 %CI 28.6 % – 59.1 %) in favor of dye-based CE [64 ]. This finding has been confirmed by a new retrospective cohort study including 78
patients with ulcerative colitis [67 ] in which CE visualized dysplastic lesions in 50 patients, including 34 new lesions
not visualized on the index SD-WLE examination. A prospective longitudinal study included
55 patients with ulcerative colitis and identified 44 dysplastic lesions in 24 patients:
6 were detected by random biopsy, 11 by WLE, and 27 by CE [68 ]. CE and targeted WLE were more likely than random biopsies to detect dysplasia,
and CE was superior to SD-WLE (OR 2.4, 95 %CI 1.4 – 4.0). One retrospective cohort
study including 2242 colonoscopies demonstrated equal dysplasia detection rates for
CE and WLE with random biopsies (11 % vs. 10 %, P = 0.80) [69 ].
Most recently CE has been evaluated for neoplasia detection and characterization in
long-standing colitis in a more real-life setting than that of a randomized controlled
trial with only expert endoscopists [70 ]. In this multicenter prospective cohort study including 350 patients, 41.5 % of
colonoscopies were performed with standard definition endoscopes. The overall dysplasia
miss rate for combined HD-WLE and SD-WLE was 40/94 (57.4 % incremental yield for CE).
The CE incremental detection yield for dysplasia was comparable between standard definition
and high definition (51.5 % vs. 52.3 %, P = 0.30) and statistically not different between expert and nonexpert endoscopists
(18.5 % vs. 13.1 %, P = 0.2).
Although this last study did not show a difference between SD-CE and HD-CE detection
of neoplasia, the additional value of high definition endoscopy in detecting ulcerative
colitis-related neoplasia has become clearer more recently, and seems to indicate
that CE increases detection only when standard definition endoscopy is used as opposed
to high definition. A recent meta-analysis of 10 studies (494 patients) compared dye-based
CE with SD-WLE and HD-WLE [71 ]. Of these 6 were RCTs (3 on SD-WLE and 3 on HD-WLE). The proportion of patients
diagnosed with dysplasia using CE was 17 % as compared with 11 % for WLE. When analyzed
separately, CE was more effective at identifying dysplasia than SD-WLE (RR 2.12, 95 %Cl
1.15 – 3.91); however CE was not more effective as compared with HD-WLE (RR 1.36,
95 %CI, 0.84 – 2.18). Based on this meta-analysis, non-RCTs demonstrated a benefit
of CE over SD-WLE and HD-WLE, whereas RCTs showed a small benefit of CE over SD-WLE,
but not over HD-WLE. In addition, two other meta-analyses comparing different advanced
techniques point in the same direction. One recent systematic review comparing CE
to other techniques (SD-WLE, HD-WLE, HD-NBI, or HD-I-SCAN), included 10 randomized
trials with 1500 participants [72 ]. CE was associated with higher detection of patients with dysplasia as compared
with other techniques. However, subgroup analyses confirmed this effect only in comparison
with SD-WLE (RR 2.12, 95 %CI 1.15 – 3.91). These findings have been confirmed by another
network meta-analysis including only 8 parallel-group RCTs with 924 patients [73 ] and comparing HD-WLE, SD-WLE, SD-CE, HD-CE, and HD-NBI for detection of neoplasia
in long-standing colitis. The network analysis did not find any single technique to
be statistically superior. CE was probably more effective than SD-WLE for detecting
any dysplasia (OR 2.37, 95 %CI 0.81 – 6.94). Finally, a recent prospective RCT compared
HD-WLE alone (n = 90) with high definition dye-based CE (n = 90), and virtual CE with
I-SCAN (n = 90) for detection of neoplastic lesions during IBD surveillance colonoscopy
[74 ]. The HD-WLE neoplasia detection rate (25.5 %) was noninferior either to dye-based
(24.4 %) or to virtual CE (15.5 %) for detection of all neoplastic lesions (P = 0.91).
Details of the abovementioned studies with SD endoscopy and HD endoscopy are available
in Tables 9 s and 10 s .
Limitations of dye-based CE in the context of long-standing colitis surveillance need
to be mentioned. There is no proof that better detection of neoplasia by CE results
in the reduction of CRC mortality or decreased risk of interval CRC. Data on cost – effectiveness
are also limited; however a reduction in the number of colonoscopies and histological
samples could be achieved by risk stratification [75 ]. One study assessed the cost–effectiveness of CE in comparison with WLE or no endoscopy
for CRC surveillance in patients with ulcerative colitis, using a decision-analytic
state-transition (Markov) model with a Monte Carlo simulation [76 ]. CE was found to be more effective and less expensive than WLE at all surveillance
intervals. However, compared with no surveillance, CE was cost-effective only at 7-year
surveillance intervals, with an incremental cost – effectiveness ratio of $77 176.
At sensitivity levels of > 0.23 for dysplasia detection and cost < $2200, CE was the
most cost-effective strategy, regardless of the level of sensitivity of WLE. The estimated
population lifetime risk of developing CRC ranged from 2.5 % (annual CE) to 5.9 %
(CE every 10 years).
Virtual CE
Three RCTs compared NBI in all cases with HD-WLE for the detection of neoplasia in
long-standing IBD. Regardless of the generation of the NBI device and the level of
definition of colonoscopes used, virtual CE did not significantly increase the detection
rate of neoplastic lesions as compared with WLE [77 ]
[78 ]
[79 ]. However, virtual CE with targeted biopsies alone yielded neoplasia detection rates
comparable to WLE with targeted and random four-quadrant biopsies (mean number of
biopsies per patient: 0.5 – 3.5 in NBI with targeted biopsies only, and 24.6 – 38.3
in WLE with targeted and random biopsies).
Two RCTs compared a HD-NBI system with high definition dye-based CE, both without
nontargeted biopsies, for the detection of neoplasia in long-standing UC. The first,
single-center, crossover RCT comparing neoplasia miss rates with HD-NBI and high definition
dye-based CE [80 ], showed a considerably higher miss rate of neoplastic lesions with HD-NBI as compared
with high definition dye-based CE (31.8 % and 13.6 %, respectively). However, this
study was not adequately powered to show a statistical significance. The second was
a recent multicenter RCT that compared HD-CE with HD-NBI in 131 patients with UC in
a 1:1 randomization [81 ]. Mean numbers of neoplastic lesions per colonoscopy were 0.47 for CE and 0.32 for
NBI (P = 0.992). The neoplasia detection rate did not differ significantly between CE and
NBI (21.2 % vs. 21.5 %, respectively). The per-lesion neoplasia detection was 17.4 %
for CE and 16.3 % for NBI (P = 0.793) and the total procedural time was on average 7 minutes shorter in the NBI
group.
One study compared I-SCAN as virtual CE with HD-WLE and dye-based HD-CE. There was
no significant difference between three groups of patients with neoplasia detection
(15.5 %, 25.5 %, and 24.4 % respectively). Although 10 % noninferiority was just passed
statistically, caution should be exercised as the difference might still be clinically
relevant [74 ]. A recent meta-analysis has highlighted the potential role of virtual CE for dysplasia
detection in IBD. For the comparison of NBI versus WLE, 4 studies with 305 patients
were included. The analysis showed no differences in per-patient neoplasia detection
(OR 0.97, 95 %CI 0.62 – 1.53) and per-neoplastic lesion detection (OR 0.94,95 %CI
0.63 – 1.4) [82 ].
Two studies (one of them an RCT) compared HD-WLE with AFI for the detection of colorectal
neoplasia in IBD [79 ]
[83 ]. A pilot study [83 ] showed that protruding lesions with a low AFI signal were significantly more likely
to be neoplastic than lesions with a high AFI signal (45.0 % vs. 13.3 %, respectively;
P = 0.043). In the RCT, the miss rate for neoplastic lesions was statistically significantly
lower with AFI compared with HD-WLE (0 % vs. 50 %, P = 0.036) [79 ]. It should be noted that inadequate bowel preparation and active inflammation interrupt
tissue AFI, resulting in discoloration on AFI and resembling neoplasia. Another recent
RCT confirmed that AFI did not meet criteria for proceeding to a large noninferiority
trial and that the existing AFI imaging technology should not be further investigated
as an alternative dysplasia surveillance method [84 ].
Details of the abovementioned studies are available in Table 11 s .
Role of biopsies
A limited diagnostic yield of four-quadrant biopsies in comparison to targeted biopsies
has already been shown in the previous Guideline. A pooled sensitivity for the detection
of neoplasia with CE-targeted biopsies only was 86 % (range 71 % – 100 %) [37 ]
[62 ]
[63 ]
[65 ]
[66 ]
[85 ]
[86 ]
[87 ]. The median numbers of targeted and targeted plus random biopsies were 1.3 (range
0.28 – 14.2) and 34.3 (range 7.0 – 42.2), respectively. Therefore, the number of biopsies
needed during dye-based CE surveillance of long-standing colitis can be significantly
reduced if targeted biopsies are taken. The yield and clinical impact of random biopsies
were also assessed in a retrospective analysis of 1010 colonoscopies [88 ]. Overall, 11 722 random biopsies (median 29) were taken in 466 surveillance colonoscopies.
Neoplasia was detected in 88 colonoscopies: in 75 (85 %) by targeted biopsies, in
8 (9.1 %) by both targeted and random biopsies, and in 5 (5.7 %) by random biopsies
in 4 patients (7.5 % of 53 with detected neoplasia). In 94 % of colonoscopies, neoplasia
was macroscopically visible. An RCT comparing the rates of neoplasia detection by
targeted versus random biopsies in 246 patients with UC found the mean number of biopsies
containing neoplastic tissue per colonoscopy to be 0.211 (24 of 114) in the target
group and 0.168 (18 of 107) in the random group [89 ]. Neoplasia was detected in 11.4 % of patients in the target group and 9.3 % of patients
in the random group (P = 0.617). Another, nonrandomized study evaluating different surveillance strategies
in 454 IBD patients showed a neoplasia detection rate of 8.2 % in the random biopsy
group compared to 19.1 % in the targeted biopsy group [90 ]. Recently, a study with 1000 colonoscopies showed neoplasia in 82 patients diagnosed
by targeted biopsies or removed lesions [91 ]. Dysplasia was detected by random biopsies in 7 patients and in 12 additional patients
by random biopsies only. The yield of neoplasia by random biopsies only was 0.2 %
per-biopsy, 1.2 % per-colonoscopy and 12.8 % per-patient with neoplasia. Dysplasia
detected by random biopsies was associated with a personal history of neoplasia, a
tubular appearing colon, or the presence of primary sclerosing cholangitis. It may
therefore be careful and advisable to combine random biopsies with dye-based or virtual
CE-targeted biopsies in these high risk patients. In addition, since it may be difficult
to locate again small lesions with dysplasia, it may be advisable in the case of lesions
< 10 mm to resect the lesion entirely to facilitate patient management.
Details of the abovementioned studies are available in Table 12 s .
Conclusions: detection of neoplasia in IBD
In conclusion, the literature on advanced imaging in the detection of colitis-associated
neoplasia is large but also heterogeneous as illustrated by the several meta-analyses.
Although several meta-analyses have been performed on the same literature and sometimes
seem to contradict each other, it seems reasonable to accept the additional value
of dye-based CE. Recent evidence with HD endoscopes point to the fact that virtual
chromoendoscopy also may be equally effective. Although the Spanish real-life study
[70 ] did not show a clear difference in dysplasia detection between expert and nonexpert
(18.5 % vs. 13.1 %, P = 0.20) and did not show a significant learning curve for CE, it is conceivable that
lesion recognition by virtual CE is facilitated by previous dye-based CE. In fact,
all investigators involved in the virtual CE trials had previous experience with dye-based
CE. In standard risk patients, the evidence clearly points to abandoning nontargeted
random biopsies. The additional value of using virtual CE lies in the fact that it
is time-saving (7 minutes less on average than dye-based CE [81 ]) and may facilitate surveillance in cases of poorer bowel preparation.
Neoplastic versus non-neoplastic lesions in IBD
2014 statement:
ESGE recommends taking biopsies from flat mucosa surrounding neoplastic lesions and
taking biopsies from or resecting all suspicious lesions identified at neoplasia surveillance
in long-standing colitis, because there is no evidence that nonmagnified conventional
or virtual chromoendoscopy can reliably differentiate between colitis-associated and
sporadic neoplasia or between neoplastic and non-neoplastic lesions (strong recommendation,
low to moderate quality evidence).
2019 statement:
ESGE recommends using advanced imaging to assess the borders of lesions in previously
colitic mucosa, to assess resectability. If optical diagnosis is used for lesion characterization
of visible lesions, ESGE recommends that the suspicion of neoplasia should be confirmed
by classical histology in the case of colitis surveillance.
Strong recommendation, low quality evidence.
Lesions can be well delineated with high definition endoscopes and advanced imaging
techniques. In an RCT comparing dye-based HD-CE with HD-NBI, no dysplasia was found
in biopsies taken next to a visible lesion, even when the lesion was flat [81 ]. This means that if lesions can be well delineated, then resectability can be defined.
However the proportion of neoplasia per suspicious lesion detected during colitis
surveillance is in general rather low, at around 15 % [70 ]
[81 ]. This means that the majority of lesions found are regenerative changes and non-neoplastic.
Especially when such lesions are larger, resection may harbor unnecessary risks. The
question therefore arises whether optical diagnosis could be used to differentiate
neoplastic from non-neoplastic lesions.
Modified pit pattern classifications have been used in three dye-based CE studies
to differentiate between neoplastic and non-neoplastic lesions in long-standing IBD
[37 ]
[62 ]
[65 ], showing high sensitivity and specificity (93 % – 100 % and 88 % – 97 %, respectively).
Kawasaki et al. evaluated the efficacy of the Japanese magnifying colonoscopy classification
(Japan NBI Expert Team [JNET]) for UC-associated neoplasia [92 ]. Lesions of JNET types IIA, IIB, and III correlated with the histopathological findings
of low grade dysplasia (LGD), high grade dysplasia (HGD)/superficially submucosally
invasive cancer, and massively submucosally invasive (mSM) carcinoma, respectively.
Lesions of Kudo types III/IV, VI low irregularity, and VI high irregularity/VN, by
pit pattern classification, correlated with the histopathological findings of LGD/HGD,
HGD, and mSM carcinoma, respectively. One more recent study evaluated the endoscopic
features of HGD in 62 patients with UC [93 ]. HGD imaged with CE and magnifying endoscopy was frequently associated with a flat/superficial
elevated area and red color. However, the use of magnifying endoscopes is still not
widespread, and total procedure times were on average 9 – 11 minutes longer. Recently,
a Spanish multicenter trial showed that predictive factors for neoplasia for dye-based
CE are Kudo pit pattern III-V, sessile morphology, loss of innominate lines, and location
in the right colon [70 ].
Previous studies evaluating the role of nonmagnified NBI in differentiating neoplastic
and non-neoplastic lesions in patients with long-standing colitis suggested that a
tortuous pit pattern and a high vascular pattern intensity may help to distinguish
neoplastic and non-neoplastic lesions in longstanding IBD [94 ]
[95 ]. However, in two RCTS, the sensitivity and specificity of NBI in predicting histology
were insufficient [79 ]
[96 ]. A more recent multicenter interobserver study [97 ] showed median sensitivity, specificity, negative predictive value, and positive
predictive value for diagnosing neoplasia, based on the presence of pit pattern other
than I or II, of 77 %, 68 %, 88 %, and 46 %, respectively. Diagnostic accuracy was
significantly higher when a diagnosis was made with a high level of confidence (77 %
vs. 21 %, P < 0.001). The agreement for differentiation between non-neoplastic patterns (I, II)
and neoplastic patterns (IIIL, IIIS, IV, or V) was moderate and significantly better
for NBI in comparison with HD-CE (κ = 0.653 vs. 0.495, P < 0.001). Another multicenter RCT compared AFI with CE for dysplasia detection in
210 patients with long-standing UC [98 ]. Overall sensitivity for real-time prediction of dysplasia was 76.9 % for endoscopic
trimodal imaging (ETMI; namely, AFI, NBE, and WLE) and 81.6 % for CE. Overall negative
predictive values were 96.9 % for ETMI and 94.7 % for CE. A total of 205 lesions in
UC were analyzed with virtual CE (flexible spectral imaging color enhancement [FICE])
in another study, by Cassinotti et al. [99 ]. Sensitivity, specificity, positive and negative likelihood ratios with the Kudo
classification were 91 %, 76 %, 3.8, and 0.12, respectively. Recently Aladrén et al.
aimed to analyze results of a CE screening program in Spain and to assess the possibility
of identifying low risk dysplastic lesions by their endoscopic appearance, in order
to avoid histological analysis [100 ]. Correlation between dysplasia and Kudo pit pattern predictors of dysplasia (Kudo ≥ III)
was low while Kudo I and II lesions were correctly identified with a high negative
predictive value of 92 %, even by nonexperts. Recently a group of international experts
has developed and validated a new classification, the Frankfurt Advanced Chromoendoscopic
IBD LEsions system (FACILE), using images from all endoscopic platforms, that might
improve performance in both trainees and experienced operators. The four characteristics
that predicted neoplastic lesions were morphology of nonpolypoid/polypoid lesion,
irregular surface pattern, vessel architecture, and signs of inflammation within the
lesion, without using Kudo pit pattern [101 ].
Details of most of the abovementioned studies are available in Table 13 s .
Based on these studies we can say that to a certain extent optical diagnosis may help
to identify typically non-neoplastic lesions with type I or II pit pattern, but that
the overall diagnostic accuracy, even in expert hands, is insufficient. Resection
of small lesions < 10 mm with a neoplastic pit pattern is probably safe and may be
more practical for determining patient management in the case that neoplasia is found.
However in larger lesions, with sessile morphology or in the right colon [70 ], a biopsy should always be taken to confirm or rule out dysplasia.
Differentiation between neoplastic and non-neoplastic small colorectal polyps
Differentiation between neoplastic and non-neoplastic small colorectal polyps
2014 statement:
ESGE suggests that virtual chromoendoscopy (NBI, FICE, i-SCAN) and conventional [dye-based]
chromoendoscopy can be used, under strictly controlled conditions, for real-time optical
diagnosis of diminutive (≤ 5 mm) colorectal polyps to replace histopathological diagnosis.
The optical diagnosis has to be reported using validated scales, must be adequately
photodocumented, and can be performed only by experienced endoscopists who are adequately
trained and audited (weak recommendation, high quality evidence).
2019 statement:
ESGE suggests that virtual chromoendoscopy and dye-based chromoendoscopy can be used,
under strictly controlled conditions, for real-time optical diagnosis of diminutive
(≤ 5 mm) colorectal polyps to replace histopathological diagnosis. The optical diagnosis
should be reported using a validated scale, must be adequately photodocumented, and
can be performed only by experienced endoscopists who are adequately trained, as defined
in the ESGE curriculum, and audited.
Weak recommendation, high quality evidence.
The vast majority of polyps detected during colonoscopy are diminutive (1 – 5 mm)
or small (6 – 9 mm) in size. Diminutive polyps represent approximately 60 % of all
polyps detected and the risk of advanced pathology or cancer incurred by these lesions
is very low [102 ]
[103 ]
[104 ]. However, based on current management protocols, all removed polyps, including diminutive
polyps, are submitted for histological analysis. This is expensive and generates a
large burden of work for pathologists and histopathology departments. Instead of sending
diminutive polyps for histological evaluation, a real-time optical diagnosis by the
endoscopist would allow diminutive polyps to be discarded after resection, and non-neoplastic
polyps located in the rectum and sigmoid to be left in situ. Furthermore, optical
diagnosis could be used to determine the interval for the next surveillance colonoscopy.
The primary goal of this strategy is to reduce the number of polyps submitted for
histopathological evaluation, which may lead to cost savings.
The optical diagnosis strategy also raises several concerns. First, when diminutive
polyps are discarded, advanced histological features (high grade dysplasia, tubulovillous
or villous morphology) or invasive growth, i. e., a cancer, are not diagnosed as such.
This could lead to a setting of suboptimal treatment and/or inappropriate surveillance
intervals. However, risk estimates for advanced pathology within diminutive polyps
are low, ranging from 0.1 % to 12 %, with most estimates at the lower end of this
range [105 ]
[106 ]
[107 ]
[108 ]
[109 ]
[110 ]
[111 ]
[112 ]
[113 ]
[114 ]
[115 ]
[116 ]
[117 ]
[118 ]
[119 ]
[120 ]
[121 ]
[122 ]
[123 ]
[124 ]
[125 ]
[126 ]
[127 ]
[128 ]
[129 ]
[130 ]
[131 ]
[132 ]
[133 ]
[134 ] (Table 14 s ). The rate of cancer in diminutive polyps is even lower, although not completely
negligible, ranging from 0 % to 0.6 %, with most estimates again at the lower end
of the range. To further reduce the risk of missing cancer, it is recommended that
an optical diagnosis should be avoided in suspicious lesions (e. g. depressed lesions,
Paris classification 0-IIc) [135 ]. The question of whether undiagnosed advanced histological features within diminutive
polyps would lead to inappropriate surveillance recommendations was recently addressed
in a large study [103 ]. In this study, data of 12 cohorts (5 FIT cohorts and 7 colonoscopy screening cohorts)
were combined, resulting in a total cohort of 64 344 individuals with 51 510 diminutive
polyps. Advanced histological features were observed in 5.6 % and cancer in 0.07 %
of all diminutive polyps. The risk of finding metachronous advanced neoplasia did
not significantly differ between patients with 1 or 2 nonadvanced diminutive or small
adenomas (low risk patients) compared with patients with diminutive polyps with advanced
histological features detected at baseline colonoscopy. This indicates that diminutive
polyps with advanced histological features do not increase the risk for metachronous
advanced neoplasia and therefore seem not to interfere with a correct surveillance
recommendation.
A second concern is that an incorrect optical diagnosis could result in a patient
being incorrectly considered at low risk for metachronous advanced neoplasia and/or
that neoplastic lesions in the rectosigmoid are left in situ. For this reason, the
American Society for Gastrointestinal Endoscopy (ASGE) published the Preservation
and Incorporation of Valuable Endoscopic Innovation (PIVI) document in which they
attempted to set standards against which a technology should be assessed in order
to be deemed suitable for use. A policy of resect and discard should have ≥ 90 % agreement
in assignment of post-polypectomy surveillance intervals when compared with decisions
based on pathology assessment, and a policy of leaving suspected non-neoplastic polyps
in place should have a ≥ 90 % negative predictive value when used with high confidence
[136 ]. A meta-analysis published in 2015 [137 ], including 20 NBI studies [138 ]
[139 ]
[140 ]
[141 ]
[142 ]
[143 ]
[144 ]
[145 ]
[146 ]
[147 ]
[148 ]
[149 ]
[150 ]
[151 ]
[152 ]
[153 ]
[154 ]
[155 ]
[156 ]
[157 ], 7 I-SCAN studies [155 ]
[158 ]
[159 ]
[160 ]
[161 ]
[162 ]
[163 ] and 8 FICE studies [164 ]
[165 ]
[166 ]
[167 ]
[168 ]
[169 ]
[170 ]
[171 ], all in vivo and published between 2008 and 2014, showed that the pooled NPV of
NBI for adenomatous polyp histology was 91 % (95 %CI 88 % – 94 %). The agreement in
assignment of post-polypectomy surveillance intervals with NBI was 89 % (95 %CI 85 % – 93 %).
Importantly, subgroup analysis indicated that the pooled NPV and the surveillance
agreement was only greater than 90 % for academic medical centers, for experts, and
when the optical assessment was made with high confidence. Comparable results were
observed for I-SCAN. For FICE the pooled NPV in this meta-analysis was 80 % (95 %CI
76 % – 85 %). Dye-based CE shows similar accuracy in differentiating between neoplastic
and non-neoplastic polyps, but because of inconvenience and costs associated with
the use of dyes it is unlikely that this technique will be adopted in routine clinical
practice [164 ]
[166 ]. From 2015 onwards, real-time differentiation studies, performed in academic centers
as well as in community hospitals, have shown conflicting results in achieving the
above mentioned PIVI thresholds [125 ]
[172 ]
[173 ]
[174 ]
[175 ]
[176 ]
[177 ]
[178 ]
[179 ]. This variability in performance may be explained by a lack of rigorous training
and/or performance measurement. However, in those studies in which the endoscopists
were adequately trained prior to the study, PIVI thresholds were also not always met
[125 ]
[174 ]
[179 ]. In conclusion, performance levels of endoscopists in correctly predicting histology
of diminutive polyps remain highly variable, underlining the necessity of a training,
auditing, and performance monitoring system when an optical diagnosis strategy is
implemented. The possible effect on optical diagnosis of the use of artificial intelligence
(AI) in the future is also unclear at this stage (see section on Role of artificial intelligence ). Details of the abovementioned studies are available in Table 15 s .
During real-time optical diagnosis, validated optical diagnostic scales, such as the
widely used NBI International Colorectal Endoscopic (NICE) classification or the Workgroup
serrAted polypS and Polyposis (WASP) classification (which also includes sessile serrated
lesions [SSLs]) should be used to improve diagnostic accuracy [145 ]
[174 ]
[180 ]. No universal training system for differentiation between neoplastic and non-neoplastic
colorectal polyps has been established yet. Several teaching modules, mostly computer-based,
have been studied and some of them are showing promising results with respect to improving
interobserver agreement; however in a substantial number of studies the interobserver
agreement was still moderate after training [180 ]
[181 ]
[182 ]
[183 ]
[184 ]
[185 ]
[186 ]
[187 ]
[188 ] (Table 16 s ).
There are currently no data to suggest what kind of documentation is needed for implementation
of an optical diagnosis strategy. As in this situation an endoscopic picture, rather
than a histology slide, becomes the record of a diminutive polyp, it seems logical
that those images are stored. At least one or two images must be stored as evidence
of adenoma detection and also for review of the optical diagnosis [136 ]. However, this strategy poses significant challenges at present, especially with
regard to logistics and the available disk space on servers in endoscopy units.
Implementation of an optical diagnosis strategy would be cost-effective, with good
evidence from large modeling studies to support this [157 ]
[170 ]
[189 ]
[190 ]
[191 ]
[192 ]
[193 ]. However, concerns associated with the data used for model analysis include: (i)
the different CRC screening programs used in these models may not be simply extrapolated
to the various screening programs in use in Europe; (ii) the assumptions are derived
from studies that have mainly been performed by experts; and (iii) the costs for implementation
of the resect-and-discard policy (training for and photodocumentation of real-time
diagnosis) are not included. It is therefore unclear whether the results of these
modeling studies can be reproduced in real-life daily practice, and this should be
further investigated in a real-life (multicenter) setting.
Role of advanced imaging in treatment of colorectal neoplasia
Role of advanced imaging in treatment of colorectal neoplasia
Prediction of deep submucosal invasion
2014 statement:
ESGE suggests the use of conventional or virtual (NBI) magnified chromoendoscopy to
predict the risk of invasive cancer and deep submucosal invasion in lesions such as
those with a depressed component (0-IIc according to the Paris classification) or
nongranular or mixed-type laterally spreading tumors (weak recommendation, moderate
quality evidence).
2019 statement:
ESGE recommends the use of high definition white light endoscopy in combination with
(virtual) chromoendoscopy to predict the presence and depth of any submucosal invasion
in nonpedunculated colorectal polyps prior to any treatment.
Strong recommendation, moderate quality evidence.
When endoscopic resection is considered for colonic lesions, it is important to assess
the lesion accurately and attempt to predict the presence and depth of submucosal
invasion, as this will aid in determining the correct treatment strategy (piecemeal
endoscopic resection, e. g. endoscopic mucosal resection [EMR]; en bloc endoscopic
resection, e. g. endoscopic submucosal dissection [ESD] or use of full thickness resection
device [FTRD], or surgery). White light characterization and virtual and dye-based
CE with and without magnification help to predict the presence and depth of submucosal
invasion.
Morphology, size, location, and recognition of gross morphological features are the
first steps in the characterization of colonic lesions with WLE, and may help to raise
suspicion of malignancy. Submucosal invasion has been shown elsewhere to be more frequent
in certain morphologies (laterally spreading tumor of nongranular type [LST-NG] pseudodepressed
lesions, and also sessile polyps), increased size, and rectal location [194 ]
[195 ]. A large prospective study of colonic lesions showed that the risk of ‘covert’ submucosal
invasion was predicted by rectosigmoid location (odds ratio 1.87, P = 0.01), combined Paris classification, surface morphology (odds ratios, 3.96 – 22.5),
and increasing size (odds ratio 1.16 /10 mm, P = 0.012) [196 ]. In particular, rectosigmoid Paris 0-Is and 0-IIa + Is nongranular lesions had a
high risk of submucosal invasion whereas proximally located Paris 0-Is or 0-IIa granular
lesions had a very low risk. In addition, the nonlifting sign, chicken skin sign,
edge retraction, depressed areas, folds convergence, induration, ulceration, polyp
over polyp, redness, tumor fullness, and spontaneous bleeding have been reported to
be associated with submucosal invasion, and also in lesions < 10 mm, but none of them
was definitive [194 ]
[197 ]. A systematic review and meta-analysis showed that sensitivities of these features
for predicting deep submucosal invasion ranged from 18 % to 68 % and specificities
from 80 % to 98 %. [198 ] The recognition of demarcated areas (clearly visualized region between two morphological
areas of a lesion, e. g., a depression, large nodule, or reddened area) is also a
key point in identifying zones that deserve close observation, because they are associated
with an increased risk of submucosal invasion [199 ].
On closer inspection of the target colonic lesion, detection and characterization
of a demarcated area where a regular neoplastic pit/vascular pattern (e. g. Kudo IV,
NICE II, Sano II) becomes disordered (e. g. Kudo V, NICE III, Sano III), often associated
with a visible depression (Paris classification 0-IIa + c) due to a fibrotic reaction
in the submucosa, is a specific marker of submucosal invasion within colonic lesions.
There are only three prospective studies evaluating in vivo CE without magnification.
The OPTICAL study [200 ] prospectively assessed 343 large nonpedunculated colorectal polyps with NBI without
magnification, using the Hiroshima classification. A total of 47 cancers were identified
(36 T1 and 11 ≥ T2), of which only 11 contained superficial sm1 invasion (23.4 % of
all malignant polyps). Sensitivity and specificity for optical diagnosis of T1 CRC
were 78.7 % (95 %CI 64.3 % – 89.3 %) and 94.2 % (95 %CI 90.9 % – 96.6 %), respectively;
corresponding values for optical diagnosis of endoscopically unresectable lesions
(i. e., ≥ T1 CRC with deep invasion) were 63.3 % (95 %CI 43.9 % – 80.1 %) and 99.0 %
(95 %CI 97.1 % – 100.0 %), respectively. Obvious advanced cancers were excluded, but
11 out of 47 were still advanced cancers (7 T2 and 4 T3), which might have increased
the sensitivity.
In a Spanish multicenter prospective study including 2123 lesions > 10 mm using NBI
and without magnification, the NICE classification system identified lesions with deep
invasion with sensitivity 58.4 % (95 %CI 47.5 % – 68.8 %) and specificity 96.4 % (95 %CI
95.5 % – 97.2 %) [194 ]. In addition, a conditional inference tree that included all variables found that
the NICE classification was the most accurate for identification of lesions with deep
invasion (P < 0.001). However, pedunculated morphology (P < 0.007), ulceration (P = 0.026), depressed areas (P < 0 .001), or nodular-mixed type (P < 0.001) also affected accuracy of identification ([Fig. 1 ]). Therefore, virtual CE without magnification is useful for predicting deep submucosal
invasion when a nonpedunculated NICE type 3 polyp is ulcerated and is useful to rule
it out when a NICE type 1 or 2 lesion has no depressed area nor nodules. Results were
comparable for identifying lesions that were endoscopically not resectable for oncological
reasons (with any risk factor for lymph node metastasis). This is consistent with
previous Japanese studies showing a higher prevalence of deep submucosal invasion
in demarcated areas [199 ]. Therefore, magnification is especially needed in nonulcerated NICE type 3 lesions
or when a demarcated area (nodule, redness, or depression) is present in a NICE type
1 or 2 lesion.
Fig. 1 Risk of submucosal invasion based on the Narrow-band imaging International Colorectal
Endoscopic (NICE) classification and polyp morphology to determine treatment options
[194 ].
There is only one study assessing the Kudo pit pattern for predicting submucosal invasion
without magnification [196 ]. Sensitivity and specificity of the Kudo pit pattern type V were 40.4 % (95 %CI
33.3 % – 47.8 %) and 97.5 % (95 %CI 96.7 % – 98.1 %) in 2106 laterally spreading lesions
> 20 mm.
In Japan, magnified NBI CE has been shown to have a sensitivity of 77 % (95 %CI 68 % – 84 %)
and a specificity of 98 % (95 %CI 95 % – 99 %) in 13 studies using different classification
systems [198 ]. Recently, type 3 JNET classification has shown a sensitivity of 55.4 % (95 %CI
48.7 % – 62.1 %) and a specificity of 99.8 % (95 %CI 99.6 % – 100.0 %) in retrospective
assessment of 2933 images [201 ]. Studies with similar results showed that JNET type 2B included a wide variety of
colorectal tumors ranging from low grade dysplasia to deep submucosal lesions and
therefore the sensitivity of JNET type 3 is low [202 ]
[203 ]
[204 ]
[205 ]
[206 ]
[207 ]. The authors suggest that direct observation of the Kudo pit pattern with crystal
violet should be performed in JNET 2B lesions.
The abovementioned systematic review and meta-analysis showed a sensitivity of 81 %
(95 %CI 75 % – 87 %) and a specificity of 95 % (95 %CI 89 % – 97 %) for magnified
CE in 17 studies [198 ]. All the studies were performed in Asian countries, mainly Japan, and with crystal
violet. A retrospective study conducted in Brazil by a single experienced endoscopist
included 123 lesions with suspicion of submucosal invasion raised by another endoscopist.
Magnifying CE with pit pattern classification had 73.3 % sensitivity and 100 % specificity
[208 ].
Details of the most important of the abovementioned studies are available in Table 17 s .
In summary: WLE may raise suspicion for submucosal invasion; virtual CE without magnification
is useful to rule out the presence of deep submucosal invasion when no demarcated
area is present; and magnifying CE may allow the differentiation between deep and
superficial submucosal invasion in highly suspicious lesions, such as those containing
demarcated areas. Based on the recent evidence, a 4-step strategy incorporating the
different roles of WLE, nonmagnifying virtual CE, magnifying virtual CE, and magnifying
dye-based CE in predicting submucosal invasion has been proposed, but it should first
be validated [209 ]. In the near future, it seems likely that AI, directed to a demarcated area by a
human observer, will significantly improve both sensitivity and specificity (see section
on Role of artificial intelligence ).
Defining the borders of colorectal lesions
2014 and 2019 statement:
ESGE recommends the use of virtual or conventional [dye-based] chromoendoscopy to
define the margins of large nonpolypoid or otherwise indistinct lesions before or
during endoscopic resection.
Strong recommendation, very low quality evidence.
No new evidence has become available regarding this statement. Because of the better
contrast, the entire extent of the lesion can be better appreciated with additional
imaging techniques to safeguard a complete resection of a lesion. Especially in IBD-related
neoplasia, demarcation of a lesion can be challenging and is facilitated by CE.
Follow-up after endoscopic resection of lesions
2014 statement:
ESGE recommends the use of virtual or conventional chromoendoscopy in addition to
white light endoscopy for the detection of residual neoplasia at a piecemeal polypectomy
scar site (strong recommendation, low quality evidence).
2019 statements:
ESGE recommends the use of virtual or dye-based chromoendoscopy in addition to white-light
endoscopy for the detection of residual neoplasia at a piecemeal polypectomy scar
site.
Strong recommendation, moderate quality evidence.
ESGE suggests that routine biopsy of post-polypectomy scars can be abandoned providing
that a standardized imaging protocol with virtual chromoendoscopy is used by a sufficiently
trained endoscopist.
Weak recommendation, moderate quality evidence.
Endoscopic piecemeal polypectomy has emerged as a safe and effective method of removing
large sessile or nonpolypoid colorectal lesions. However, because of a relatively
high rate of adenoma recurrence, estimated at 15 % – 30 %, [210 ]
[211 ], it is recommended to perform a surveillance colonoscopy at 4 – 6 months after endoscopic
resection [212 ]
[213 ].
It has been shown that using HD-WLE alone allows the identification of 69 % to 83 %
of recurrences revealed by performing targeted and random biopsies [141 ]
[214 ]. Recent studies have provided new evidence for the efficacy of advanced endoscopic
imaging in the detection of post-polypectomy/post-EMR scars and residual/recurrent
colorectal neoplasia. A prospective single-center study, which analyzed 183 scars
after a median of 3.9 months from the endoscopic polypectomy, showed a significantly
higher sensitivity for endoscopic residual neoplasia detection for a combination of
HD-WLE and NBI compared with HD-WLE alone (93.3 % vs. 66.7 %). The NPV for the combination
of HD-WLE and NBI was 98.6 % (95 %CI 95.1 % – 99.8 %) [215 ]. Another prospective multicenter study, which evaluated 255 colorectal scars after
a median of 7 months following a colorectal piecemeal EMR, showed a NPV of 100 % (95 %CI
98 % – 100 %) and sensitivity of 100 % (95 %CI 93 % – 100 %) for NBI with near-focus
imaging [216 ]. However, slightly lower values were observed in a study of 112 scars, which showed
that the accuracy of NBI for the detection of residual neoplasia at the resection
site was 86.8 %, compared to 81.6 % for WLE and (P = 0.15) [217 ]. This study has however several limitations, including the single operator, high
recurrence rates, and non-blinded pathologist. Another study, comparing the combination
of HD-WLE and virtual or dye-based CE against histological verification in recurrence
assessment, revealed biopsy evidence of residual/recurrent lesions in 16 of 228 macroscopically
inconspicuous polypectomy scars (7 %) [218 ]. This study had, however, very high rates of recurrence (31.7 %) and used argon
plasma coagulation to complete or ascertain completeness of resection in 50 % of patients.
The high sensitivity and NPV (93 % – 100 %) of HD-WLE combined with virtual CE in
identifying residual and/or recurrent colorectal neoplasia justifies abandoning biopsy
of macroscopically normal EMR or piecemeal polypectomy scars.
Role of artificial intelligence in detection and characterization of colorectal polyps
Role of artificial intelligence in detection and characterization of colorectal polyps
2019 statement:
ESGE suggests the possible incorporation of computer-aided diagnosis (detection and
characterization of lesions) into colonoscopy, if acceptable and reproducible accuracy
for colorectal neoplasia is demonstrated in high quality multicenter in vivo clinical
studies. Possible significant risks with implementation, specifically endoscopist
deskilling and over-reliance on artificial intelligence (AI), unrepresentative training
datasets, and hacking, need be considered.[* ]
Weak recommendation, low quality evidence.
Computer-aided diagnosis in medical imaging has been revolutionized by the advent
of artificial intelligence (AI) “deep learning” based on neural networks that simulate
to some degree the workings of the human brain. It seems likely that such systems
will have a major place in clinical practice in the future, with more than 20 systems,
in particular in radiology and pathology, having received regulatory approval [219 ]. Video endoscopy provides a further opportunity for the application of AI systems
to support and enhance clinical practice and endoscopist performance. However despite
the potential benefits, there are also risks associated with the clinical adoption
of AI.
Endoscopist – AI interaction
AI can support clinicians in endoscopy in a number of ways. We consider below two
major scenarios for colonoscopy, looking at lesion detection and lesion characterization;
however the endoscopist can interact with computer-aided diagnosis systems in different
ways. This interaction can be active, where we find a polyp and ask the AI system
to confirm our diagnosis as a “second reader,” or passive, where AI is running continuously
in the background, for example for polyp detection, providing a “concurrent read”
alongside the endoscopist. There may be situations where AI acts completely autonomously
to make a decision without any endoscopist input, and it is unknown how the AI output
is determined [220 ] ([Fig. 2 ]). An expert group set up by the European Commission has recently proposed that algorithms’
“black boxes” should be deconvoluted before they can be used for patient care [221 ]. The levels of endoscopist–AI interaction have similarities to self-driving cars.
For example, humans can monitor the environment but can be aided by automated speed
control and braking; self-driving may also allow the AI system to monitor the environment,
with limited human input, or even to be fully autonomous. However it seems unlikely
that fully autonomous “black box” AI will feature widely in medicine [219 ].
Fig. 2 Different possibilities for endoscopist – artificial intelligence (AI) interaction
[220 ].
Diagnostic performance of AI in colonic polyp detection
Substantial variation exists among endoscopists in terms of polyp detection and effectiveness
in preventing CRC with colonoscopy [4 ]
[11 ]. This variability has been attributed to many factors, but a significant cause seems
to be that potentially detectable polyps are missed [179 ]
[222 ]
[223 ]
[224 ]
[225 ]. The limitations of human visual perception and other human factors, such as fatigue,
distraction, and level of alertness during examination, increase such recognition
errors, and their mitigation may be the key to improving polyp detection and further
reduction in CRC mortality. Computer-aided detection (CAD) could address these limitations
[226 ] Recent advances in AI, deep learning, and computer vision have shown potential for
assisting polyp detection during colonoscopy.
Preliminary studies of deep learning-based CAD systems have reported sensitivities
from 70 % to 90 % and specificities from 60 % to 90 % for detecting polyps [227 ]
[228 ]
[229 ]
[230 ]
[231 ]
[232 ]. There are insufficient data to establish whether there is effective detection of
sessile serrated or relatively flat and depressed lesions (Paris 0-II).
Although CAD could be useful for polyp detection in clinical practice [228 ], some limitations remain. A major drawback of current CAD systems is the relatively
large number of false-positive detections, which could adversely affect the application
of CAD in clinical practice. A large rate of false positives is likely to confound
the endoscopist’s task of image interpretation and reduce the efficiency of colonoscopy.
In addition, endoscopists may lose confidence in CAD as a useful tool. The speed of
CAD for image analysis and output presentation may also be an issue. Fast processing
times are required for image analysis and on-screen labeling, so that the endoscopist
is alerted in real time to the presence of a polyp.
Details of the more important of the abovementioned studies are available in Table 18 s .
Diagnostic performance of AI in polyp characterization
AI for characterization of colorectal lesions might have potential advantages in:
(i) improving the endoscopist’s learning phase; (ii) predicting neoplastic and non-neoplastic
tissue (e. g. to support a resect-and-discard strategy); and (iii) guiding endoscopic
therapy (e. g. by prediction of submucosal infiltration). So far, no randomized controlled
trials have assessed this rapidly emerging technology.
Specifically, no data are yet available on the effect of AI on the learning curve
of endoscopists. Regarding prediction of adenomatous and hyperplastic polyp histology,
recent data have highlighted that AI based on deep learning models can accurately
predict polyp histology with sensitivities and NPVs exceeding 90 % [233 ]
[234 ]. Similar results have also been shown for AI based on traditional machine learning
[235 ]
[236 ]. AI based on machine learning has also been evaluated for predicting the need for
additional surgery after endoscopic resection of T1 colorectal cancer; it was found
that it could significantly reduce unnecessary additional surgery [237 ]. Finally AI based on a deep learning model has been used to assist in diagnosis
of submucosal CRC showing an accuracy of 81 % [238 ]
[239 ].
Beyond colonic polyps there may be a role for AI in scoring inflammation in IBD, with
preliminary data supporting distinction between Mayo 0 – 1 levels of inflammation
and higher Mayo 2 – 3 levels (area under receiver operating characteristic [AUROC]
0.98) [240 ]. In addition, AI may potentially help in automatically registering quality indicators
for colonoscopy (withdrawal time, cecal intubation, bowel preparation).
Details of the most important of the abovementioned studies are available in Table 19 s .
The role of add-on standalone systems versus AI that is integrated into commercially
available endoscopy systems remains unclear. However either approach seems to have
significant potential to enhance practice and facilitate optical diagnosis or resect-and-discard
strategies [220 ].
Risks of AI in clinical practice
Whilst many previous publications have exclusively mentioned the strengths and advantages
of the use of AI in medicine, there are potential drawbacks to using AI in colonoscopy.
In seven prospective studies on AI in colonoscopy [231 ]
[235 ]
[241 ]
[242 ]
[243 ]
[244 ]
[245 ], none addressed the downsides of AI as one of the main outcome measures, except
for assessment of the time required for using AI; results varied from an increase
of 35 – 47 seconds per polyp assessed with AI [235 ] to no additional withdrawal time [231 ].
Outside the field of colonoscopy, recent review articles have warned of unintended
consequences that possibly arise from the use of AI in health care [219 ]
[246 ], namely over-reliance on AI, deskilling, biased datasets for machine learning, and
the risk of hacking, all of which seem to be applicable to AI in colonoscopy. In the
short term, endoscopists’ diagnoses can be affected by incorrect AI predictions. Some
previous studies on decision support systems for mammography [247 ] and electrocardiography [248 ] demonstrated this negative effect in practice. According to these studies, experienced
radiologists and residents, respectively, tended to adopt wrong decisions when they
were given an incorrect AI prediction.
The problem of biased data for machine learning should be addressed when AI is widely
implemented into colonoscopy practice. Currently, no colonoscopy AI systems have used
learning data from different countries, although the status of colonic mucosa, morphologic
pattern of polyps, and quality of bowel preparation may differ significantly among
countries. Similarly, differences in endoscopic technology among regions of the world
(e. g. between the Olympus Lucera Elite and the Exera III) or between endoscopy manufacturers
may significantly affect AI performance if the training sets had not included a full
range of data. In this regard, international validation should be required before
global use of the developed AI. Small, unrepresentative data sets can lead to unintended
outcomes, as happened with the IBM Watson for Oncology software [249 ]. Wide adoption of such data sets in healthcare systems can have far-reaching negative
consequences.
The risk of hacking is also an inevitable concern. Deliberate hacking of a computer
with AI installed could lead to large-scale harm to patients. For example, use of
AI which provides wrong histological predictions because of malware could lead to
serious consequences, such as neoplastic polyps being left in situ.
A more specific concern is that endoscopists using CAD are obliged to pay attention
to the CAD output at the same time as making their own assessment. Thus the CAD output,
especially if it was inaccurate, might distract the endoscopist, leading to missing
or mischaracterization of polyps [250 ]. On the other hand, no serious adverse event such as perforation has been reported
that was due to such distraction, according to two prospective studies [235 ]
[244 ]. Detection algorithms may produce many false positives which require careful mucosal
inspection; this can increase the time and mental load when performing colonoscopy,
leading to a lessening of concentration.
There is also an assumption that effects of CAD (e. g. improved adenoma detection)
will automatically lead to a reduction in missed CRC, because of the association between
ADR and post-colonoscopy CRC [11 ]. However, changes in ADR produced by AI are in effect improvements in detection
of polyps within the visual field, and AI cannot detect polyps in non-inspected mucosa.
Therefore if improved ADR is in fact a surrogate measure of enhanced mucosal visualization,
with better re-inspection of flexures, suctioning and pressing down mucosal folds,
factors not changed by application of AI, the link between enhancement of ADR and
fewer missed cancers may not hold true.
Although the evidence on the risks of AI for colonoscopy is limited, nevertheless
various risks of AI such as prolonged procedure time, over-reliance on AI, and distraction
caused by AI, should be considered, and quality assurance measures instituted [251 ]
[252 ]. Future prospective studies should assess the impact of these downsides of AI in
addition to its efficacy.
Disclaimers
The legal disclaimer for ESGE guidelines [10 ] applies to this Guideline.
The views expressed by J.E. East and M. Iacucci are those of these authors and not
necessarily those of the National Health Service (England and Wales), the National
Institute for Health Research or the UK Department of Health.
Funding
J.E. East was funded by the National Institute for Health Research (NIHR) Oxford Biomedical
Research Centre, and M. Iacucci receives funding support from the National Institute
for Health Research (NIHR) Birmingham Biomedical Research Centre. R. Bisschops was
funded by a grant of the Research Foundation Flanders (FWO). Y. Mori was funded by
the Japan Society for the Promotion of Science. E. Coron, H. Neumann, and Y. Hazewinkel
received no funding.
Bisschops R, East JE, Hassan C et al. Advanced imaging for detection and differentiation
of colorectal neoplasia: European Society of Gastrointestinal Endoscopy (ESGE) Guideline
– Update 2019 Endoscopy, DOI 10.1055/a-1031-7657 In the above-mentioned article, the author Serguei Mouzyka and his institution have
been included. This was corrected in the online version on December 18, 2019