The purpose of this editorial is not to detail all the technical aspects of advanced
diagnostic endoscopy for colorectal tumors but to focus on some of the issues raised
by diagnostic endoscopy, with which impressive progress has been made in recent years,
at least regarding its use for therapeutic colonoscopy.
Challenges regarding detection
Electronic chromoendoscopy
The most recent literature shows that techniques using electronic chromoendoscopy
(EC), particularly from the latest generation, improve adenoma detection rates by
enhancing the contrast between the lesion and the surrounding mucosa [1]
[2]
[3]
[4]. But these techniques, if they are simple to implement, are not widely used, especially
in Europe, and there are no recommendations that impose these techniques first-line
for those at average-risk of colorectal cancer (CRC). For example [5], The European Society of Gastrointestinal Endoscopy (ESGE) suggests that “virtual
chromoendoscopy can be used to increase the endoscopist’s adenoma detection rate (ADR).
Their routine use must be balanced against costs….”. Why not “has to be used” instead
of “can be used”? Possible challenges preventing its use, such as lack of adequate
training, user-experience, and additional withdrawal time, should be addressed. More
endoscopist-friendly EC alternatives may be game-changing. By mixing the blue component
with red and green components, new lights mimicking the bright appearence of the usual
white-light have been developed, while preserving the same efficacy as traditional
EC. The challenge is whether and when these techniques will be recommended as a first
line and will replace white light imaging (WLI) for detection in the colon.
Artificial intelligence
The most recent publications show that artificial intelligence (AI) techniques improve
the adenoma detection rate (ADR). All published trials comparing AI to WLI reported
an effect on ADR in the AI groups and a recent meta-analysis [6] concluded that ADR with AI is much higher than ADR with WLI. The first challenge
concerning AI is to continue to evaluate AI. We need more data on the rate of false-positive
results, more data for non-polypoid lesions and on cost-effectiveness due to effect
of withdrawal time, additional costs for polypectomy and for surveillance. In addition,
the issue of de-skilling, especially for trainees, must be considered. Will AI users
be able to achieve adequate standards when AI is not available?
The second challenge is whether the promising data apply to all AI algorithms, which
are numerous. In other words, will there be an evaluation and labeling of the different
AI techniques? A PIVI (Preservation and Incorporation of Valuable endoscopic Innovations)
statement, such as the one recommended by the American Society of Gastrointestinal
Endoscopy, could be useful for AI “to avoid widespread use before clinical studies
documenting effectiveness have been performed.” Technical comparisons across different
systems with a rigorous ground-truth may be expected to prevent excessive variability
in AI-assisted detection. This is critical when implementing AI in population-based
programs in which a consistent standard is especially required.
The third challenge is to better evaluate whether AI can be combined adequately with
electronic chromoendoscopy techniques. That is not certain, but possible. Will the
gold standard be a combination of AI and electronic chromoendoscopy or will the latter
be made obsolete by AI?
The fourth challenge is medico-legal and consists of determining precisely what the
level of responsibility is related to AI. So far, AI has been presented as assisting
the operator and the operator as responsible for detection, i. e., AI is a low-risk
medical device. But a day will come when, in the event of a medico-legal case brought
after a lesion has been missed that the respective responsibilities of the operator
and the AI will be clarified. This also addresses the need to document AI output at
each examination. Should all the AI triggers be saved, only those relevant from a
clinical perspective, or neither of them?
The fifth challenge will be commercial. How can any hospital equip itself with so
many different systems, particularly if the systems evolve frequently? Maybe the best
business model will be to lease different systems that can be upgraded regularly.
Devices to explore the colonic mucosa behind the folds
Currently, distal attachment devices, especially Endocuff, have been shown to be effective
in detecting lesions behind the folds in areas that are difficult to explore, and
in increasing ADRs [6]
[7]. As for electronic chromoendoscopy, the first challenge for scientific societies
such as ESGE is to decide if the use of these devices (and which ones?) is only suggested
(“can be used”) or strictly recommended (“has to be used”). Possible barriers may
prevent add-on implementation. These devices usually represent an additional cost
per colonoscopy that is currently not reimbursed by several health systems. Despite
their overall efficacy, there is variability in the intrinsic mechanism of adds-on,
i. e. cap vs. cuffs, generating uncertainty about the best way to flatten the fold.
The educational context seems relevant. The use of the cap is reported in a substantial
proportion of screening colonoscopies in Japanese series, while this approach is missing
from Western-based series. The second challenge does once again concern AI: AI will
help to determine whether the exploration of the colonic mucosa has been complete
or not and whether unexplored areas persist. The challenge will then be to combine
attachment devices and AI.
Challenges regarding characterization (optical diagnosis)
Limiting the number of classifications
Indeed, Western endoscopists tend to analyze fewer details than Japanese endoscopists
and to use fewer classifications. This is primarily related to the reluctance of Western
endoscopists to use optical magnification, which in contrast is on the standard for
Japanese endoscopists, and Japanese-driven classifications. Therefore, in order for
a Western endoscopist to adhere to the classifications, the number of them should
be limited. There has already been an attempt to limit the number of classifications
by establishing international classifications, such as the NBI International colorectal
endoscopic classification (NICE) or Japan NBI Expert Team (JNET). In fact, the problem
is that the fewer classifications there are, the more complex they should be to cover
all situations. Finding a balance between the number of classifications and their
respective complexity is a difficult challenge. An additional chalenge is AI development
in this arena. Will classifications adhere to the histologically-mimicking AI-output,
such as neoplastic vs. non-neoplastic or adenoma vs. non-adenoma?
Comparing different endoscopes from different companies
Indeed, it is necessary to avoid having a classification for each endoscope company
(NICE, BASIC, SIMPLE, …). Conversely, after demonstration, scientific societies should
be able to indicate that a particular technology is better and more adapted than another
technology. Currently, there is a certain policy of juxtaposition of technologies
and non-aggression between the different endoscope companies. It would be preferable
to move towards a policy of homogenization but to maintain a certain degree of competition
[7]. A difficult challenge politically speaking!
Distinguishing between hyperplastic polyps and adenomas (NICE type 1 vs type 2)
Distinguishing these two types of polyps is the basis of the discard policy (to resect
but not perform histopathology). This policy has a rationale: it reduces costs and
is feasible and safe. But the first part of the challenge is whether it concerns diminutive
polyps ( < 5 mm) or small polyps (6–10 mm). The second part is whether the challenge
should be restricted to the “leave-in-situ” strategy for rectosigmoid diminutive hyperplastic
polyps or extended to the “resect-and-discard” for adenomatous or more proximal yeperplastic
lesions. The former is easier but clinically less relevant, if at all, while the second
may pose additional challenges, such as reimbursement for optical diagnosis and legal
and financial issues, as histology is mandatory for reimbursement in some health systems,
and general reluctance by patient population. Part of the challenge is for scientific
societies like the ESGE to promote or not promote this policy. This is not the case
at the moment because routinely the results are not as good as with the experts and
because there is the problem of sessile serrated lesions (SSLs) that behave like adenomatous
lesions but look like hyperplastic lesions. Despite various attempts, incorporating
SSLs into current classifications remains difficult. To date, ESGE states that it
is possible to discard polyps, as long as photodocumentation and certified training
are organized, but it is not recommended. With regard to photodocumentation, it means
having defined quality criteria for pictures and having defined quality criteria for
the preservation and storage of documents (how long? and on what medium?). For certified
training, this means having a certification program in place. This is not simple and
is not yet the case for ESGE. By offering very high accuracy values, AI may further
push these challenges. Will health systems and patients be happy to rely on an AI-driven
diagnosis? Will an endoscopist with no skills in optical diagnosis be ready to passively
accept AI output? Will scientific societies like ESGE be ready to rely on AI-driven
certified diagnosis?
Improving detection of a carcinomatous component within an adenoma and to better predict
the level of parietal invasion (e. g., type 2A vs type 2B vs type 3 of the JNET classification)
With the development of endoscopic submucosal invasion, the question arises in all
countries, including Western countries. But unlike NICE 1 vs NICE 2, the challenge
is to analyze a continuous variable: from regular to irregular, there is no clear
limit. Even in Japan, the study results are not perfect. The level of expertise is
less in Europe, so the challenge is to set up quality training with evaluation. This
may have annoying consequences such as excessive referral to surgery for benign lesions.
Moreover, the situation is complicated by the fact that Europeans do not use crystal
violet with optical magnification for pit pattern evaluation, which is very useful
and widely used in Japan. On one hand, Europeans tend to make little use of dyes although
they have in their operating rooms Lugol, indigo carmine and acetic acid and they
will not accept so easily a fourth dye. On the other hand, crystal violet is considered
potentially toxic and carcinogenic in Europe. Endocytoscopy also appears to be a very
interesting tool to better diagnose a carcinomatous area within an adenoma [8] but the routine use of endocystoscopy will be limited by three factors : (1) the
need to apply methylene blue, another potentially carcinogenic dye, prior endocystoscopy;
(2) the duration of a complete analysis of the polyp surface ; and (3) the cost of
the system. A cost-effectiveness analysis is mandatory before any routine use of endocystoscopy.
AI linked to image-enhanced endoscopy is likely to be the best option to solve challenges
3 and 4 in Europe in the near future.
Conclusions
To conclude, diagnostic colonoscopy has made impressive progress in recent years,
thanks to implementation of quality criteria and to the availability of new technology.
But of course, the arrival of new technology gives rise to major challenges, which
must be quickly resolved through rigorous evaluation and well-balanced statements
by scientific societies. This should be done before AI implementation. To bypass all
of these pre-AI challenges, while missing the benefit of them, may be the worst start
for AI implementation in diagnostic colonoscopy!
