Introduction
Large (≥20 mm) nonpedunculated colonic polyps (LNPCPs) are effectively and safely
resected by endoscopic mucosal resection (EMR), with 98% of lesions cured and surgery
avoided [11]
[22]
[33]. Despite this, recognition of submucosal invasive cancer (SMIC) remains a challenge,
particularly in bulky lesions [44]
[55]. Ideally lesions with SMIC are resected en bloc to facilitate accurate histologic
assessment and satisfy criteria should low risk superficial SMIC be present. Recently
it has emerged that individual LNPCP variables including location, morphology, granularity,
and size influence the risk of SMIC [44]. When the complex interaction between the effects of these variables is examined,
certain LNPCP subtypes demonstrate significantly greater risk of SMIC. The application
of this knowledge is inherently complex, hindering its accessibility in a real-world
context. This problem remains a major challenge for endoscopists of all levels of
experience and proficiency. A simplified method of identifying LNPCP subtypes with
different risks of SMIC would serve as a tool to facilitate targeted surface optical
evaluation and selective use of en bloc resection.
Methods
Study design and patient selection
From September 2008 until November 2022, consecutive participants were enrolled from
a single tertiary referral center. All LNPCPs were detected by an accredited endoscopist
(gastroenterologists and surgeons) and referred to our center for consideration of
endoscopic resection. Lesions were not proven cancers prior to referral and it was
the belief of the referring endoscopist that the LNPCP was most likely benign. Exclusion
criteria included serrated histopathology and resections that were not attempted for
technical reasons. LNPCPs not attempted owing to poor submucosal lifting and significant
submucosal fibrosis remained in the cohort when invasive cancer was strongly suspected.
Institutional ethics approval was obtained for study registration and informed consent
was collected. All authors had access to the study data and reviewed and approved
the final manuscript.
Procedure
Colonoscopy was performed using Olympus 180/190 high definition, variable stiffness
colonoscopes (Olympus, Tokyo, Japan). All endoscopic procedures were performed by
either a study investigator (accredited gastroenterologist with advanced training
and experience in colorectal endoscopic resection and LNPCP characterization) or a
senior interventional endoscopy fellow under direct supervision.
EMR was performed in a standardized fashion, with subsequent innovations adopted as
supporting data emerged [66]
[77]
[88]
[99]. A subgroup of LNPCPs underwent endoscopic submucosal dissection (ESD) as part of
a selective ESD protocol (NCT02198729) [1010]. Optical evaluation was performed under high definition white-light, narrow-band
imaging (NBI), and near focus, once these became available, to identify surface features
of SMIC. A study conducted in our center found dye-based chromoendoscopy did not provide
incremental benefit in the optical diagnosis of SMIC [1111]. As such, this was not incorporated into standard practice in our unit.
LNPCPs were consistently described using the Kudo classification throughout the duration
of the study. Procedural documentation and LNPCP morphology descriptions were reported
in a standardized fashion to maintain consistency between proceduralists and minimize
incomplete data.
Specimens were collected and processed for histopathology review, in accordance with
Australasian Gastrointestinal Pathology Society guidelines [1212]. Histopathology review was completed by expert gastrointestinal pathologists and,
where SMIC was identified, consensus opinion was obtained. Furthermore, cases of identified
cancer were discussed at a multidisciplinary meeting with both pathologists and gastroenterologists
in attendance.
Collected data, definitions, and outcomes
Prospectively collected data included LNPCP location, size, Paris classification,
surface granularity, and histopathologic diagnosis.
SMIC was defined as neoplasia invading beyond the muscularis mucosae into the submucosa.
SMIC was deemed overt if Kudo Vi or Vn features were evident on surface optical evaluation.
Endoscopic resection of SMIC was performed either on LNPCPs that were optically suspicious
for potentially curable superficial SMIC (Kudo Vi) or in cases of piecemeal EMR in
which SMIC was “covert” and not recognized prior to resection owing to the absence
of Kudo Vi or Vn features. LNPCPs not resected owing to optical evidence of deeply
invasive cancer had malignancy confirmed through cold forceps biopsies. If biopsies
could not confidently provide a diagnosis of cancer, lesions were only deemed to have
SMIC following evaluation of the surgical outcomes. Unless the patient was deemed
medically unfit or the lesion was curatively resected with en bloc techniques, all
cases of SMIC underwent surgical resection.
LNPCP location was divided into the proximal colon and rectosigmoid. The rectosigmoid
was defined as the rectum and sigmoid colon; the proximal colon as proximal to and
including
the descending colon. LNPCP morphology was described according to the Paris classification
and was further grouped as “flat” vs. “nodular” or “depressed.” Flat LNPCPs were Paris
0-IIa
or IIb, nodular LNPCPs contained a sessile component (Paris 0-Is or 0-IIa+Is with
a nodule
exceeding 2.5 mm in size), and depressed lesions contained Paris 0-IIc foci. LNPCP
size was
measured using endoscopic devices of known size, such as snares, as a frame of reference.
Size was considered as both a continuous and binary variable (<40 or ≥40mm). Granularity
was classified as granular or nongranular in appearance. Mixed granularity LNPCPs
with both
granular and nongranular features were classified as nongranular. High SMIC prevalence
was
defined as a probability exceeding 10%, as described in previous studies [44].
The primary aim of the study was to identify combinations of LNPCP characteristics
(“LNPCP subtypes”) that were associated with an increased prevalence of SMIC in the
study cohort and display these in a simplified decision-making algorithm that stratifies
the risk of potential cancer in LNPCPs otherwise thought to be benign by the referring
endoscopist.
Statistical analysis
IBM SPSS Statistics version 29.0 (IBM, Armonk, New York, USA) was used to analyze
the data. All analyses were exploratory and per LNPCP. Two-tailed tests with a significance
level of 5% were used throughout. Continuous variables are summarized as the median
(interquartile range [IQR]) or mean (SD). Categorical variables are summarized as
frequency and percentage. LNPCP size is summarized using median (IQR). The dichotomized
size variable (≥40 vs. <40 mm) maximized the sum of sensitivity and specificity when
classifying SMIC status (present vs. absent).
Chi-squared tests were used to test for univariable association between each categorical
variable and the presence of SMIC; the Mann–Whitney test was used for size. Odds ratios
(ORs) with 95%CIs, estimated using logistic regression analysis, were used to quantify
the strength of the univariable associations. We identified the best fitting multivariable
logistic regression model using backward stepwise variable selection from the main
effects (size, granularity, morphology, and colon side) and their pairwise interactions
with P value for removal <0.1.
Classification trees were then used to explore whether a simple decision tree based
on
only four predictor variables and constrained to have a minimum number of 200 cases
in any
parent node and 100 in any child node could better communicate a clinical decision
pathway.
The analysis was used to classify the LNPCPs into groups based on the predictor variables
morphology, granularity, location, and size according to the dependent variable SMIC
status
(present vs. absent). The chi-squared automatic interaction detection (CHAID) method
was
used to grow the tree to a maximum depth of five levels, ensuring the minimum number
of
LNPCPs at parent and child nodes were 200 and 100, respectively. This guaranteed the
95%CI
for percentage SMIC within any terminal node was no wider than ±10% about the observed
value. P values associated with the splits were Bonferroni
adjusted.
All lesions and the overall SMIC rate appear at the first level of the tree. The variable
with the greatest impact on the dependent variable (SMIC rate) is then identified
and breaks down the population into child nodes. This process of selection of the
most influential variable is repeated at each child node at each level using the remaining
variables. Five-fold cross-validation was used to confirm the structure of the classification
tree and the stability of the proportions. The “gain” (number with SMIC in the node)
summary table was produced for the terminal nodes and used to further simplify the
potential risk of SMIC for LNPCPs into four categories. The Nagelkerke R
2 value, which quantifies the goodness of fit of logistic regression models, was used
to compare the fit for this simple four category “potential SMIC” variable to that
of the best fitting logistic regression model and to that using the subtype categories
identified by the terminal nodes of the decision tree.
Results
A total of 3039 LNPCPs were assessed in the study period; 512 serrated lesions were
excluded, along with 19 LNPCPs that were not resected because of technical difficulties
including poor endoscopic access. There were 23 patients who had missing histology
or lesion data, and granularity was unclassified for 22 LNPCPs; 12 LNPCPs had insufficient
tissue or nonadenomatous histology (lipoma, colitis, lymphoma).
Overall, 2451 LNPCPs from 2260 enrolled patients were included, with a median size
of 35
mm (IQR 25–50mm), predominant proximal colon location (n=1669; 68.1%), flat morphology
(n=1289; 52.6%), and granular appearance (n=1603; 65.4%) ([Table 1Table 1]). SMIC was identified in 273/2451 LNPCPs (11.1%). Covert SMIC was present in 42.8%
of
all cancers (Table 1s, see online-only Supplementary material) and
4.7% of all LNPCPs. Overt SMIC was present in 6.3% of all LNPCPs.
Table 1
Table 1 Baseline characteristics of the 2451 large nonpedunculated colonic polyps included
in the study.
|
Characteristic
|
n (%) unless otherwise specified
|
|
* Missing data for Paris type, n=1.
|
|
Histopathology
|
Tubular adenoma
|
699 (28.5%)
|
|
Tubulovillous adenoma
|
1461 (59.6%)
|
|
Villous adenoma
|
18 (0.7%)
|
|
Submucosal invasive cancer
|
273 (11.1%)
|
|
Size, median (IQR), mm
|
35 (25–50)
|
|
Location
|
Rectosigmoid
|
782 (31.9%)
|
|
Proximal colon
|
1669 (68.1%)
|
|
Morphology*
|
Flat
|
1289 (52.6%)
|
|
Nodular
|
1043 (42.6%)
|
|
Depressed
|
118 (4.8%)
|
|
Granularity
|
Granular
|
1603 (65.4%)
|
|
Nongranular
|
848 (34.6%)
|
LNPCP characteristics associated with SMIC
SMIC was associated with: depressed (73/118; 61.9%) and nodular (144/1043; 13.8%)
vs.
flat (56/1289; 4.3%) morphology (OR 35.7 [95%CI 22.6–56.5] and OR 3.5 [95%CI 2.6–4.9],
respectively; P<0.001); rectosigmoid (154/782; 19.7%) vs.
proximal (119/1669; 7.1%) colonic location (OR 3.2 [95%CI 2.5–4.1]; P<0.001); nongranular (146/848; 17.2%) vs. granular (127/1603; 7.9%) appearance
(OR 2.4 [95%CI 1.9–3.1]; P<0.001); and size ≥40mm (160/1135;
14.1%) vs. <40mm (113/1316; 8.6%) cohort (OR 1.7 [95%CI 1.4–2.3]; P<0.001) ([Table 2Table 2]).
Table 2
Table 2 Distribution of each large nonpedunculated colonic polyp characteristic by submucosal
invasive cancer (SMIC) status, together with odds ratios (ORs) and 95%CIs derived
from multivariable regression analysis.
|
Variable
|
SMIC present, n (%) unless otherwise stated
|
P value
|
Odds ratio (95%CI)
|
|
No
(Total 2178)
|
Yes
(Total 273)
|
|
* Per 10-mm increase in size.
† Missing data for Paris type, n = 1 in the no SMIC group.
|
|
Size, median (IQR), mm
|
35 (25–45)
|
40 (30–50)
|
<0.001
|
1.12* (1.05–1.19)
|
|
Size ≥40mm (vs. <40mm)
|
975 (44.8%)
|
160 (58.6%)
|
<0.001
|
1.75 (1.35–2.26)
|
|
Size ≥35mm (vs. <35mm)
|
1212 (55.6%)
|
174 (63.7%)
|
0.01
|
1.40 (1.08–1.82)
|
|
Morphology† (vs. flat)
|
|
|
899 (41.3%)
|
144 (52.7%)
|
<0.001
|
3.53 (2.56–4.86)
|
|
|
45 (2.1%)
|
73 (26.7%)
|
35.7 (22.6–56.5)
|
|
Nongranular (vs. granular)
|
702 (32.2%)
|
146 (53.5%)
|
<0.001
|
2.42 (1.87–3.12)
|
|
Rectosigmoid location (vs. proximal)
|
628 (28.8%)
|
154 (56.4%)
|
<0.001
|
3.19 (2.47–4.13)
|
Decision tree analysis targeting SMIC ([Fig. 1Fig. 1]) identified eight terminal nodes (“LNPCP subtypes”): SMIC prevalence was 62% in
118
depressed LNPCPs; 22% in 363 nodular rectosigmoid LNPCPs ≥40mm; 12% in 160 nodular
rectosigmoid LNPCPs <40mm; and 20% in 125 nodular proximal colon nongranular LNPCPs.
The
lowest prevalence (1%) occurred in 596 flat proximal colon granular LNPCPs. The decision
tree that used LNPCP size as a continuous variable was identical to that which used
dichotomized size and had a depth of 3. Table 2s shows the
percentage SMIC prevalence, with 95%CI, within each LNPCP subtype for the study
cohort.
Fig. 1
Fig. 1 Decision tree analysis targeting submucosal invasive cancer (SMIC) in the study cohort
and illustrating the terminal nodes (“large nonpedunculated colonic polyp subtypes”).
Left colon, rectosigmoid; right colon, proximal colon.
[Table 3Table 3] shows the “gain” (number with SMIC in node) summary table for the terminal nodes
(subtypes) and four broad categories of “potential SMIC” risk, based on SMIC prevalence
for
each LNPCP subtype: high (>50%); elevated (>10%–50%); unlikely (2.5%–10%); and very
unlikely (<2.5%). The Nagelkerke R
2 value,
quantifying goodness of fit of the logistic regression model, for SMIC status using
this
four-level “potential SMIC” categorical independent variable was 0.257. This is very
similar
to that using the eight-level subtype variable (Nagelkerke R
2=0.265) and that for the best fitting logistic regression model,
which incorporated four main effects and colonic location by granularity interaction
(R
2=0.274).
Table 3
Table 3 Large nonpedunculated colonic polyp (LNPCP) and submucosal invasive cancer (SMIC)
frequency (“gain”) by terminal node subtype, together with percentage SMIC within
each subtype, index ratio, and “potential SMIC” category.
|
Terminal node
|
LNPCP subtype
|
LNPCP frequency,
n (%)
|
Gain (number with SMIC),
n (%)
|
Percentage SMIC within subtype
|
Index ratio*
|
Potential SMIC category
|
|
* Ratio of percentage with SMIC within subtype compared with percentage with SMIC
for the total LNPCP cohort.
|
|
3
|
Depressed
|
118 (4.8%)
|
73 (26.7%)
|
61.9%
|
5.6
|
High risk (>50%)
|
|
13
|
Nodular, rectosigmoid, ≥40mm
|
363 (14.8%)
|
81 (29.7%)
|
22.3%
|
2.0
|
Elevated risk (>10%)
|
|
10
|
Nodular, proximal colon, nongranular
|
125 (5.1%)
|
25 (9.2%)
|
20.0%
|
1.8
|
|
12
|
Nodular, rectosigmoid, <40 mm
|
160 (6.5%)
|
19 (7.0%)
|
11.9%
|
1.1
|
|
4
|
Flat, nongranular
|
558 (22.8%)
|
43 (15.8%)
|
7.7%
|
0.7
|
Unlikely (2.5%–10%)
|
|
9
|
Flat, rectosigmoid, granular
|
136 (5.5%)
|
7 (2.6%)
|
5.1%
|
0.5
|
|
11
|
Nodular, proximal colon, granular
|
395 (16.1%)
|
19 (7.0%)
|
4.8%
|
0.4
|
|
8
|
Flat, proximal colon, granular
|
596 (24.3%)
|
6 (2.2%)
|
1.0%
|
0.1
|
Very unlikely (<2.5%)
|
|
Total LNPCP cohort
|
2451 (100%)
|
273 (100%)
|
11.1%
|
1
|
|
The decision tree results were further simplified into a SMIC decision-making algorithm
that highlights high risk LNPCPs with a risk of potential SMIC >10% ([Fig. 2Fig. 2]). [Table 3Table 3] provides a more detailed breakdown of the LNPCP subtype risk categories.
Fig. 2
Fig. 2 Decision-making algorithm highlighting large nonpedunculated colonic polyp subtypes
where the potential SMIC risk is >10%.
Discussion
Detection of covert and overt SMIC in LNPCPs remains a considerable challenge. This
is a problem that impacts endoscopists of all levels of experience and proficiency.
We have recently shown that SMIC can be reliably detected in flat lesions through
the expression of invasive features on the surface of the lesion [1313]; however, the accuracy of this is greatly diminished in bulky lesions. While EMR
is proven to be safe and effective for the treatment of LNPCPs, those containing SMIC
are not considered cured by piecemeal resection, according to accepted criteria [11]
[22]
[33]
[1414]
[1515]
[1616]. Various factors feed into SMIC estimation and can be used to improve optical diagnosis
and optimize the resection modality. Current methods are challenging to apply in a
real-world setting. A simple approach that can be easily applied by all endoscopists
is needed.
SMIC has previously been reported in 7.6%–8.5% of endoscopically resected LNPCPs and,
when
identified and appropriately resected, may be cured [44]
[55]. En bloc resection of SMIC is curative when favorable histologic features are present,
including superficial invasion (<1000µm), no lymphovascular invasion, and absence
of poor
differentiation. Under these circumstances, surgical resection is generally not necessary
[1717]. Even when SMIC is inadvertently resected piecemeal, it seems the same curative
criteria that are used for en bloc resection apply, although data are limited. If
SMIC is well
differentiated with no lymphovascular invasion, the risk of lymph node metastasis
is
negligible. Furthermore, if the deep margin is clear, despite piecemeal resection,
it seems
the risk of local recurrence is insignificant [1818].
Optical evaluation of the surface vascular and pit patterns of LNPCPs can accurately
detect SMIC in flat lesions. The probability of optical evaluation not detecting SMIC
in a
flat LNPCP was 0.6% in a large prospective study; in contrast, because of their bulky
morphology limiting inspection, optical evaluation missed 5.9% of cases of SMIC within
nodular
LNPCPs (P<0.001) [1313]. A prospective series of 2277 LNPCPs found the sensitivity of Kudo pit pattern for
identifying SMIC was only 40.4%, with 138/171 histologically confirmed cancers demonstrating
benign surface features [44]. These cases of “covert” SMIC present a significant challenge in the detection and
management of SMIC in LNPCPs and highlight the importance of risk characterization
independent
of the surface features.
The individual LNPCP characteristics of size, morphology, location, and granularity
inform the baseline risk of SMIC, irrespective of the surface pit and vascular patterns.
Compared with flat lesions, nodular or depressed morphology demonstrates a greater
probability of SMIC [1919]
[2020]. Overall, nongranular LNPCPs are more likely to contain SMIC compared with their
granular counterparts [44]. Whereas LNPCPs located in the proximal colon have a low risk of SMIC, lesions in
the distal colon, particularly the rectosigmoid, are at far greater risk [44]
[2121]. While these individual risk factors are widely reported, their collective influence
on SMIC risk is now also well recognized [44]
[2222]
[2323]. The interaction of these characteristics is however complex and difficult to apply
in a real-world setting. Moreover, a simple algorithm estimating risk in a given LNPCP
subtype has been lacking.
Given the complex interaction of LNPCP characteristics when predicting SMIC risk,
we used a decision tree approach to identify subtypes with different risks of SMIC
based on morphology, granularity, colonic location, and polyp size. Paris morphology
has been shown to have poor interobserver agreement among experts [2424] and was therefore not included as an independent predictor in the decision tree
analysis.
Our novel algorithm offers readily available lesion-specific risks of potential SMIC
according to the LNPCP subtype ([Fig. 2Fig. 2] and [Fig. 3Fig. 3]). This can assist endoscopists in stratifying the risk of cancer in an LNPCP that
is otherwise thought to be benign. Endoscopists first need to characterize the LNPCP
as depressed, flat, or nodular. Depressed LNPCPs require no further characterization
and have a 62% prevalence of SMIC. For such lesions, in the absence of surface features
of deep submucosal invasion, endoscopists should proceed with an en bloc resection
modality. LNPCPs that are not depressed are then divided into flat or nodular. Irrespective
of granularity or location, flat LNPCPs confer a low prevalence of SMIC (4.3%: granular,
1.8%; nongranular, 7.7%). Endoscopists must be mindful of this risk profile prior
to commencing optical assessment, which has been demonstrated to be highly accurate
in flat lesions [1313]. For lesions that are nodular, location assumes primacy as a discriminatory variable.
In the proximal colon, nodular nongranular LNPCPs have a high probability (20%) of
SMIC, whereas nodular granular LNPCPs have a low prevalence of SMIC (5%). In contrast,
all rectosigmoid nodular LNPCPs are high risk, irrespective of their granularity or
size (19%). Once high risk lesion subtypes have been identified, fastidious optical
evaluation of their surface features should be conducted to exclude the presence of
deep submucosal invasion. If these features are absent, an en bloc resection modality
is required.
Fig. 3
Fig. 3 Examples of risk assessment of large nonpedunculated colonic polyps using the simple
algorithm for potential SMIC showing: a–c step 1 = nodular lesion; step 2 = location, rectosigmoid → potential SMIC risk 19%
(en bloc resection performed); d–f step 1 = flat → potential SMIC risk 4% irrespective of location (resected by piecemeal
endoscopic mucosal resection after surface optical evaluation).
This decision-making algorithm is a tool that can be used by endoscopists of all levels
of experience and training to stratify the risk of potential cancer in an LNPCP that
is otherwise thought to be benign. Risk calculations, which necessitated prerequisite
knowledge of twelve LNPCP subtypes, can now be conducted on individual lesions using
binary questions addressing LNPCP size, morphology, location, and granularity. The
flowchart presentation is intuitive, with potential SMIC risk often calculated using
between one and three LNPCP characteristics ([Fig. 2Fig. 2]). A detailed understanding of SMIC risk of an LNPCP is necessary before conducting
optical evaluation of the lesion’s surface. This particularly applies to flat LNPCPs.
Conversely, the inaccuracy of optical evaluation within nodular LNPCPs requires accurate
risk stratification to inform the resection strategy.
With the use of our algorithm, high risk nodular lesions can be easily identified
and subsequently targeted in validated en bloc selective resection protocols. One
such protocol performed ESD on high risk lesions of the rectum, which included Paris
0-Is or 0-IIa+Is nongranular LNPCPs or 0-IIa+Is granular LNPCPs with a dominant nodule
≥10 mm. Selective resection accurately captured cases of SMIC, with curative oncologic
resection achieved in all cases satisfying the criteria for low risk SMIC [1010]. Despite the utility of this algorithm, not all lesions with SMIC are detected,
emphasizing the importance of high quality optical evaluation during lesion assessment.
This algorithm is specific to adenomatous LNPCPs, with serrated lesions excluded from
the final analysis. In contrast to adenomatous polyps, serrated lesions demonstrate
a unique carcinogenesis pathway and have distinct endoscopic appearances and resection
approaches [2525]. Unlike adenomas, serrated lesions with cytologic dysplasia are well described precancerous
lesions that are endoscopically detectable [2626]. Furthermore, it must be emphasized that all LNPCPs referred for endoscopic resection
in our study were thought to be benign by the referring endoscopists. Cases of overt
cancer were only subsequently diagnosed by our service. Consequently, covert and overt
cancers were retained within the study to accurately construct a tool that stratifies
the risk of potential cancer in LNPCPs otherwise thought to be benign by the referring
endoscopists.
The strength of this study is the large cohort of LNPCPs that were prospectively collected
and characterized at a single expert referral center with substantial expertise in
LNPCP characterization, assessment, and resection.
A theoretical limitation is the real-world variability in classifying lesion morphology
and size; however, our algorithm has simplified this process and does not rely on
variables such as Paris classification, which may have poor agreement between proceduralists
[2424]
[2727]. We do acknowledge however that this study lacks data on the level of agreement
between our proceduralists regarding lesion description. This is a derivation study
without external validation. To ascertain the generalizability of the results, external
validation in an independent dataset is required to obtain data on algorithm accuracy.
We also recognize that our rate of SMIC (11%) exceeds that of previous studies; however,
this is likely a consequence of including overtly cancerous lesions that were referred
to us for assessment and endoscopic treatment, which we then sent directly to surgery.
Being a tertiary referral center there is unavoidable referrer bias as we recognize
that some endoscopically resectable cancers detected by referrers may not have been
considered for resection and referred directly to surgery. While subject to debate
among experts, a recent study highlighted incomplete retrieval of piecemeal resection
specimens as a theoretical source of missed foci of SMIC [2828]. We recognize a large proportion of our EMR cases were piecemeal resections and
as such lack the histologic accuracy obtained from an en bloc specimen.
In conclusion, this study has defined an algorithm to stratify the likelihood of potential
SMIC in an LNPCP otherwise thought to be benign. The algorithm encourages systematic
LNPCP assessment prior to endoscopic resection and provides all endoscopists with
greater confidence in decision-making around treatment selection.
