Keywords Endoscopy Lower GI Tract - Polyps / adenomas / ... - CRC screening - Colorectal cancer
Introduction
Colorectal cancer (CRC) stands as one of the most prevalent malignancies globally.
Implementation of CRC screening programs may reduce CRC-related mortality [1 ]. CRC has been the most common cancer for 15 years consecutively in Taiwan. According
to statistics from the Taiwanese Ministry of Health and Welfare, it was the most common
cancer
in males and the third most common in females in 2020 [2 ]. The age-standardized mortality rate for CRC has increased and it was the third
leading cause of cancer deaths, with 6,853 deaths in 2022.
Current knowledge recognizes three distinct pathways through which CRC can emerge:
the
adenoma-carcinoma sequence, the serrated pathway, and the inflammatory pathway [3 ]. The majority of cases are sporadic, influenced by factors such as dietary lifestyle,
polyp size, polyp number and presence of dysplasia.
According to current guidelines from the European Society of Gastrointestinal Endoscopy
(ESGE) and American Society for Gastrointestinal Endoscopy, clinical decision-making
regarding
CRC treatment modalities is heavily influenced by polyp size and characteristics of
invasiveness [4 ]
[5 ]
[6 ]. However, visual estimation of polyp size remains high variable among endoscopists,
even those with extensive experience [7 ]. In light of advancements in artificial intelligence (AI) technology, AI techniques
for polyp detection and characterization are becoming a reality, as well as the exciting
prospect of reliable AI-technique polyp size measurement.
Presently, Fujifilm Europe [8 ] and Argus, powered by EndoSoft [9 ], have introduced software with size estimation capabilities. However, both solutions
require a reference object for measurement. Addressing this, EndoAim [10 ], developed by ASUSTeK Computer Inc., is now one of the AI systems for colonic polyp
detection and characterization. Our objective was to enhance EndoAim by incorporating
a new
application for polyp size measurement without need for a reference object. This enhancement
involves installing verified images or video into EndoAim to generate accurate size
measurement.
We aimed to develop an AI model for estimating colorectal polyp size without the need
for
a reference object, intended for use in screening colonoscopy.
Methods
Construction of preliminary AI system and development of model for polyp size
estimation
Image segmentation is a computer vision task, by producing a dense pixel-wise
segmentation map of an image. Each pixel is then dedicated to a specific class or
object. We
used two open datasets, bkai-igh-neopolyp and kvasir-seg, for segmentation model training.
For this technique, a polyp was outlined from a colonoscopy image and sketches were
made of
the two endpoints of its longest axis ([Fig. 1 ]). Still figures with image segmentation were captured from a dynamic video
colonoscopy. For real-time image segmentation, we employed SegFormer, a robust and
efficient
approach for semantic segmentation, to develop two independent models for monocular
depth
estimation [11 ] and polyp detection. SegFormer integrates a hierarchical transformer encoder with
a
lightweight multilayer perceptron (MLP) decoder, delivering both simplicity and
computational efficiency while maintaining top-tier performance across various benchmarks.
As shown in [Fig. 2 ], the framework processes images cropped and resized to 720 × 576 pixels, which are
then used to train two separate SegFormer models: one for depth estimation and the
other for
polyp segmentation. Output features from both models are concatenated to form a regression
model for estimating polyp size.
Fig. 1 Colorectal polyp with bounding box (left) and its segmentation (right). The blue
line represents two endpoints of longest axis.
Fig. 2 Proposed incorporated framework of the model.
Snare as a tool for ground truth
The ground truth of this model is based on the fully opened snare, which was not
deformed. Clear images of the polyp within the snare ([Fig. 3 ]) were selected. Both snare edges should be visualized for measurement. A Boston
Captivator II single-use snare was chosen in the study, either 10 mm or 15 mm, due
to its
stiffness. The model was trained with a learning rate of 0.001 over 60,000 iterations.
Actual polyp size was calculated in millimeters by contrasting the two coordinates
– maximum
snare edges and the polyp – with the former serving as a standard reference ([Fig. 4 ]). In our study, the AI model was used to measure polyp size during routine
colonoscopies without any reference object. The development flowchart, including the
training and testing processes, is depicted in [Fig. 5 ].
Fig. 3 Schematic of an in vitro polyp with a fully opened snare (left) and a polyp
surrounded by a 15-mm Boston Captivator II snare (right).
Fig. 4 The yellow line represents the maximum length of both snare edges, 10 mm; the white
line indicates digital size of the polyp, 6.13 mm.
Fig. 5 Flowchart of training and testing process.
AI system construction
Model training
The initial model was trained on four detachable polyp phantoms measuring 7.5 mm,
8
mm, 10.5 mm, and 12 mm ([Fig. 6 ]). Size estimation was performed according to the aforementioned system. Because
polyp phantoms are not representative of real polyps, transfer learning was conducted.
From May to July, 2022, 172 colon polyps from 50 patients in Landseed International
hospital were retrospectively reviewed. A total of 1304 images were used for learning.
The
study was approved by the Ethics Committee of Landseed International hospital (approval
number: IRB-23-063-C0).
Fig. 6 Polyp phantoms.
Model testing
From May to October, 2023, 52 polyps were collected retrospectively from 257 routine
colonoscopies by a single endoscopist, Yii CY. In total, 178 images with polyps centered
by a snare were selected for validation ([Fig. 7 ]). The study was approved by the Ethics Committee of Landseed International
hospital (approval number: IRB-23-055-C0). An Olympus 290 series colonoscope, CF-H290L/I,
was used in this model.
Fig. 7 Fifty-two polyps were selected with 178 usable images for validation.
Polyp size obtained by snare was then compared with auto-estimation of size by the
AI
model. [Fig. 8 ] shows the process of AI auto-estimation.
Fig. 8 AI model of auto-detection, segmentation, depth estimation, and automated size
estimation.
Colonoscope
An Olympus 290 series colonoscope with s 170° field of view (FOV) was used, such as
CF-H290L/I, CF-HQ290L/I, CF-HQ290ZL/I, PCF-H290L/I, PCF-H290DL/I, PCF-H290ZL/I. In
addition, optical or digital zoom and near focus function of the colonoscope were
not
allowed.
Statistical analysis
Polyps were allocated to three size groups: ≤ 5 mm, 5 to 10 mm, and ≥ 10 mm. Error
rates represented the size difference between the snare method and the AI model. It
was
calculated by the following formula,
s
i
z
e
e
s
t
i
m
a
t
e
d
b
y
s
n
a
r
e
m
e
t
h
o
d
-
s
i
z
e
e
s
t
i
m
a
t
e
d
b
y
A
I
m
e
t
h
o
d
s
i
z
e
e
s
t
i
m
a
t
e
d
b
y
s
n
a
r
e
m
e
t
h
o
d
×
100
%
. Each of the three size groups was calculated separately. Precision indicated
the proportion of predicted true positive rate, comparing the snare method and the
AI
model, in each of the three size groups. Estimated polyp sizes, as determined by both
the
AI model and the snare method, were summarized using means and standard deviations.
To
assess the size difference between these two methods, a Bland-Altman plot was constructed.
In addition, sizes measured by the AI model and the snare method were statistically
compared using a paired t -test. Polyp sizes were categorized
into three groups (≤ 5 mm, 5–10 mm, and ≥ 10 mm). Cohen's Kappa coefficient was calculated
to evaluate agreement between the two methods for these categorical size ranges. All
statistical analyses were performed using R software, version 4.2.1. A P < 0.05 was considered to indicate statistical significance.
Sample size estimation
Sample size for our study was determined through a power analysis. This analysis
utilized the mean and standard deviation of the differences between the size estimates
in
the phantom obtained by the AI model and the snare method, or absolute percentages
of 7%
and 10%, respectively. The analysis aimed for a significance level (alpha) of 0.05
and a
power of 90% in a two-tailed paired t -test. This approach
calculated the minimum required sample size to ensure at least 20 polyps.
Results
Performance for polyp measurement
[Table 1 ] shows that errors rate for polyps ≤ 5 mm, 5 to 10 mm, and ≥ 10 mm were 11.36%,
11.46%, and 23.24%, respectively, for the training model, and 10.74%, 12.36%, and
9.89%,
respectively, for the testing model. The average error rate was 11.80% (95% confidence
interval [CI] 10.29–13.32) for the training model, and 11.47% (95% CI 10.27–12.68)
for the
testing model.
Table 1 Error rates for the snare method and AI model for three polyp size ranges: ≤ 5
mm; 5 to 10 mm, and ≥ 10mm.
CI, confidence interval of standard error of the mean.
Training model
Polyp size (mm)
≤ 5 mm
5–10 mm
≥ 10 mm
Overall
Error (%)
11.36
11.46
23.24
11.80
Testing model
Polyp size (mm)
≤ 5 mm
5–10 mm
≥ 10 mm
Overall
Error (%)
10.74
12.36
9.89
11.47
Variation in accuracy of size classifications
If a predicted polyp size by the AI method and the snare method were within the same
size group, it was considered as true. As shown in [Table 2 ], the precision for polyps ≤ 5 mm, 5 to 10 mm, and ≥ 10 mm was 0.870, 0.911, and
0.857, respectively, with an average of 0.879 in the testing group. Recall in the
testing
group represented size measurements that were correctly predicted by the AI model
for the
snare method. Recall for polyps ≤ 5 mm, 5 to 10 mm, and ≥ 10 mm was 0.941, 0.847,
and 0.750,
respectively, with an average of 0.846. Calculating the harmonic mean of the precision
and
recall, the overall F1 score was 0.861.
Table 2 Precision, recall, and F1 score of testing model in three polyps size
groups: ≤ 5 mm; 5–10 mm, and ≥ 10 mm
Snare method
≤ 5 mm
5–10 mm
≥ 10 mm
All
≤ 5 mm
80
12
0
92
5–10 mm
5
72
2
79
≥ 10 mm
0
1
6
7
All
85
85
8
178
Precision
0.87
0.911
0.857
0.879
Recall
0.941
0.847
0.75
0.846
F1 Score
0.904
0.878
0.8
0.861
Variability in size measurement between the AI model and the snare
Cohenʼs kappa value for the comparison between the snare and the AI model across the
three polyp size groups was 0.792, indicating strong consistency between these two
distinct
approaches. Moreover, the Bland-Altman analysis demonstrated a mean bias of –0.03
mm between
the methods, with no significant differences observed in the paired t -test (P = 0.638). Limits of agreement were calculated
as mean difference ± 1.96 standard deviations of the differences, ranging from –1.654
mm to
1.596 mm, as depicted in [Fig. 9 ]. Of 178 analyzed images, six images (3.4%) fell outside the 95% agreement limits,
with a maximum absolute difference of 3.47 mm.
Fig. 9 Bland-Altman plot showed that there were six cases (3.4%) out of range of the 95%
confidence interval, ± 1.6 mm (the green dashed line), whereas 33 cases (18.5%) exceeded
±1 mm (the blue dashed lines).
Discussion
In this study, we developed a novel AI model to measure colon polyp size without need
for
a reference object. Initially trained on colonoscopic images of polyp phantoms, the
model
underwent transfer learning with real-world data. We systematically compared AI-based
model
performance against the traditional snare method, which served as the ground truth
in our
testing dataset. This AI system has the potential to accurately measure colorectal
polyp size
and assess polyp characteristics, aiding in selection of appropriate polypectomy techniques
and determining surveillance intervals, ultimately contributing to reduced CRC risk.
A European study by Erlangen Group [12 ] proved that adenoma size was the most important factor related to invasive carcinoma
in a database of 11,188 adenomas from 1978 to 1993. A total of 5027 adenomas (44.9%)
that were
less than 5 mm carried no cancer risk. In addition, the cancer risk increased proportionate
with adenoma size, especially for right-sided colon adenomas [12 ]
[13 ]. Polyp size is one of the major determinants of the resection plan and interval
for
surveillance [13 ]
[14 ]
[15 ]. Recent ESGE guidelines [16 ] emphasize different and appropriate polypectomy techniques in relation to polyp
characteristics and size. For instance, cold snare polypectomy is recommended for
polyps less
than 9 mm, hot snare polypectomy for polyps 10 to 19 mm, and endoscopic mucosal resection
for
polyps larger than 20 mm.
Given the high volume of colonoscopy workload, a few studies have suggested a “discard”
strategy for polyps ≤ 5 mm without presence of unfavorable features endoscopically
[17 ]
[18 ]. The 5-mm threshold is crucial to this “resect and discard” strategy and aids
endoscopist decision about cold or hot polypectomy [19 ]. Some studies set polyp size ≥ 10 mm as a cut-off reference because of the potential
cancer risk [20 ]. Although advanced histology has been identified, frequent follow-up is recommended
in
these cases [4 ]
[14 ]. The number of adenomas also dictates the surveillance interval. According to the
American guideline, patients with only one or two low-grade dysplastic tubular adenomas
≤ 10
mm should have the next follow-up in 5 to 10 years; those with three to 10 adenomas,
or any
adenoma > 10 mm, or any adenoma with high-grade dysplasia or villous histology should
have
follow-up in 3 years [21 ]. Polyps less than 10 mm may be resected immediately during routine colonoscopy [22 ]
[23 ]. Polyps that are ≤ 20 mm need to be evaluated for immediate snare with endoscopic
mucosal resection or scheduled endoscopic submucosal dissection [24 ]. It is hoped that polyp size measurement with AI can allow the examiner to decide
on
the spot whether to remove a polyp or leave it, making the examination smoother and
faster.
Categorization of polyp size into ≤ 5 mm, 6 to 9 mm, and ≥ 10 mm is widely recognized
as
clinically important, as supported by several studies [17 ]
[18 ]
[25 ]
[26 ]. In our study, we grouped polyps into three different size categories: ≤ 5 mm, 5
to 10
mm, and ≥ 10 mm, aiming for model precision of one to two decimal places, especially
for
smaller polyps. Our observations revealed that four of six images exceeded the CI,
likely due
to factors such as surrounding feces, orientation, distance, and direction of the
open snare,
as shown in Supplementary Fig. 1. In addition, we noted that larger snares could cause
compression and deformation, introducing measurement errors. Discrepancies between
the
AI-based model and snare measurements were more pronounced for larger polyps. Specifically,
two of the six images (Supplementary Fig. 2) depicted polyps approximately 20 mm in
size, with
discrepancies greater than 1.6 mm between the AI model and the snare method (1.72
mm and 1.85
mm, respectively). However, the error rate remained below 10% due to their large size.
Furthermore, lateral spreading tumors, which span one to two colonic folds and are
significantly large, also present challenges for accurate AI measurement. Given the
infrequency of polyps exceeding 15 mm, the limited number of such images in our test
model
might also diminish confidence in measurements of larger polyps.
There are now two commercial AI products for polyp size measurement, Fujifilm Scale
Eye
and EndoSoft Argus, with laser and snare tip as reference objects, respectively. From
the
literature review, several tools were used as reference for size estimation, such
as forceps
[27 ]
[28 ], ruler [7 ], calibrated hood [29 ], graduated injection needle [30 ], or snare [26 ]. A study also has been reported of use of a novel system called Poseidon for
measurement by using the auxiliary waterjet as ground truth [31 ]. Kwak MS et al. reported on an AI technique for measuring polyp size by
bifurcation-to-bifurcation distance of colon vessels [32 ]. Until now, there was no ideal ground truth for polyp size measurement. Every tool
encounters some bottlenecks and inaccuracy. In this study, we used the snare as ground
truth
because of its availability, convenience, and its common use in daily practice. We
selected
images in which the snares were not distorted or compressed. We used only the maximum
width of
the snare that was officially written on the product cover, such as 10 mm, as our
ground
truth.
To ensure consistency of snare-based polyp measurement across different physicians,
we
asked three physicians to independently measure the size of snare-based polyps using
the same
coordinate-based method. To assess agreement among them, Fleiss' kappa was used and
a Single
Score Intraclass Correlation (ICC) was calculated across 178 subjects. The resulting
ICC was
0.942 (95% CI 0.923–0.956), demonstrating a very high level of consistency. In addition,
the
kappa value of 0.882 indicates almost perfect agreement (P <
0.001). Both the ICC and Fleiss' kappa confirm that the three raters achieved a very
high
level of consistency in their measurements and categorizations.
Our study has several limitations that warrant acknowledgment. First, the snare was
used
as the ground truth for polyp size measurement. Future studies could benefit from
incorporating additional measurement methods for comparison, such as direct measurement
of the
resected specimen. Second, having a single endoscopist in our testing model caused
selection
bias and could lead to poor generalizability of the model. Third, retrospective review
of
endoscopic pictures introduces a potential bias. Performance of a prospective study
conducted
in real-time colonoscopy would have enhanced the reliability of our findings. Fourth,
the
study would benefit from a larger sample size of polyps, ideally in a multicenter
setting.
This approach aims to mitigate potential biases associated with a limited number of
cases.
Fifth, this method is currently tailored only to an Olympus 290 series colonoscope
with a 170°
FOV. Extending the study to include a 140° FOV and exploring compatibility with other
endoscope manufacturers is mandatory for a comprehensive understanding. Sixth, we
used a
Boston Captivator II snare as the sole reference tool, which may limit the generalizability
of
our results. Considering alternative reference tools, such as a hexagonal snare, on-site
polyp
measurement by caliper, snare tip, and forceps, would enhance the comparability and
robustness
of our study.
Conclusions
In summary, our model demonstrated streamlined efficiency in screening colonoscopy
through
accurate polyp size measurement with no reference object required. This significant
advancement aligns with increasing integration of AI in healthcare.
Bibliographical Record
