Keywords
EndoBRAIN - endocytoscopy - polyp
Introduction
Colorectal cancer (CRC) is a formidable health problem worldwide. In India, the annual
incidence rates (AARs) for colon cancer and rectal cancer in men are 4.4 and 4.1 per
1,00,000, respectively. The AAR for colon cancer in women is 3.9 per 100000.[1]
Artificial intelligence (AI) known as computer vision in computer-aided diagnosis
(CAD) and detection (CADe) helps in identifying health-related conditions based on
medical imaging. Convolutional neural network (CNN) is a type of deep machine learning
algorithm that uses convolutions of the input image to extract the most relevant information
that helps to classify the image into different entities. Based on the accumulated
data features, a deep CNN can diagnose newly acquired clinical images prospectively.[2]
[3]
Precise optical diagnosis of colorectal polyps could improve the cost-effectiveness
of colonoscopy and reduce polypectomy-related complications. It is difficult for community-based
nonexperts to obtain sufficient diagnostic performance. CAD has potential for better
accuracy and lower interobserver variability. Nonexpert endoscopists may more easily
achieve accuracy levels sufficient to meet the preservation and incorporation of valuable
endoscopic innovations (PIVI) threshold.[4] The EndoBRAIN technology has a potential in this regard with studies showing improved
adenoma detection rate (ADR) and diagnostic accuracy reaching the PIVI thresholds.
Removing precancerous polyps from the bowel during a colonoscopy is the cornerstone
of CRC screening and prevents polyps developing into bowel cancer. Many polyps never
grow into cancer and it can be difficult for the clinicians performing the procedure
(endoscopists) to tell which ones are precancerous. This means many polyps are removed
unnecessarily, with a considerable waste of resources. The EndoBRAIN system uses optical
diagnostic technologies like endocytoscopy (EC) and narrow band imaging (NBI). EC
enables in vivo observation of cells and nuclei at 520x ultramagnification using methylene
blue staining, and combined with NBI, can observe microvessels in detail.[5]
[6]
[7]
[8] EndoBRAIN may prove to be cost-effective by reducing biopsies and histopathology
examinations. Usage of these technologies, especially in a high-volume center, may
help us improve patient care, at the same time with cost-effectiveness.[9]
Methods
This was a prospective, observational study conducted to estimate the accuracy of
visual inspection alone and of EndoBRAIN (EC-CAD) in identifying a lesion as neoplastic
or nonneoplastic using EC. The study population included individuals 18 years or older
who were scheduled for screening, surveillance, diagnostic, or therapeutic colonoscopy.
Patients with inflammatory bowel disease, polyposis syndrome (e.g., familial adenomatous
polyposis, serrated polyposis), history of chemotherapy or radiation therapy for colorectal
lesions, and inability to undergo polypectomy (e.g., intake of anticoagulants, comorbidities,
or patient refusal) were excluded from the study. Subjects who were scheduled for
colonoscopy and found to have a polyp on white light endoscopy were included. Patients
underwent colonoscopy with Olympus colonoscope (CF-HQ290, Olympus, Tokyo, Japan) equipped
with EndoBRAIN technology and performed using EVIS LUCERA ELITE CV-290 processor (Olympus,
Tokyo, Japan). Experienced endoscopist who performed more than 5,000 colonoscopies
have performed the procedure. Endoscopic diagnosis of polyp was done under white light
imaging and NBI using Japan NBI expert team classification. EC-NBI and/or EC-stain
images were acquired from the polyps. The acquired images were assessed by endoscopist
in real-time and was asked to give a diagnosis (nonneoplastic/neoplastic) who was
blinded to the EndoBRAIN diagnosis and histopathological diagnosis. The EndoBRAIN
diagnosis of the polyp on EC NBI (nonneoplastic/neoplastic) and/or EC stain (nonneoplastic/neoplastic)
images was documented by the assistant. Resected polyps were sent for histopathological
assessment who were blinded to the endoscopic diagnosis. Pathological assessment of
polyps was performed by senior pathologist with experience in the gastrointestinal
histopathology. The number of polyps from which good quality EC images could be acquired
is calculated (image acquisition rate). The sensitivity, specificity, accuracy, positive
predictive value (PPV), negative predictive value (NPV), positive likelihood ratio,
and negative likelihood ratio in identifying a neoplastic lesion are calculated. Polyps
from which EC images could not be acquired were excluded from this analysis.
Statistical Analysis
Baseline characteristics of the polys were described using descriptive statistics.
Categorical data are described using percentages and frequencies and compared using
Fisher's exact test or chi-squared test. The normality of continuous data was assessed
by Kolmogorov–Smirnov test and represented as mean (standard deviation) or median
(interquartile range). Comparison of the continuous data was done by independent Student's
t-test for parametric data and Mann–Whitney U-test for nonparametric data. Statistical
analysis was performed at 5% level of significance and p less than 0.05 was considered as statistically significant.
Results
This is a pilot study conducted from January 2021 to June 2021. Institute review board
and ethical board clearance was obtained prior to initiating the study (AHF/AIGH-IRB:02/46/2021).
Study was conducted in accordance with ethical principles for human subjects as stated
in the declaration of Helsinki. Informed consent was obtained from all the participants.
Baseline characteristics showing number of polyps, image acquisition rate, and histopathological
details were elucidated in [Table 1]. A total of 55 polyps were studied. Most of the polyps were diminutive and most
of them were located in rectum. Good quality EC images using either EC-NBI or EC-stain
mode were acquired from 43 out of 55 polyps (78.2%). However, the image acquisition
rate was lower in diminutive polyps at 66.7% (24 out of 36 polyps). Histopathological
examination of the polyps from which EC images were acquired showed 22 (51%) nonneoplastic
and 21 (49%) neoplastic polyps. EndoBRAIN (EC-CAD) detects a polyp as neoplastic or
nonneoplastic using EC in real-time ([Supplementary Video S1]). Nonneoplastic polyp on EC showed narrow serrated lumina and dense pattern of small
roundish nodules ([Fig. 1]). Neoplastic polyp showed slit like smooth lumina and regular pattern of fusiform
or roundish nuclei ([Fig. 2]). The sensitivity, specificity, PPV, NPV, and accuracy of endoscopist in identifying
a neoplastic polyp based on EC were 90.48% (95% confidence interval [CI], 69.2–98.8),
81.81% (95% CI, 59.7–94.8), 82.61% (95% CI, 65.95–92.1), 90% (95% CI, 70.36–97.15),
and 86% (95% CI, 72.07–94.70), respectively, with a positive likelihood ratio of 4.98
and negative likelihood ratio of 0.12. The sensitivity, specificity, PPV, NPV, and
accuracy of EndoBRAIN in identifying a neoplastic polyp based on EC were 100%, 81.82%
(95% CI, 59.7–94.8%), 84%(95% CI, 68.4–92.72%), 100%, and 90.7% (95% CI, 77.86–97.41%),
respectively, with a positive likelihood ratio of 5.5 and negative likelihood ratio
of 0. The sensitivity and NPV of EndoBRAIN were significantly better than that of
endoscopist (p < 0.05). Though diagnostic accuracy is more with EndoBRAIN, it did not reach statistical
significance (p = 0.5). Specificity and PPV were similar in both groups ([Table 2]).
Fig. 1 Nonneoplastic polyp. (A) White light imaging showing sessile polyp of size 5 mm (Paris 1s). (B) Narrow band imaging (NBI) showing sessile polyp with invisible vascular pattern
with regular dots (Japan NBI expert team type 1). (C and D) Endocytoscopy with NBI and stain showing narrow serrated lumina and dense pattern
of small roundish nodules (EC1b).
Fig. 2 Neoplastic polyp. (A) White light imaging showing sessile polyp of size 5 mm (Paris 1s). (B) Narrow band imaging (NBI) showing sessile polyp with regular vessel distribution
and caliber with regular surface (Japan NBI expert team type 2A). (C and D) Endocytoscopy with NBI and stain showing slit-like smooth lumens with uniform fusiform
or roundish nuclei (EC2).
Table 1
Baseline characteristics
|
Total no of polyps
|
55
|
|
Size
|
|
Diminutive polyps
|
36 (65.4%)
|
|
Polyps of size 5 mm–1cm
|
10 (18.1%)
|
|
Polyps of size > 1 cm
|
9 (16.3%)
|
|
Location
|
|
Rectum
|
21 (38.1%)
|
|
Left colon
|
15 (27.2%)
|
|
Right colon
|
19 (34.5%)
|
|
Image acquisition rate
|
43/55 (78.2%)
|
|
Image acquisition rate from diminutive polyps
|
24/36 (66.7%)
|
|
Histopathology of polyps from which endocytoscopy images acquired (n = 43)
|
|
Hyperplastic
|
20
|
|
Inflammatory
|
2
|
|
Low-grade adenoma
|
16
|
|
High-grade adenoma
|
5
|
Table 2
Comparison of evaluation of endocytoscopy between endoscopist and EndoBRAIN (n = 43)
|
Endoscopist
(confidence interval)
|
EndoBRAIN (EC-CAD)
(confidence interval)
|
p-Value
|
|
Sensitivity
|
90.48% (69.2–98.8)
|
100%
|
0.03
|
|
Specificity
|
81.81% (59.7–94.8)
|
81.82% (59.7–94.8)
|
0.9
|
|
Positive predictive value
|
82.61% (65.95–92.1)
|
84% (68.4–92.72)
|
0.86
|
|
Negative predictive value
|
90% (70.36–97.15)
|
100%
|
0.03
|
|
Accuracy
|
86% (72.07–94.70)
|
90.7% (77.86–97.41)
|
0.5
|
Abbreviation: EC-CAD, endocytoscopy-computer-aided diagnosis.
Discussion
In the past decade, development in AI and its applications in the medical field were
exponential. Being third-leading malignancy, technical, operator, and human dependent
limitations are missing out on a significant proportion of polyps during colonoscopy
in CRC patients. These errors ultimately affect the patients and their overall CRC
management. It was also reported that with each 1% increase in ADR, an equivalent
3% decrease in the subsequent risk of cancer was reported.[10]
[11] In view of this, the highest level of accuracy is highly essential and much needed
to deal with such unmet problems with minimal errors that can be feasible only with
AI. AI has its own advantages in diagnosing the polyp characteristics easily, early,
accurately, and economically than the existing conventional ex-vivo microscopic analysis
methods.[12] Considering the facts, the AI powered CADe and diagnosis (CADx) systems were developed
to improve the nonhistological polyp evaluation with better accuracy and reduced intra-
and interobserver variability.[13]
To overcome the limitations with the existing CADx systems, Kudo et al[14] and Mori et al[15] have collaborated, designed, and developed an advanced novel AI technology-based
CADe (EndoBRAIN-EYE) and CADx tool—EndoBRAIN to help the surgeons in real-time (in-vivo)
to differentiate the nonneoplastic lesions from neoplastic and help in avoiding unnecessary
resection.[16] To the best of our knowledge, this is the first of its kind study in India to use
EndoBRAIN (an AI software tool) to differentiate neoplastic versus nonneoplastic polyp
in real-time during a colonoscopy.
The first step during the procedure is identification and differentiation of cancerous
lesions from noncancerous lesions in vivo using high-quality images acquired and analyzed
in real time. The image acquisition rate in our study (78.2%) was highly efficient
and it was in line with a large multicenter study conducted by Mori et al (83.6%).[17] This parameter is important in the context that unless a high-quality EC image is
acquired, EndoBRAIN does not give us an output. The image acquisition was difficult
in diminutive polyps at high magnification because minimal movement at the time of
freezing the image can lead to false results. Images were also difficult to obtain
polyps located in traditionally difficult locations like hepatic and splenic flexure.
Better bowel preparation, higher procedure volumes, and strategies such as examining
the polyp at 12'0 clock position are few ways to improve the image acquisition in
our experience.
Study by Mori et al is one of the most significant and the first benchmark study ever
conducted using EndoBRAIN to clarify the value of an AI-assisted colonoscopy system
in identifying cancerous lesions under the strictly controlled environment.[17] Another important parameter to be considered as a benchmark for optical diagnosis
in adoption of AI systems as a clinical decision support device for diminutive polyp
management is PIVI thresholds. Higher the accuracy of optical polyp diagnosis, higher
will be the PIVI acceptance. Based on the PIVI threshold, anyone of the paradigm will
be opted in the end clinically—“resect and discard” or “leave in situ.” In the present
study, the accuracy rate of EndoBRAIN was 90.7%, exceeding the initiative threshold
of more than 90% for the “resect and discard” strategy as proposed by the American
Society for Gastrointestinal Endoscopy.[4] Diagnostic accuracy reports of our study (90.7%) were observed to be similar to
the reports obtained from Misawa et al[6] (87.8%) and Shin et al.[14] In our study, an improvement in accuracy of EndoBRAIN (90.7%) over endoscopist (86.05%)
was also observed to be same as Jin et al where the use of CADx has improved the overall
accuracy of optical polyp diagnosis from 82.5 to 88.5% (p < 0.05).[18] With advancement of technology like this, many nonexpert endoscopists around the
world can now easily achieve accuracy levels sufficient to meet the PIVI threshold.
Another major parameter to consider is NPV, where the NPV results (100%) from our
study were observed to be better than multiple studies from the literature allowing
diminutive hyperplastic polyps to be left in situ without a pathological diagnosis.
In studies conducted by Mori et al[17] and Shin et al,[14] the NPV in both the stain mode and NBI mode was observed to be only more than or
equal to 93%.
Whereas with the other parameters concerned such as EndoBRAIN sensitivity, specificity,
and PPV values were also reported and they were observed to be better than values
from endoscopists. Reports from our study were almost similar with the results published
by the team who developed the EndoBRAIN itself. However, the sensitivity, specificity,
PPV values of our study and EndoBRAIN team were reported to be 100, 81.82, 84 and
96.9, 100, 100%, respectively.[15] Reports from Shin et al also showed almost similar results as our study in both
stained EC and EC-NBI.[14] Overall, results from our study suggest that the sensitivity and NPV are statistically
significant and better in EndoBRAIN than that of endoscopists group suggesting the
efficiency of the EndoBRAIN and its unlikely nature to miss a neoplastic polyp. In
addition, EndoBRAIN is a good alternate to conventional methods in terms of cost-effectiveness,
time-saving, and the trauma involved throughout the process. In future, these AI-based
diagnostic systems like EndoBRAIN can be a game changer in reducing the unnecessary
surgeries/resections because of their high accuracy, NPV, and specificity. In future,
these AI systems also have a high potential to transform clinical endoscopic practice
positively forever over the existing conventional procedures.
There are some limitations to our study. First, sample size is very small; hence,
it is difficult to generalize the findings to community. Second, sessile serrated
adenoma (SSA) that appears similar to hyperplastic polyps on digital chromoendoscopy
was not studied in our study. As such the incidence of SSA is low and predominant
distal location of polyps in our study may be the reason for not having SSA. Further
polyp surveillance studies with EndoBRAIN involving SSA are required to conclude on
“do not resect” strategy. Third, objective assessment of additional time required
to perform procedure and cost-effective analysis was not performed.
Conclusion
Optical diagnosis using EC and EC-CAD has a potential role in predicting the histopathological
diagnosis. The diagnostic performance of CAD seems to be better than endoscopist using
EC for predicting neoplastic lesions. Large-scale data analysis in Indian population
is needed prior to community practice.
Supplementary Video S1 Title slide: Real-world experience of artificial intelligence-assisted endocytoscopy
using EndoBRAIN—An observational study from a tertiary care center. Nonneoplastic
polyp showing white-light imaging, narrow band imaging (NBI), endocytoscopy with NBI
and methylene staining. Neoplastic polyp showing white-light imaging, NBI, endocytoscopy
with NBI, and methylene staining.
