Introduction
The development of capsule endoscopy (CE) enhanced examination of the small bowel
[1]. Today, it is commonly used as the initial exam in situations of suspected mid-gastrointestinal
bleeding, after normal upper and lower endoscopy [2]. It is minimally invasive, has a higher diagnostic yield than other noninvasive
methods, and it has proven to be cost-effective in these clinical scenarios [3]
[4]
[5]. Nonetheless, it is time-consuming and error-prone, because crucial frames might
be overlooked, especially if there are just a few of them [6].
Vascular lesions are the most common cause of gastrointestinal bleeding, not only
in the small bowel, but also in other locations [7]. Although the assessment of the small bowel is the primary focus of CE, detection
of upstream or downstream lesions in other areas of the gastrointestinal system may
also be clinically significant. In fact, because the majority of gastrointestinal
bleeding occurs beyond the duodenal ampulla or distal to the ileocecal valve, it might
be considered a second examination of the upper or lower digestive tract, especially
if the initial examination did not yield conclusive results [7].
The introduction of dual-camera capsule has prompted to discussion a CE-based panendoscopic
evaluation of the digestive tract, especially for colorectal cancer screening and
for Crohn’s assessment [8]
[9]. It would be beneficial if CE allowed for complete assessment of the whole digestive
tract, excluding all potential vascular lesions while avoiding repeat exams. Nonetheless,
it is associated with an even greater reading burden which, along with its significant
cost and lack of experience in the majority of centers, may limit its use in clinical
practice, mainly in low-volume ones [9].
Convolutional neural networks (CNNs) have revolutionized image pattern recognition.
This type of deep learning technology was inspired by the neural architecture of the
human cortex and it emulates the neurobiological process of accomplishing complex
tasks combining multiple layers of interconnected neurons [10]. Many articles have been published using this type of artificial intelligence (AI)
system in different image-base procedures, including CE. Currently, there is published
research in the field of automatic detection of vascular lesions during CE, in the
small bowel, which has high overall accuracy [11]
[12]
[13]
[14]
[15]. These algorithms can not only identify different types of vascular lesions (red
spots, angiectasia and/or varices), but also predict their likelihood of bleeding,
according to Saurin classification [14]
[16]. Nonetheless, no studies on panendoscopic assessment of vascular lesions have been
published.
This study aimed to develop and test a CNN-based algorithm for panendoscopic automatic
detection of vascular lesions during CE.
Patients and methods
Study design
A multicenter retrospective cohort study was conducted in two different centers (Centro
Hospitalar Universitário de São João and ManopH Gastroenterology Clinic, both in Porto,
Portugal), which included 1188 CE and colon CE (CCEs) which were performed between
June 2011 and August 2023.
The project was developed without direct intervention on patients; therefore, their
clinical management was not affected. To protect patient identity, identifying information
was omitted and random numbers were allocated to each one. A legal team with Data
Protection Officer (DPO) certification (Maastricht University) ensured data protection,
regarding to its non-traceability as well as compliance with General Data Protection
Regulation.
Capsule endoscopy protocol
Seven different CE devices were used for CE procedures: PillCam COLON (Medtronic Corp.,
Minneapolis, Minnesota, United States), PillCam Crohn's Capsule (Medtronic Corp.,
Minneapolis, Minnesota, United States), PillCam SB1 (Medtronic Corp., Minneapolis,
Minnesota, United States), PillCam SB3 (Medtronic Corp., Minneapolis, Minnesota, United
States), OMOM HD Capsule (JINSHAN Co., Yubei, Chongqing, China), Olympus Endocapsule
10 (Olympus Corp., Tokyo, Japan), and MiroCam (Intromedic Corp., Seoul, South Korea).
PillCam COLON 1, PillCam Crohn’s, PillCam SB1 and PillCam SB3 images were examined
with PillCam Software version 9 (Medtronic, Minneapolis, Minnesota, United States),
while OMOM HD images were reviewed with Vue Smart Software (Jinshan Science & Technology
Co, Chonqing, Yubei, China), Olympus with EC-10 System (Olympus) and MiroCam with
MiroView Software. To protect patient identification, image processing was used to
erase personal information (name, operation number, and procedure date). Each frame
was then labeled with a sequential number.
The European Society of Gastrointestinal Endoscopy recommendations were followed for
bowel preparation [7]. Patients were advised to have a clear liquid diet the day before taking the capsule,
and to fast the night before the examination. Prior to ingestion, patients underwent
bowel preparation, which involved taking 2 L of polyethylene glycol (PEG) solution.
For the PillCam Crohn’s capsule, patients were given 2L of PEG solution the night
before the procedure and another 2L on the morning of the procedure. An anti-foaming
agent, namely simethicone, was used, and if the capsule remained in the stomach for
more than 1 hour after ingestion (which implied image review on the patient’s data
recorder), domperidone 10 mg was given.
Categorization of lesions
The existence of vascular lesions, defined as angiectasias (tortuous and clustered
capillary dilatations, resulting in well-defined brilliant red lesions) or varices
(elevated venous dilatations with serpiginous appearance), was subsequently assessed
in each frame. The images were separated into two groups: those with normal mucosa
and those with vascular lesions. A consensus among three experienced gastroenterologists
in CE was required for the final inclusion of each frame. A total of 152,312 frames,
from seven types of CE devices, were used to develop the CNN, of which 14,942 contained
pleomorphic vascular lesions.
Development of the CNN and performance analysis
We constructed a deep learning CNN to automatically detect vascular lesions, allowing
for panendoscopic assessment of the presence of this lesion throughout the gastrointestinal
system. This was accomplished by a two-step process. First, we used 90% of the dataset
to perform a 5-fold cross-validation, during training and validation, to ascertain
the robustness and assess the global performance of the CNN. Second, the remaining
10% of the dataset was used for testing with the average model resulting from the
five training sessions of the cross-validation. During this phase, the test set was
used to screen eventual discrepancies in the algorithm. The whole process was iterated
five times in total, using different combinations. [Fig. 1] shows a graphical flowchart of the research design.
Fig. 1 Flowchart illustrating the study design. AUC-PR, area under the precision-recall curve;
AUC-ROC, area under the conventional receiver operating characteristic curve; CE,
capsule endoscopy; CCE, colon capsule endoscopy; CNN, convolutional neural network;
N, normal mucosa; NPV, negative predictive value; PPV, positive predictive value;
PV, vascular lesion.
The CNN was built using the RegNetY model [17]. Weights between units were trained using ImageNet, a large-scale image dataset
created for object software recognition. We kept its convolutional layers in order
to transfer its learning to our model. We deleted the last fully connected layers
from our own classifier of dense and dropout layers. Each of the two blocks we utilized
had completely connected layers first, followed by dropout layers with a 0.2 drop
rate. After that, we added a dense layer the size of which was defined the number
of classification groups (two: normal or vascular lesions). Trial and error were used
to determine the learning rate (ranging between 0.0000625 and 0.0005), batch size
(128) and the number of epochs (20). PyTorch and scikit libraries were used to prepare
the model. During training standard data augmentation techniques, such as picture
rotations and mirroring were used. A 2.1 GHz Intel Xeon Gold 6130 processor (Intel,
Santa Clara, California, United States) and double NVIDIA Quadro RTX 80000 graphic
processing unit (NVIDIA Corp, Santa Clara, California, United States) powered the
computer.
The algorithm calculated the probability of each frame being considered normal and
the probability of having a vascular lesion. Each frame was assigned to one of the
aforementioned categories, and the one with the highest probability was selected (Supplementary Fig. 1). We generated heatmaps to identify the features that contributed the most to the
CNN prediction ([Fig. 2]). The algorithm’s final classification was compared with the equivalent evaluation
supplied by the three expert gastroenterologists, with the latter considered the gold
standard.
Fig. 2 Examples of generated heatmaps showing how CNN distinguishes a vascular lesion. 1-esophagus,
2-stomach, 3-small bowel, 4-colon.
Training and validation dataset
First, we performed a 5-fold cross-validation, to assess robustness and assess the
global performance of the CNN. From the total dataset, 90% of data (n=1069 exams)
was divided 5-fold with equivalent dimensions. For this division, we utilized the
StratifiedGroupKFold method, ensuring that images from the same procedure were grouped
together within each fold, while also ensuring that lesions were diversely represented.
We conducted a total of five separate runs. In each of these runs, four folds were
designated to train the model, while the remaining one was used to validate it. The
folds used to train and validate the CNN changed within each run. This process was
iterated a total of five times. [Table 1] lists the number of frames, patients, devices, regions (esophagus, stomach, small
bowel and colon), and pleomorphic vascular lesions contained in each fold. [Table 2] lists the number of exams and corresponding number of frames for each device during
the 5-fold cross-validation experiment and the test set.
Table 1 Number of frames, patients, devices, regions (esophagus, stomach, small bowel and
colon) and pleomorphic vascular (PV) lesions presented within each fold, during the
5-fold cross-validation experiment (five iterations total) and the test set (five
iterations total).
|
|
Frames (n)
|
Patients (n)
|
Devices (n)
|
Regions (n)
|
PV lesions (n)
|
|
PV, pleomorphic vascular lesions.
|
|
Iteration 1
|
Fold 1
|
30889
|
226
|
4
|
4
|
2634
|
|
Fold 2
|
23848
|
222
|
4
|
4
|
5662
|
|
Fold 3
|
33063
|
228
|
5
|
4
|
1834
|
|
Fold 4
|
23324
|
237
|
5
|
4
|
2155
|
|
Fold 5
|
28045
|
156
|
6
|
4
|
1516
|
|
Test set
|
13143
|
119
|
4
|
4
|
1141
|
|
Iteration 2
|
Fold 1
|
23951
|
231
|
4
|
4
|
3898
|
|
Fold 2
|
24024
|
225
|
4
|
4
|
2016
|
|
Fold 3
|
31039
|
214
|
5
|
4
|
3459
|
|
Fold 4
|
40274
|
245
|
3
|
4
|
3243
|
|
Fold 5
|
21617
|
154
|
6
|
4
|
1665
|
|
Test set
|
11407
|
119
|
5
|
4
|
661
|
|
Iteration 3
|
Fold 1
|
29035
|
231
|
5
|
4
|
1684
|
|
Fold 2
|
24753
|
222
|
4
|
4
|
2658
|
|
Fold 3
|
37927
|
233
|
4
|
4
|
4454
|
|
Fold 4
|
17612
|
155
|
4
|
4
|
1428
|
|
Fold 5
|
27551
|
228
|
5
|
4
|
3775
|
|
Test set
|
15434
|
119
|
5
|
4
|
943
|
|
Iteration 4
|
Fold 1
|
40446
|
213
|
5
|
4
|
4490
|
|
Fold 2
|
25285
|
219
|
4
|
4
|
2927
|
|
Fold 3
|
26678
|
201
|
4
|
4
|
1109
|
|
Fold 4
|
21997
|
207
|
4
|
4
|
3054
|
|
Fold 5
|
22657
|
229
|
5
|
4
|
1883
|
|
Test set
|
15249
|
119
|
6
|
4
|
1479
|
|
Iteration 5
|
Fold 1
|
27056
|
226
|
4
|
4
|
2323
|
|
Fold 2
|
32663
|
226
|
6
|
4
|
2506
|
|
Fold 3
|
28752
|
229
|
5
|
4
|
4955
|
|
Fold 4
|
28836
|
234
|
5
|
4
|
3004
|
|
Fold 5
|
18127
|
154
|
4
|
4
|
1511
|
|
Test set
|
16878
|
119
|
4
|
4
|
643
|
Table 2 Number of exams and corresponding number of frames for each device during the 5-fold
cross-validation experiment (five iterations total) and the test set (five iterations
total).
|
|
Fold 1 Exams (frames)
|
Fold 2 Exams (frames)
|
Fold 3 Exams (frames)
|
Fold 4 Exams (frames)
|
Fold 5 Exams (frames)
|
Test set Exams (frames)
|
|
Iteration 1
|
PillCam COLON 1
|
1
(28)
|
5
(1636)
|
4
(206)
|
3
(653)
|
1
(34)
|
2
(128)
|
|
PillCam Crohn’s
|
43
(14818)
|
24
(5557)
|
31
(6577)
|
33
(9154)
|
27
(11508)
|
17
(8056)
|
|
PillCam SB1
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(10)
|
1
(10)
|
0
(0)
|
|
PillCam SB3
|
158
(12825)
|
169
(13721)
|
161
(25080)
|
166
(12576)
|
111
(15075)
|
88
(4834)
|
|
OMON
|
24
(3218)
|
24
(2934)
|
31
(1197)
|
34
(931)
|
15
(1391)
|
12
(125)
|
|
Olympus
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(27)
|
0
(0)
|
|
Mirocam
|
0
(0)
|
0
(0)
|
1
(3)
|
0
(0)
|
0
(0)
|
0
(0)
|
|
Iteration 2
|
PillCam COLON 1
|
3
(319)
|
4
(563)
|
4
(331)
|
0
(0)
|
3
(883)
|
2
(589)
|
|
PillCam Crohn’s
|
30
(9480)
|
41
(10735)
|
30
(8468)
|
41
(15776)
|
27
(10236)
|
6
(975)
|
|
PillCam SB1
|
0
(0)
|
0
(0)
|
1
(10)
|
0
(0)
|
1
(10)
|
0
(0)
|
|
PillCam SB3
|
161
(9618)
|
163
(11404)
|
153
(20322)
|
179
(24006)
|
101
(10242)
|
96
(8519)
|
|
OMON
|
37
(4534)
|
17
(1322)
|
26
(1908)
|
25
(492)
|
21
(219)
|
14
(1321)
|
|
Olympus
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(27)
|
0
(0)
|
|
Mirocam
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(3)
|
|
Iteration 3
|
PillCam COLON 1
|
5
(1146)
|
2
(178)
|
1
(150)
|
3
(56)
|
5
(1155)
|
0
(0)
|
|
PillCam Crohn’s
|
37
(6590)
|
33
(10704)
|
44
(14021)
|
20
(6703)
|
28
(11020)
|
13
(6632)
|
|
PillCam SB1
|
1
(10)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(10)
|
0
(0)
|
|
PillCam SB3
|
160
(19723)
|
160
(12180)
|
162
(21353)
|
107
(9918)
|
170
(12323)
|
94
(8614)
|
|
OMON
|
28
(1566)
|
27
(1691)
|
26
(2403)
|
25
(935)
|
24
(3043)
|
10
(158)
|
|
Olympus
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(27)
|
|
Mirocam
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(3)
|
|
Iteration 4
|
PillCam COLON1
|
1
(588)
|
3
(453)
|
1
(150)
|
6
(1333)
|
4
(152)
|
1
(9)
|
|
PillCam Crohn’s
|
35
(15992)
|
33
(10215)
|
27
(9392)
|
26
(5951)
|
35
(8600)
|
19
(5520)
|
|
PillCam SB1
|
1
(10)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(10)
|
0
(0)
|
|
PillCam SB3
|
154
(19985)
|
153
(13481)
|
149
(14536)
|
155
(14254)
|
158
(13158)
|
84
(8697)
|
|
OMON
|
22
(3871)
|
30
(1136)
|
24
(2600)
|
20
(459)
|
31
(737)
|
13
(993)
|
|
Olympus
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(27)
|
|
Mirocam
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
1
(3)
|
|
Iteration 5
|
PillCam COLON1
|
2
(408)
|
4
(1193)
|
3
(257)
|
4
(696)
|
2
(130)
|
1
(1)
|
|
PillCam Crohn’s
|
44
(14649)
|
30
(10093)
|
38
(13128)
|
30
(7186)
|
21
(7960)
|
12
(2654)
|
|
PillCam SB1
|
0
(0)
|
0
(0)
|
1
(10)
|
1
(10)
|
0
(0)
|
0
(0)
|
|
PillCam SB3
|
163
(11546)
|
161
(20614)
|
158
(10535)
|
169
(17910)
|
110
(9579)
|
92
(13927)
|
|
OMON
|
17
(453)
|
29
(733)
|
29
(4822)
|
30
(3034)
|
21
(458)
|
14
(296)
|
|
Olympus
|
0
(0)
|
1
(27)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
|
Mirocam
|
0
(0)
|
1
(3)
|
0
(0)
|
0
(0)
|
0
(0)
|
0
(0)
|
After completing five iterations of this 5-fold cross-validation, we calculated the
mean sensitivity, specificity, accuracy, positive predictive value (PPV), negative
predictive value (NPV). We also computed the mean area under the conventional receiver
operating characteristic (AUC-ROC) curve and area under the precision-recall curve
(AUC-PR) for each one. We chose to calculate booths (precision-recall calculation,
in addition to conventional ROC curve), due the higher proportion of normal mucosa
frames (true negatives) over frames containing vascular lesions (true positives),
which could lead to a misinterpretation of the ROC curve [18].
Test dataset
In a second step, the testing phase involved the remaining 10% of the dataset (n=119
exams), employing the average model resulting from the five training sessions in the
cross-validation. Frames from a single exam were assigned to either the training/validation
or testing set, ruling out the possibility of their inclusion in both. We repeated
this process in five iterations, with different random combinations. In this phase,
the algorithm was scrutinized for potential discrepancies through the examination
of the independent test set.
During the testing phase, we calculated the mean sensitivity, specificity, accuracy,
NPV and PPV. In addition, the mean AUC-ROC and AUC-PR were also calculated. Furthermore,
we determined the CNN computational performance through the determination of processing
time of all frames within the test set. We performed statistical analysis using scikit-learn
v0.22.2 [19]. All the outcomes derived from this five-iteration process are presented as means
along with their respective 95% confidence intervals.
Results
A total of 1188 exams were included for the development and testing of the CNN, 1097
corresponding to small bowel CE (PillCam SB3, n=941 OMOM HD Capsule, n=152; Olympus
Endocapsule, n=1; PillCam SB1, n=2; MiroCam, n=1), and 210 to devices allowing CCE
(PillCam Crohn’s Capsule, n=192; PillCam COLON, n=18). From these exams, 152312 images
were ultimately validated and incorporated in the dataset, from which 14942 showed
vascular lesions (angiectasias or varices).
Training and validation dataset
[Table 3] shows the results obtained from the five iterations of this five-fold cross-validation
experiment. The mean sensitivity was 87.5% (IC95% 81.5–93.6%) and median specificity
was 99.5% (IC95% 99.3–99.7%). The mean PPV and NPV were 94.9% (IC95% 93.1–96.8%) and
98.6% (IC95% 98.0–99.3%), respectively. Mean global accuracy was 98.4% (IC95% 97.7–99.1
%). The mean AUC-ROC was 0.987 (IC95% 0.980–0.995) while the mean AUC-PR was 0.998
(IC95% 0.997–1.000) ([Fig. 3]).
Table 3 Five-fold cross-validation with exam split (repeated a total of five iterations).
|
|
Sensitivity %
|
Specificity %
|
PPV %
|
NPV %
|
Accuracy %
|
AUC-ROC
|
AUC-PR
|
|
AUC-PR, area under the precision-recall curve; AUC-ROC, area under the conventional
receiver operating characteristic curve; PPV, positive predictive value; NPV, negative
predictive value.
|
|
Iteration 1
|
87.9
|
99.4
|
93.4
|
98.3
|
98.1
|
0.984
|
0.998
|
|
Iteration 2
|
91.1
|
99.5
|
95.8
|
98.9
|
98.5
|
0.990
|
0.998
|
|
Iteration 3
|
93.3
|
99.7
|
96.1
|
99.4
|
99.2
|
0.996
|
1.000
|
|
Iteration 4
|
83.2
|
99.4
|
92.8
|
98.1
|
97.7
|
0.986
|
0.998
|
|
Iteration 5
|
82.0
|
99.7
|
96.2
|
98.4
|
98.2
|
0.980
|
0.998
|
Fig. 3 Representative example of an area under the conventional receiver operating characteristic
curve (AUC-ROC) (1) and an area under the precision-recall curve (AUC-PR) (2) of CNN
performance in detection of vascular lesions in iteration 3 of the training/validation
phase. Precision (on the y axis), also known as positive predictive value, is related
to the proportion of cases in which the CNN algorithm was correct. Recall (on the
x axis), also known as sensitivity, is related to the proportion of frames containing
vascular lesion that were retrieved by the CNN model. A higher precision indicates
a lower false-positive rate, whereas a higher recall means a lower false-negative
rate. The higher the precision and recall, the bigger the AUC-PR.
Test set
The testing dataset comprised an independent group of images (10% of the full dataset).
Supplementary Table 1 displays the metrics of the test set for each of the five iterations performed.
The model’s mean sensitivity and specificity were 72.8% (IC95% 55.8–89.6%) and 99.0%
(IC95% 98.5–99.5%), respectively. The PPV was 83.3% (IC95% 76.1–90.4%), while the
NPV was 97.8% (IC95% 96.0–99.8%). The algorithm’s overall accuracy was 97.0 (IC95%
94.8–99.2%). The AUC-ROC value was 0.984 (IC95% 0.9772–0.991), while AUC-PR value
was 1.000.
There were two main reasons why the model was typically incorrect: existence of large
air bubbles and inadequate cleansing during CE (Supplementary Fig. 2).
The CNN algorithm processed each frame in 26±3 milliseconds.
Discussion
This was the first study to evaluate the application of AI deep learning models in
panendoscopic automatic detection of vascular lesions, not only in the small intestine,
but also in other gastrointestinal topographies. This model not only performed well
in all of the evaluated outcomes, but the results also suggest that it could possibly
be used effectively with various types of CE devices. We believe that these results
are promising and might contribute to implementation of AI-assisted panendoscopic
CE in routine clinical practice, independent of the device brand.
There are a few methodologic details concerning this study that should be highlighted.
Because each exam’s frames were assigned to a single fold in the cross-validation
experiment and to a single dataset (training or testing) in the subsequent phase of
assessing CNN global performance, the risk of overfitting was reduced. When frames
from the same patient are given to both groups, the probability of having similar
images and producing too accurate prediction measures increases. Exam split design
improves the external validity of the results, as well as inclusion of CE exams from
two distinct high-volume centers. In addition, the CNN was developed using frames
from different types of CE devices, including not only one but even two camera capsules,
which may improve its effectiveness in real-world clinical practice. Furthermore,
in the 5-fold cross-validation experiment involving different patients and device
distributions, the model demonstrated excellent median diagnostic performance metrics.
This implies that the CNN performance remains robust, regardless of the type of CE
device employed. The development of a proficient deep learning model with this many
(seven) different brands of CE devices marks a noteworthy achievement which, to the
best of our knowledge, has not previously been documented. Addressing this important
interoperability barrier may increase the technology readiness level (TLR), allowing
for earlier implementation of AI-assisted gastroenterology procedures into routine
clinical practice.
The study has some limitations. First, it was conducted retrospectively, which may
introduce selection bias because the studied sample may not be as representative as
it should be. Second, the study included a relatively small number of frames, mainly
in the test dataset, which could also compromise the external validity of our findings.
To corroborate these results, prospective and multicenter studies are required before
introducing these deep learning models in clinical practice. Third, achieving excellent
performance outcomes with still frames may not guarantee comparable performance with
video segments or full-length videos. Nonetheless, we hypothesize that the algorithm
computational performance, with a reading rate of approximately 38 frames per second,
gives it the capacity to adjust to real-life settings. Although our results look promising,
more studies are needed to determine whether the use of AI models is cost-effective.
Fourth, by comparing the performance metrics obtained from cross-validation during
training/validation with those derived from the testing set, we observed a slight
decrease in sensitivity and PNV in the latter. This discrepancy could be attributed
to various factors. On the one hand, despite our efforts to mitigate overfitting during
training, it cannot be entirely ruled out. On the other hand, differences in representation
between the validation and test sets may also contribute to this variation.
Research on AI and CE is increasing exponentially. However, most studies focus on
the development of deep learning models in automatic identification of a specific
type of lesion in either the small bowel or the colon. In the small bowel, there are
very accurate deep learning models capable of detecting different types of vascular
lesions, as well as predicting their bleeding risk accuracy [14]. In the colon, although the vast majority of retrospective studies focus on the
detection of protruding lesions, there are already published AI algorithms not only
for automatic detection of blood or hematic residues [20]. In addition, there is also a published trinary network aiming to detect and differentiate
blood from normal colonic mucosa and from mucosa lesions (including ulcers and erosions,
vascular lesions and protruding lesions) with high sensitivity, specificity, and accuracy
[21].
Panendoscopic evaluation of the entire gastrointestinal tract is still in the developmental
stage, even though it has a wide range of potential and exponential growth is anticipated
in it. To our knowledge, there are no published papers that reporting on development
of a deep learning algorithm to detect vascular lesions, not only in the small bowel
and colon, but also in the esophagus and stomach, allowing a true panendoscopic evaluation
of the entire digestive tract mucosa. This may be important in clinical practice in
patients who present with overt gastrointestinal bleeding. Our results demonstrated
not only exceptional CNN robustness, but also high global performance levels with
98% overall accuracy, supporting AI use in a live healthcare practice environment.
Conclusions
In conclusion, this was the first proof-of concept AI deep learning model, worldwide,
that was developed and validated for panendoscopic automatic detection of vascular
lesions during CE. The high diagnostic performance of this CNN in multibrand devices
addresses an important issue of technological interoperability, allowing it to be
replicated in multiple technological settings. The enhancement in diagnostic efficiency
of CE provided by AI, combined with increased interest in minimally invasive techniques,
may contribute to increased access to this diagnostic method, thus promoting its performance
when a purely diagnostic endoscopic exploration is expected.
