Keywords Endoscopy Upper GI Tract - Precancerous conditions & cancerous lesions (displasia
and cancer) stomach - Endoscopic resection (ESD, EMRc, ...) - Endoscopy Lower GI Tract
- Endoscopic resection (polypectomy, ESD, EMRc, ...)
Introduction
Endoscopic submucosal dissection (ESD) is a modern minimally invasive technique for
resection of large superficial neoplasia of the gastrointestinal tract [1 ]. The method is technically challenging and requires a long training period.
Complications such as bleeding and perforation are caused by inadvertent dissection
through
submucosal blood vessels and the muscularis propria, especially when visualization
of
anatomical structures and exposure of the submucosal cutting plane are insufficient.
The need
for trainee support in the form of observation, expert supervision as well as animal
training
is promoted in international consensus statements [2 ]. Artificial Intelligence-assisted clinical decision support solutions (AI-CDSS)
for
colonic polyp detection have been certified by international authorities [3 ]. detection and delineation of neoplasia in the upper gastrointestinal tract is the
focus of current research [4 ]
[5 ]
[6 ]. The expected role of AI in endoscopy is predominantly one of standardization and
quality improvement of diagnostic procedures [7 ]. However, research data on AI support during surgery [8 ] show that interventions may also benefit from this new technology. Preliminary
research into structure identification during third-space endoscopy showed promising
results
[9 ]. Therefore, we aimed to develop an algorithm for detection and delineation of relevant
anatomical structures during ESD and subsequent testing in a real-life setting.
Methods
Objectives and study design
The primary endpoint of this study was feasibility of application of a newly developed
algorithm for intraprocedure structure identification during porcine ESD procedures.
The
design was a one-arm exploratory study. Feasibility was defined as continuous technically
successful application during resection. Secondary endpoints included the user experience,
Dice scores, and pixel accuracies of the algorithm in an internal and external validation,
as well as lesion and procedure characteristics during the porcine ESDs. User experience
was
measured in a binary way (appropriate, score 1, not appropriate, score 0).
Algorithm development
Thirty full-length videos of ESD procedures (2 esophageal, 26 gastric, 2 colorectal)
were extracted from the database of the Prince of Wales Hospital, Hong Kong, SAR.
Training
data were generated using GIF-HQ290 gastrosocopes and EVIS Lucera elite processors
(Olympus,
Tokyo, Japan). Subsequently 1011 frames were selected for training. The categories
“submucosal layer,” “muscle layer,” “submucosal blood vessel,” and “background” were
delineated in each training image by an expert in ESD (> 500 procedures in human ESD).
This data set was used to training a deep learning algorithm using a combination of
a U-Net
and transformer network. The model employed a hybrid CNN-transformer architecture
to
leverage both detailed localization information and global context. The model was
trained
and evaluated on an HP Z4 G4 workstation powered by an Intel(R) Xeon(R) W-2223 CPU
operating
at 3.60 GHz, featuring four physical cores and supporting hyper-threading to provide
eight
logical processors, in conjunction with a single Nvidia 3090 GPU equipped with 24
GB of
video random access memory. All implementations were carried out using PyTorch 1.13.1
with
CUDA 11.7 and Python 3.7. For pre-training the base model, the Adam optimizer with
an
initial learning rate of 1e-4 was employed, training the model for 10 epochs with
a batch
size of four. During the training process, 20% of the dataset was allocated as a dedicated
test set, while the remaining 80% was split further into 80% for training and 20%
for
validation. Before processing by the model, video data are preprocessed by resizing
to 256 ×
256 resolution and normalized to zero mean and unit variance to ensure consistency
across
input features. The resulting algorithm was evaluated for Dice score and pixel accuracy
in
an internal cross-validation, as well as an external validation (220 images, 4 ESD
cases, 2
esophageal, 1 gastric, 1 colorectal) obtained from Augsburg University Hospital, Augsburg,
Germany. Test data were generated using GIF-EZ1500 gastroscopes with EVIS X1 processors
(Olympus, Tokyo, Japan).
Porcine model study
An in vivo porcine model was then constructed using two live pigs. Herein, one expert
and one novice in ESD (no experience in human ESD) performed esophageal, gastric,
and rectal ESDs with real-time support from the AI system. For this purpose, a real-time
endoscopic image was transferred to an AI station with a separate monitor equal in
size and resolution to the conventional monitor, and the endoscopic image was shown
continuously with real-time segmentation of the algorithm ([Fig. 1 ], [Video 1 ]). Standard ESDs were performed for 12 lesions with a target size of 2 cm without
use of traction techniques. In all procedures, DualKnife J (Olympus Medical Corporations,
Tokyo, Japan) was used. Procedure times, en bloc resection rates, and intraprocedure
complications such as profuse bleeding and transmural perforation were recorded. Bleeding
was graded according to severity (Grade 1: Relevant oozing bleeding that requires
the endoscopist to intervene, hemostasis successful by coagulation with the tip of
the knife; Grade 2: Relevant oozing or spurting bleeding that rapidly covers the resection
plane and causes impaired view, hemostasis successful by coagulation with the tip
of the knife and/or hemostatic forceps and/or clipping; Grade 3: Fulminant bleeding
causing termination and failure of the procedure or requiring hemostatic measures
for more than 5 minutes). Perforation was defined as a full-thickness perforation
with visibility of underlying tissue outside of the muscle layer such as the serosa,
adventitia, peritoneum or mediastinum. Ninety-five percent confidence intervals (CIs)
are shown in brackets. Institutional review board approval for the study was obtained
from the Animal Experimentation Ethics Committee (AEEC), the Chinese University of
Hong Kong (Ref. No. 22–145-MIS).
Fig. 1 Images from the animal trial. Far left: Two endoscopic images during submucosal dissection.
Left: The same images with AI prediction in a semitransparent colored overlay. Mucosa
(background) is in pink, the submucosal layer is in yellow, the muscle layer is in
green, and the blood vessels are in red. Right: Photo of the trial setup with the
regular endoscopic screen on the right and the screen showing the AI picture in real
time on the left.
Demonstration video of an ESD procedure with (left side) and without (right side)
AI overlay. The green semitransparent overlay indicates the muscle layer. The yellow
border tracing indicates the submucosal layer. The red semitransparent overlay indicates
the submucosal blood vessels.Video 1
Results
The AI system was continuously and successfully applied during all 12 ESD procedures
([Table 1 ]). Feedback from both endoscopists showed an appropriate function of the algorithm
for each ESD procedure by both interventionists. Internal cross-validation showed
mean Dice scores of 91% (95% CI 90%-93%), 77% (95% CI 74%-80%), 60% (95% CI 54%-67%),
82% (95% CI 78–85), and mean pixel accuracies of 9% (95% CI 94%-96%), 73% (95% CI
70%-76), 60% (95%CI 53%-66%), 82% (95% CI 79%-86%) for background, submucosal layer,
submucosal blood vessels, and muscle layer, respectively. On the external dataset,
Dice scores of 90% (95% CI 89%-91%), 68% (95% CI 65%-71%), 56% (95% CI 49%-62%), and
92% (95% CI 89%-96%) and pixel accuracies of 95% (95% CI 94%-96%), 63% (95% CI 61%-67%,
55% (95% CI 49%-62%), and 92% (95% CI 89%-95%) were calculated for the same classes.
The algorithm generated real-time annotation of the trained anatomical structures
with a latency of 0.04 seconds at a frame rate of 25 frames per second. All ESDs were
successfully performed with a 100% en bloc resection rate and no major intraprocedure
complications. Mean resection time per specimen size was 14.8 minutes (95% CI 9.8–19.8).
Bleeding occurred 4.8 times per ESD on average (95% CI 3.4–6.1) (grade 1: 4.2, 95%
CI 3.0–5.3); grade 2: 0.6, 95% CI 0.2–1.0); grade 3: 0 (95%CI 0–0). The rate of bleeding
per specimen size was 2.0 per cm (95% CI 1.1–2.8). There were no full-thickness perforations.
Table 1 Results from the in vivo porcine model AI application trial.
Expert
Novice
Overall
Resection characteristics
2
2
4
3
2
5
2
1
3
7
5
12
33
48
39
Feasibility test
100
100
100
7
5
12
Specimens
7
5
12
0
0
0
2.9
2.6
2.8
Complications
4.9
4.6
4.8
4.3
4.0
4.2
0.6
0.6
0.6
0
0
0
0
0
0
Discussion
In this preliminary study, a novel AI algorithm for real-time delineation of anatomical
structures during ESD was developed. Technical feasibility for usage in a porcine
model could be demonstrated for the expert and novice endoscopist in a real-life setting.
Thus, the primary goal of this study was achieved, that is, proof of concept of real-time
application of an AI-CDSS for ESD. User experience evaluation showed appropriate usability
of the algorithm. The in vivo porcine model is considered the most appropriate experimental
model for ESD procedures [10 ]. Results from the internal validation were deemed satisfactory, whereas reduced
performance for certain classes was measured in external validation. The decrease
in Dice scores and pixel accuracies for the classes submucosa and submucosal blood
vessel on the external dataset show the as-yet-limited robustness of the algorithm
when confronted with images from different endoscopy systems. To reduce possible overfitting,
different image modalities from different endoscopes will be included in the training
data for future versions of the algorithm. Feasibility testing was conducted with
a limited number of animal procedures and interventionists. However, in this early
stage of development of a new technology, a test with a limited number of animal subjects
was deemed ethically appropriate, while still large enough for a meaningful result.
Further development of the algorithm with more training data as well as more concise
preclinical evaluation of performance should precede human trials.
Conclusions
In conclusion, although real-time decision support during ESD seems viable and may
offer valuable assistance especially for trainees, confirmatory studies are needed
in order to corroborate these hypotheses.
Bibliographical Record
