Introduction
Artificial intelligence (AI) involves computer programs that perform functions normally
attributed to human intelligence, such as learning and problem solving.[1] AI has gradually evolved over the decades since its inception in the 1950s as a
display of intelligent behavior indistinguishable from that of a human being and has
come to incorporate concepts like machine learning (ML) and deep learning (DL). DL,
a relatively new concept, has emerged as a revelation in the realm of computer technology.
A substantial amount of research has been done using DL application in the domain
of image analysis in various medical fields including gastroenterology and hepatology.
AI has been applied in the endoscopic analysis of nonmalignant lesions such as polyps,
ulcers, lymphangiectasia, angiectasia, etc. In addition to cancer detection and analysis
of inflammatory lesions or localizing the lesion in obscure bleeding during wireless
capsule endoscopy (WCE), AI has been used to good effect in diagnosis making and prognostication.
The application of AI has waxed and waned over the past six decades with seemingly
little improvement but various ML-based models like support vector machine (SVM),
artificial neural network (ANN), and convolutional neural network (CNN) have proven
to be useful in different branches of medicine with outstanding performance in image
recognition and analysis. However, the optimal performance of DL-based methods requires
a huge amount of properly labeled training data. This issue has been addressed by
combining DL methods with reinforcement learning principles.[2] A graphical summary of the concepts of AI, ML, and DL is depicted in [Fig. 1].
Fig. 1 Summary of the development of artificial intelligence, machine learning, and deep
learning.
Nevertheless, AI and its newer models are not fool-proof or free from errors. The
major drawbacks with newer techniques are overfitting and a lack of explainability.
While the DL-based models perform much better than other models, these are intrinsically
dependent on the training dataset. The lack of diversity and the presence of unidentified
bias in the training dataset may hinder generalization to real-life situations and
may lead to problems in proper model validation. Moreover, lack of explainability
(black-box nature) is a major concern in AI-based models.
Until now, most of the studies have stressed on improving the explainability of AI-based
models.[3] In the field of luminal gastroenterology, most reviews related to AI have focused
on AI-assisted endoscopy, either in the form of automatic polyp detection in a colonoscopy[4] or malignant lesion detection in the stomach and esophagus.[5] In contrast, studies on AI-assisted endoscopy in diagnosing or localizing inflammatory,
malignant, or bleeding lesions in the small bowel have been sparse. Studies on WCE
have dealt only with the technical improvements of AI models in lesion detection,
classification, and characterization. Surprisingly, data on the use of AI on single/double-balloon
enteroscopy (DBE) are lacking. Furthermore, real-time applications of AI on any form
of small bowel enteroscopy are lacking.
We conducted an extensive search in the internet using Google Scholar, Google search
engine, and PubMed database on most recent and relevant articles using keywords: AI
and small bowel endoscopy, computer-aided diagnosis (CAD), and capsule endoscopy,
computer-based diagnosis in small bowel diseases, ML in small bowel endoscopies, and
application of AI techniques in small bowel endoscopy. No exclusions were made on
study designs and all articles were in English. In this present review, we have tried
to summarize the different modalities of application of various AI models in diagnosing
small bowel diseases with special emphasis on WCE and have dwelt on the pros and cons
of such applications including their prospects.
Small Bowel Endoscopy and AI Application
WCE and balloon (single or double) enteroscopy have revolutionized the field of small gut imaging. It is no secret that the invention
of WCE by Gavriel Iddan et al[6] has brought in a sea change in the management of small-bowel diseases, including
occult gastrointestinal (GI) bleeding, Crohn’s disease, polyposis syndromes, malignancy,
and celiac disease. Although its use is limited to the nonobstructed bowel, the advent
of balloon enteroscopy (BE) to “chase” the findings of WCE has ushered in a new revolution
in the realm of therapeutic interventions in the small bowel.
Advancements like suspected blood indicator,[7] adaptive frame rate technology,[8] and the quick-view algorithm[9] based on AI for CAD have been developed to reduce the long review time and increase
the accuracy in diagnosis. However, it is important to keep a note of the fact that
these have yielded mixed results in terms of diagnosis and have reported high missing
rates too.[10]
Current Use of WCE in Practice
Analysis of Malignant and Premalignant Lesions
As mentioned above, there have been only a few studies that have focused on small
gut malignancies in particular.[11]
[12]
[13]
[14] Technical variations on detection and characterization of polyps in WCE images have
been tested in patients with suspected or previously known polyposis syndromes such
as familial adenomatous polyposis and Peutz-Jeghers syndrome.[15]
[16] WCE has also been found to be beneficial in the detection of polyps in the jejunum
and ileum.[15] WCE is found comparable to magnetic resonance enterography (MR enterography)[17] whereas computed tomography (CT) enterography and DBE have shown a superior sensitivity
compared with WCE in the detection of small bowel tumors.[18]
[19]
Therefore, the presence of multiple constraints like lack of accuracy in image reading
by an observer, lengthy capsule processing, and concerns regarding the image quality
in WCE procedures coupled with the fact that the importance and applicability of AI
have been steadily increasing led scientists to devise models that would be swift
and would be less erroneous in detecting polyps and tumors in the small intestine.
This, in turn, led researchers to focus primarily on domains like automatic polyp
detection and characterization in the GI tract. Detection and characterization of
a polyp are done using features (color, shape, edge, and texture) with various AI
classifiers with a sensitivity of approximately 95% and accuracy varying from 91 to
98%. The different methods of AI to detect and characterize polyp in WCE images/videos
are summarized in [Table 1].
Table 1
Summaries of studies on polyp detection (overall) and tumor localization/characterization
(in small bowel) involving AI
Study
|
Disease/localization
|
Design
|
Feature/technique
|
Classifier
|
Image/videos
|
Accuracy
|
Sensitivity
|
Abbreviations: AI, artificial intelligence; ANN, artificial neural network; CV, cross
validation; HMM, hidden Markov model; KNN, k-nearest neighbors; LBP, local binary
patterns; MLP, multilayer perception network; RS, retrospective study; SB, small bowel;
SVM, support vector machine; WCE, wireless capsule endoscopy.
|
Nawarathna et al[14]
|
Polyp(LB)
|
RS
|
Texton histogram
|
KNN, SVM
|
400
|
95.27%
|
–
|
Zhao et al[20]
|
Polyp(LB)
|
RS
|
HMM
|
Boosted SVM
|
1,200
|
90%
|
–
|
Li et al[21]
|
Polyp(LB)
|
RS
|
Uniform LBP + wavelet transform
|
SVM
|
1,200
|
91.6%
|
–
|
Condessa et al[22]
|
Polyp(LB)
|
RS
|
Local polynomial approximation
|
SVM
|
3 videos
|
–
|
92.31%
|
Constantinescu et al[11]
|
Polyp (SB)
|
RS
|
WCE (SB)
|
ANN
|
54 videos/90 images
|
97.7%
|
93.8%
|
Li et al[12]
|
Tumor (SB)
|
RS
|
WCE (SB)
|
KNN, MLP
|
900/300 images
|
90.5%
|
92.3%
|
Dinevari et al[23]
|
Tumor (SB)
|
RS
|
WCE (SB)
|
SVM
|
600/200 images
|
93.5%
|
94.04%
|
Liu et al[13]
|
Tumor (SB)
|
RS
|
WCE (SB)
|
SVM
|
1,800 images
|
97.3%
|
97.8%
|
A few studies have also highlighted the methods of detecting polyp/malignant lesions
in WCE images/videos concerning the small bowel using only various AI-based models,
viz., ANN, KNN (K-nearest neighbors), multilayer perception network (MLP), and SVM[11]
[12]
[13]
[14] with a reported sensitivity of 92 to 98% and a diagnostic accuracy of 92 to 97%.
The results of the AI studies involving small gut malignancies are summarized in [Table 1].
Inflammatory and Other Nonmalignant Lesions
Crohn’s Disease
WCE is useful in the evaluation of Crohn’s disease in the small intestine in cases
where there is a diagnostic dilemma. Many studies have established the role of WCE
as a valuable adjunct to conventional endoscopy and colonoscopy with ileoscopy with
a reported sensitivity and specificity of 89 to 93% and 84 to 100%, respectively.[20]
[21] WCE has also been shown to be superior to CT enteroclysis[22] and MR enterography,[23] especially in terminal ileal disease and proximal small-bowel disease.[22] Overall, most studies suggest a superior sensitivity of WCE with varying degrees
of specificity over other radiological tests in the detection of small bowel Crohn’s
disease. It is, however, pertinent to remember that a lack of validated capsule criteria
and the inability to obtain biopsy specimens for confirmation of diagnosis are significant
limitations in the diagnosis of Crohn’s disease.[24] While DBE was found to be superior to WCE,[25] false-positive results in a few asymptomatic patients raise concerns regarding accuracy
in Crohn’s disease.[26] Further, severity scales for Crohn’s disease using WCE: the Lewis score and the
capsule endoscopy Crohn’s disease activity index have also been developed which are
undergoing validation and these may prove useful in diagnosing Crohn’s disease of
the small bowel.[27]
[28]
To overcome the various limitations of WCE, researchers have tried to develop and
modify AI-based models with considerable success. Various methods to determine features
(color, edge, and texture), viz., mean shift algorithm[29] and local binary pattern[30] have been used to characterize inflammation in Crohn’s disease. ML technique has
also been used successfully in risk assessment of Crohn’s disease and ulcerative colitis.
Overall sensitivity and accuracy reported in the above studies are to the tune of
71 to 95% and 80 to 94%, respectively.[29]
[30]
[31]
[32]Studies showing the application of AI in patients of small gut IBD are summarized
in [Table 2].
Table 2
Summaries of studies on Crohn’s disease and Celiac disease in small bowel involving
AI
Study
|
Disease/localization
|
Design
|
Diagnostic modality
|
AI classifier
|
Videos/Images in training/testing
|
Accuracy
|
Sensitivity/specificity
|
Abbreviations: AI, artificial intelligence; ANN, artificial neural network; BI, Bayesian
inference; CD, Celiac disease; CV, cross-validation; EMR, electronic medical records;
IBD, inflammatory bowel disease; KNN, K nearest neighbor; MNN, Multilayer neural network;
NBI, narrow band imaging; RS, retrospective study; SB, small bowel; SVM, support vector
machine; SVM, support vector machines; WCE, wireless capsule endoscopy; WL, white
light.
|
Girgis et al[29]
|
Crohn’s disease
|
RS
|
WCE (SB)
|
SVM
|
467/277 images
|
87
|
80/93
|
Kumar et al[30]
|
Crohn’s disease
|
RS
|
WCE (SB)
|
SVM
|
533 images
|
80.2
|
81.1/93.6
|
Wei et al[31]
|
Crohn’s disease
|
RS
|
Genetics
|
SVM
|
53,279/22,442
|
AUROC-0.86
|
71/83
|
Charisis and Hadjileontiadis[32]
|
Crohn’s disease
|
RS
|
WCE (SB)
|
SVM
|
800/102 images
|
93.8
|
95.2/92.4
|
Ciaccio et al[35]
|
Celiac disease
|
RS
|
WCE(SB)
|
SVM
|
8,600/10,000 images
|
76.7
|
88/80
|
Tenório et al[36]
|
Celiac disease
|
RS
|
EMR
|
BI, KNN
|
178/38 images
|
80
|
78.8/80
|
Gadermayr et al[37]
|
Celiac disease
|
RS
|
WL/NBI
|
SVM
|
2,835 images
|
99.6
|
NA
|
Zhou et al[38]
|
Celiac disease
|
RS
|
EMR
|
GoogLeNet
|
8,800/8000 images
|
NA
|
100/100
|
Chen and Lee[41]
|
Ulcer
|
RS
|
WCE(SB)
|
SVM
|
250/930 images
|
96.3
|
91.7/99.4
|
Charisis et al[40]
|
Ulcer
|
RS
|
WCE(SB)
|
SVM, MNN
|
156/18 images
|
95
|
96.6/93.5
|
Eid et al[42]
|
Ulcer
|
RS
|
WCE(SB)
|
SVM
|
260 images
|
86.5
|
84.5/88.6
|
Yuan et al[43]
|
Ulcer
|
RS
|
WCE(SB)
|
SVM
|
271/68 images
|
92.7
|
94.1/91.2
|
Celiac Disease
Celiac disease (CD), with a worldwide incidence of 1%, manifests as loss/scalloping
of duodenal folds with nonspecific mucosal lesions (fissures, crevices, grooves, micro
nodules, or a mosaic pattern) in the small intestine.[33]
Duodenal biopsies using standard endoscopy together with serological testing have
been the cornerstone of diagnosis in CD.[34] The invasiveness of endoscopic biopsy and the expense of serological tests have
resulted in the search for alternative economical, feasible, and noninvasive methods.
Duodenoscopy, despite being convenient for inspecting and assessing villous atrophy
in the duodenum, has significant limitations that it cannot examine the extent and
severity of the disease. In this context, WCE may well be a suitable noninvasive,
patient-friendly, and feasible alternative, which could visualize the entire small
bowel for a detailed evaluation of the mucosal villous architecture with sufficient
resolution in patients of suspected CD.[35] Overall sensitivity and accuracy of 87 to 89% and 97%, respectively are reported
for the diagnosis of CD using WCE.
Over the last decade, AI has been evolving in diagnosing and classifying disease severity
in CD using WCE. The features (color, texture, and shape) of a lesion are being utilized
for diagnosis and assessment of extent and severity of CD.[35] A web-based clinical decision support system that uses AI techniques to diagnose
CD cases has also been reported.[36] A few researchers assessed a hybrid approach that incorporated expertise and technical
knowledge into the computer-based classification, which showed a very high accuracy
in diagnosing CD.[37] A 22-layered deep CNN named GoogLeNet achieved 100% sensitivity and specificity
in diagnosing CD from WCE clips.[38] However, another group of researchers who built a series of predictive models to
diagnose CD utilizing a variety of statistical approaches met with unsatisfactory
results, yielding poor discriminatory performance with AUCs ranging from 0.49 to 0.53.
Overall sensitivity and accuracy reported in the above studies are 78 to 100% and
76 to 99%, respectively.[35]
[36]
[37]
[38]
[Table 3] summarizes the results of studies on CD using AI.
Table 3
Summaries of studies on identification of non-inflammatory lesions and obscure gastrointestinal
bleed in small bowel involving AI
Study
|
Disease/localization
|
Design
|
Diagnostic modality
|
AI Classifier
|
Videos/Images in training/testing
|
Accuracy
|
Sensitivity/specificity
|
Abbreviations: AI, artificial intelligence; CNN, convolutional neural network; GI,
gastrointestinal; MLP, multilayer perceptron network; MNN, multilayer neural network;
RS, retrospective study; SB, small bowel; SSMD, single shot multibox detector; SVM,
support vector machines; WCE, wireless capsule endoscopy.
|
Cui et al[45]
|
Lymphangiectasia
|
RS
|
WCE(SB)
|
Threshold SVM
|
7,218 images
|
97.9
|
48.8/NA
|
He et al[47]
|
Hookworm
|
RS
|
WCE(SB)
|
CNN
|
20,000–30,000
|
88.5
|
84.6/88.6
|
Wu et al[48]
|
Hookworm
|
RS
|
WCE(SB)
|
Rusboost SVM
|
401,476/40,148 images
|
78.2
|
77.2/77.9
|
Li and Meng[56]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
MLP
|
2,700/900
|
NA
|
87.8/88.6
|
Pan et al[57]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
PNN
|
14,630 images/150 videos
|
87.4
|
93.1/85.8
|
Lv et al[58]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
SVM
|
280/280
|
97.9
|
97.8/98
|
Fu et al[59]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
SVM
|
30,000/30,000 pixels
|
94
|
97/92
|
Sainju et al[60]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
MLP
|
100 images
|
93
|
96/90
|
Hassan and Haque[61]
|
Obscure GI Bleed
|
RS
|
WCE(SB)
|
SVM
|
1,200/1,720 images
|
99.2
|
99.4/99
|
Leenhardt et al[63]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
CNN
|
600/600 images
|
98
|
100/96
|
Tsuboi et al[64]
|
Obscure GI bleed
|
RS
|
WCE(SB)
|
CNN
|
2,237/10488 images
|
NA
|
98.8/98.4
|
Ulcer
A vast majority of WCE-related literature is concerned with the reduction of the examination
time of WCE data in the detection of certain disorders in the small gut. However,
there have been very few studies that have dealt with the detection of ulcers (7%)
and Crohn’s lesions (2%).[39] Detection of such lesions is difficult owing to the inherent challenges like nonspecific
characteristics of such lesions and the huge diversity in appearance. Time, expertise,
and feasibility also remain a matter of concern for definitive identification and
localization. Various methods have been utilized to extract the color and texture
of ulcers while a few studies have focused on the salient region identification. Color-texture
extraction using segmentation scheme,[40] ulcer salient map redefined with Gabor filter,[41] texture only extraction method based on discrete curvelet transform,[42] and saliency map using super-pixel region[43] are some of the techniques that have been successfully used. These methods have
been reported to achieve a sensitivity of 84 to 97% while the overall diagnostic accuracy
remains 86 to 96%. The results of the AI-based studies on ulcers are summarized in
[Table 2].
Other Non-Inflammatory and Nonmalignant Lesions
Lymphangiectasia
Lymphangiectasia is the pathologic dilation of lymphatic channels. When it occurs
in the intestines, it is known as intestinal lymphangiectasia, also simply called
lymphangiectasis. WCE is of use in the detection of these lesions in patients presenting
with features of protein-losing enteropathy or chronic malnutrition.[44] Here again, the variability in color, shape, and textural characteristics of lymphangiectatic
lesions often makes it extremely difficult to characterize the lesion using WCE or
BE, thus making the role of AI even more important.[45] Algorithms using luminance information in hue, saturation, and intensity colors
space and Commission Internationale de l’éclairage-Laboratory representation have
reported a sensitivity and accuracy of approximately 48 and approximately 98%, respectively
in the detection of lymphangiectatic lesions in the small bowel.
Hookworm Infestations
Intestinal hookworms are difficult to find with direct visualization because of their
small tubular structures and semitransparent features, which make it tough to distinguish
them from the intestinal mucosa. Moreover, the presence of intestinal secretions makes
detection even more difficult. The role of WCE in detecting hookworms in the small
intestine has been mentioned in the literature, albeit with highly variable detection
rates.[46] Interestingly, AI methods have been utilized to detect hookworms in the small bowel
with a sensitivity of 77 to 84% and accuracy of 78 to 88%. The flip side of this is
the reportedly high missing rates of around 23%.[47]
[48]
[Table 3] summarizes the results of the studies on noninflammatory lesions of the small gut
using AI-based methods.
Obscure GI Bleed
Obscure gastrointestinal bleeding (OGIB) is defined as the bleeding from the digestive
tract, which recurs or persists after a negative initial evaluation, using both upper
and lower GI endoscopy and a negative result on radiologic imaging using small bowel
follow-through or enteroclysis.[49] Accounting for approximately 5% of overall GI bleeding, OGIB bleed has been shown
to arise from the small bowel distal to the Ampulla of Vater and proximal to the ileocecal
valve in more than 80% of the cases, rendering it relatively inaccessible to traditional
endoscopy.[49] While the efficacy and reliability of WCE have been tested over many years, the
detection rate is variable ranging from 35 to 77%.[50]
[51]
Various studies comparing WCE to other methods in unraveling the causes of OGIB have
shown it to be superior to the other investigations. WCE performed better than CT
angiography,[52] CT-enteroclysis, intraoperative enteroscopy,[53] and push enteroscopy[54] in the detection of the lesion. However, the detection rate is comparable to DBE.[55]
Despite the obvious advantages of WCEs over the traditional techniques, issues like
exhaustive evaluation of images and interobserver/intraobserver variability including
subjective humane error have plagued this novel technique. In a bid to overcome these
problems, the application of different AI methodologies has been tried with variable
success rates. Combining color and spatial information of bleeding lesions, viz.,
chromaticity moment ([HIS Color space],[56] [RGB Color space][57]) and color descriptors (pyramid of color invariant histograms, pyramid of hue histograms
, and pyramid of transformed color histograms),[58] researchers have applied AI models to detect and localize bleeding lesions in the
small bowel. Besides, a few studies have also explored the use of textural features
such as pixel segmentation and pixel grouping[59] while a few others have utilized statistical features like a color histogram.[60] A novel algorithm that operates on normalized gray level co-occurrence matrix using
the frequency spectrum of WCE images has also successfully been used.[61] Further, AI methods like CNN, MLP, SVM, and PNN have also been used to classify
the above image patterns for lesion detection and localization with a sensitivity
of 87 to 100%, specificity of 85 to 99%, and a diagnostic accuracy ranging from 87
to 99%.
Small-bowel angiectasia comprises the majority of small-bowel vascular lesions and
is diagnosed in 30 to 40% of OGIB cases.[62] Although detection of angiectasia using WCE is well established, computer-aided
detection methods have not been validated as yet. CNN algorithm using still frames
featuring annotated angiectasia[63] and CNN algorithm based on single shot multibox detector using WCE images of angiectasia[64] have been applied with a reported sensitivity of 98 to 100%, specificity of 96 to
98%, and accuracy of 98%. [Table 3] summarizes the results of AI-based studies on small gut for the evaluation of OGIB.
Challenges in AI and Future Directions
Soon, AI is envisaged to play a major role in helping establish diagnoses, devising
treatment protocols, and in the prediction of treatment outcomes. Over the last seven
decades, AI has been studied extensively and important developments have been made
which show promising results. However, a major drawback of all these endeavors is
the retrospective nature of most of these studies which have heavily banked upon data
chosen from specific endoscopic modalities limited to a fewer number of institutions.
In such a situation, there is a high likelihood of selection bias creeping in, therefore,
it is crucial to validate the performance of AI using different population-based models
in a “real world” setting. Overfitting and spectrum bias have also been observed to
impact negatively on AI performance in terms of reproducibility. Overfitting occurs
when the learning model is dependent too much on a training dataset, resulting in
unsuitable generalization to newer dataset leading to overfitting.[65] There have indeed been attempts to find a way out to solve this problem but with
limited success. Besides, datasets in case–control design studies are readily vulnerable
to spectrum bias. Spectrum bias occurs when the training dataset does not adequately
represent the target population.[66] Because overfitting and spectrum bias may overestimate the accuracy of a model,
external validation of unused datasets is mandatory. Additionally, robust clinical
verification, as well as properly designed multicenter prospective studies with adequate
criteria (inclusion/exclusion) representing the target population, is required. Furthermore,
a lack of interpretability or explainability (black box nature) is a major concern
in AI technology where the decision-making mechanism of AI models may not be clearly
understood and corrected if needed. Some techniques have been developed to address
“black box” limitations such as the attention mapping and saliency region identification;
these, however, require further studies regarding their applicability.[67] As the accuracy and efficiency of ML model is proportional to the input data, developing
an efficient ML model is challenging due to the paucity of human-labeled data. Data
augmentation strategies have been proposed to address this problem.[68] Of note Spiking neural networks, which closely mimic the biological mechanisms of
neurons, can potentially replace the present ANN models, bringing in higher and more
sophisticated computational ability.[69]
The diagnostic precision of AI does not always reflect the exact picture in real clinical
practice. The actual benefit in terms of clinical outcome, viz., physician’s satisfaction,
cost-effectiveness, etc. must be proven by appropriate methods. AI-based models having
either inaccuracies or those that deliver results divorced from clinical reality are
likely to cause ethical issues owing to misdiagnosis or misclassification. Thus, the
impact of AI application on the traditional doctor–patient relationship, which is
the essence of the practice of medicine, should be looked into carefully. Ethical
principles relevant to AI-based models (akin to Asimov’s laws of robotics!) might
need to be developed to tackle problems concerning medical ethics and AI. Finally,
the formulation of reasonable regulatory guidelines and devising a proper reimbursement
policy keeping in mind the economic aspects of health care are essential before integrating
AI technology into the current health care structure. It is to be remembered that
AI is not perfect and error-free. That is why the concept of “augmented intelligence”
has emerged that emphasizes on improving and enhancing human intelligence rather than
replacing it. The future challenges in the field of AI-based technology are described
in [Table 4].
Table 4
Future challenges in artificial intelligence
Abbreviation: AI, artificial intelligence.
|
Variations in performance levels:
-
Great heterogeneity and lack of high-quality datasets.
-
Wide variety of performance metrics (sensitivity, specificity, and accuracy).
-
Lack of proper validation techniques in multiple studies.
|
Lack of randomized controlled trials (RCT) comparing AI vs. non-AI based approaches:
|
Limitations of AI techniques that require further investigation:
|