Role of artificial intelligence in the detection and characterization of gastrointestinal premalignant and early malignant lesions

Abstract

Artificial intelligence (AI) is revolutionizing the field of gastrointestinal (GI) endoscopy, a technology that relies heavily on images and optical data. Precancerous lesions and early cancers of the GI tract can be subtle and easily missed even on high-definition endoscopy and chromoendoscopy. The advancements in machine learning and deep learning led to the development of computer-aided models of high performance in image analysis. The convolutional neural networks of these models are trained to analyze large datasets of endoscopic images through the supervised learning approach. Their utilization enhances lesion detection and visibility. This aids in real-time classification and risk stratification of GI luminal lesions, thus assisting endoscopists in making more accurate and timely decisions. AI has shown promising results in the detection and characterization of premalignant and early malignant lesions of the GI tract, such as Barrett’s esophagus, gastric atrophy, intestinal metaplasia, small bowel and colonic polyps, as well as early esophageal, gastric and colon cancers. This positive impact of AI is more established in the esophagus and stomach than in the colon. However, the impact of AI on patients’ outcomes such as mortality and interval cancer incidence remains to be seen. This review highlights the breakthroughs and clinical applications of AI in the detection and characterization of premalignant lesions and early cancers of the GI tract.

Key Words: Artificial intelligence; Premalignant; Early cancer; Neural networks; Polyps; Barrett’s esophagus; Colon cancer; Video capsule

Core Tip: Artificial intelligence (AI) is revolutionizing healthcare. There is evidence supporting the role of AI in the detection/characterization of premalignant lesions and early cancers of the gastrointestinal (GI) tract, especially in the esophagus, stomach and colon. The utilization of AI may eventually decrease the incidence of interval cancers which are mainly attributed to missed lesions during endoscopy. It benefits mostly novice endoscopists and trainees. AI may be superior to experienced endoscopists in the detection of diminutive lesions. The effect of AI on clinically meaningful outcomes such as incidence of interval GI cancers, morbidity, mortality and cost effectiveness requires further research.

INTRODUCTION

Overview of artificial intelligence technologies

The field of artificial intelligence (AI) has witnessed exponential growth during the last few years. It has influenced various aspects of life including business, cybersecurity and education. Its transformative effect on healthcare is not an exception[1].

AI evolves around the development and deployment of computer systems capable of performing tasks that require human-level intelligence such as learning, problem solving and decision-making[2]. It is a wide-ranging term that encompasses multiple disciplines including machine learning (ML) and deep learning (DL) (Figure 1).

Figure 1 Relationship between artificial intelligence, machine learning, deep learning, convolutional neural networks and recurrent neural networks. CNN: Convolutional neural networks; RNN: Recurrent neural networks.

ML is a subset of AI that enables computer programs to learn from datasets and subsequently improve their performance through experience. The key to this learning process, unprogrammed decision making ability, and intelligence enhancement is the application of ML algorithms[3]. Effectuality of ML relies principally on the type and characteristics of the datasets and the selection of an appropriate learning algorithm. Outcomes of learning algorithms from a single category are different even when applied to one dataset. Thus, choosing a suitable algorithm that serves the intended purpose remains challenging. As an example, two different ML algorithms, decision tree and random forest from the tree-based category were compared (Figure 2). The accuracy and precision of predicting users’ mobile phone activities were higher for the latter despite using the same training data set and belonging to the same category of algorithms[3].

Figure 2  Categories of machine learning algorithms.

There are four main categories of ML algorithms (Figure 2): Supervised, unsupervised, semi-supervised and reinforcement learning[4]. Supervised learning is the most important methodology in ML. The distinguishing feature of supervised learning is the utilization of labelled training datasets to learn the function of calculating and modifying errors to achieve the desired outcome. This will induce computer models capable of mapping input variables to output variables. It has two main functions: (1) Data classification: Organizing data into categories; and (2) Data regression: Establishing the relationship between changes of an independent variable and its effect on other dependent variables. Examples of supervised learning are face recognition and image tagging feature in social networking sites[4,5]. In unsupervised learning, models learn to discover common hidden patterns among unlabeled datasets without explicit guidance from humans. It is used for data clustering where data is grouped into clusters based on their inherent similarities. Data in each cluster share more common features together than those in a different cluster. This is of great utility in reducing large amounts of data into smaller analyzable groups for the extraction of relevant information and the detection of data points that deviate from the usual pattern[6].

DL is a subset of ML[7]. A distinctive feature of DL is the outstanding performance with large amounts of data in comparison to traditional ML[8]. The structure and processing function in DL are inspired by the human brain and neuronal interaction. DL is based on the older fundamental concept of artificial neural network (ANN) where it uses multiple layers of algorithms organized in several layers of fully connected ANN, each of which interprets the data it receives differently[9]. Neural networks consist of interconnected nodes called neurons. Each neuron has a specific weight (strength) and threshold. Activation of a neuron and subsequent effect on other neurons occurs only when the output exceeds the threshold value, otherwise no data is propagated to the next layer[10]. A DL neural network comprises a minimum of 3 layers: An input layer, an output layer, and hidden layers in between. The most common types of DL networks are convolutional neural networks (CNN) and recurrent neural networks (RNN)[11].

CNN is among the most used DL networks and a main contributor to its popularity nowadays. They are capable of learning image features extraction from raw datasets without any human intervention, and that makes it superior to traditional ANN. CNNs are specifically designed to process a large array of two-dimensional and three-dimensional figures by utilizing the information across the pixels of an image, and then creating feature maps that are later on pooled together to capture a larger field of the image. Eventually, features from all CNN layers are combined to produce the final intended outcome. This mechanism explains their application in medical imaging analysis, visual recognition and natural language processing[9,12].

RNN is another familiar type of DL networks. It uses sequential or time-series data, i.e., it processes data where a specific order and arrangement of elements are required for a meaningful outcome. Memory is a distinguishing feature of RNN. This memory allows RNN to influence current input by experience learnt from previous inputs. Speech recognition, natural language processing and prediction problems (such as forecasting stock prices and predicting flood levels) are the most common applications of RNN[13].

The process of training AI models occurs over several steps. Initially, the targeted raw data goes through the process of data annotation, where the data gets prepared and structured into training data sets to feed the AI models later. Next, and depending on the type and complexity of data, suitable algorithms are deployed. The next step is validating the trained model with new data, the validation dataset. This step is essential to assess if the performance of the model aligns with the expectations. It also allows transfer learning, defined as retraining and fine-tuning of the model with new data exposure. The last step is testing the model’s classification or prediction ability with a separate set of new data, the testing data set (Figure 3).

Figure 3 Steps to train an artificial intelligence model. AI: Artificial intelligence.

AI in gastroenterology

AI has revolutionized various image-dependent medical fields such as radiology, gastroenterology, dermatology and pathology[14]. However, with the growing bulk of digital images and advances in imaging modalities comes the need for advanced AI models that have the ability to efficiently process and learn from these large optical datasets. The evolution of ML through DL and specifically CNN provided the computer systems with these fundamental features facilitating the implementation of AI in the aforementioned fields.

All disciplines of gastroenterology are subjects for AI research and application. Nevertheless, endoscopy [esophagogastroduodenoscopy, colonoscopy, capsule endoscopy and device assisted enteroscopy (DAE)] remain the cornerstone diagnostic and therapeutic procedures in gastroenterology. They yield enormous digital datasets that represent an ideal substrate to feed and train computer models. As a result, the main focus of AI research is drawn to endoscopic image analysis rather than data analysis as in the fields of hepatology, pancreaticobiliary diseases and inflammatory bowel diseases[15].

Gastrointestinal (GI) malignancies remain a leading cause of mortality worldwide. Despite the availability of well-established screening programs and the decreased mortality associated with their implementation, colorectal, gastric and esophageal cancers are still responsible for more than 3.5 million new cases and around 2 million deaths yearly[16]. Established risk factors contributing to the high percentage of cancers post endoscopy (interval cancers) are high miss rates of early cancers during the procedure, as well as failed recognition and sampling of high-risk premalignant lesions[17,18]. Research has been directed towards the development of AI algorithms to optimize the effectiveness of endoscopic procedures by eliminating inter-endoscopist variability. This led to the development of clinically meaningful computer assisted devices (CAD) for lesion detection [computer aided detection (CADe)], diagnosis [computer-aided diagnosis (CADx)] and quality assessment[19]. CADe helps in the detection of mucosal abnormalities through recognition of changes in shape, color and texture whereas CADx is used for optical diagnosis and prognostication of the detected lesions without the need for histopathological assessment[20].

This review sheds light on the role of AI assisted endoscopy in the detection and characterization of premalignant and early malignant lesions of the upper GI tract, small intestine as well as the colorectum.

METHODOLOGY

A comprehensive literature search was conducted using PubMed databases to identify relevant studies on the application of AI in the detection and characterization of GI premalignant and early malignant lesions. The search covered publications from January 2003 to April 2025. The following keywords were used and combined using Boolean operators: “Artificial intelligence”, “deep learning”, “machine learning”, “convolutional neural networks”, “CADx”, “CADe”, “gastrointestinal”, “colorectal”, “gastric”, “esophagus”, “small bowel”, “Barrett’s esophagus”, “endoscopy”, “colonoscopy”, “gastroscopy”, “upper endoscopy”, “neoplasia”, “precancerous”, “premalignant” and “dysplasia”. In addition, the “related articles” feature in PubMed was used to expand the search and identify further relevant studies. Eligible studies were those involving endoscopic applications of AI in luminal GI neoplasia. Studies focusing on biliary, hepatic, or pancreatic malignancies, and those not involving endoscopic procedures were excluded. Only peer-reviewed articles published in English were considered. Reference lists of relevant reviews and included articles were also screened to identify additional eligible studies.

AI IN THE DETECTION AND CHARACTERIZATION OF PREMALIGNANT LESIONS AND EARLY CANCERS OF THE UPPER GI TRACT

Precancerous lesions and early cancers of the esophagus and stomach are often subtle, flat, multifocal and easily missed[21,22]. By improving tumor detection rate, AI effectively enhances the diagnostic yield in upper GI cancer screening[23,24].

AI in early esophageal neoplasia

In the esophagus, the premalignant lesion is Barrett’s esophagus (BE)-associated dysplasia, a precursor of esophageal adenocarcinoma (EAC), while squamous dysplasia is the precursor of esophageal squamous cell carcinoma (ESCC). Furthermore, early-stage or superficial ESCC or EAC is defined histopathologically as high-grade intraepithelial neoplasia or cancer confined to the mucosa and submucosa. Most esophageal cancer cases are still diagnosed at advanced stages, contributing to poor prognosis, higher mortality, and increased economic burden[25,26]. In light of this, AI may improve early detection and characterization of early esophageal cancers and precancerous lesions, enabling timely intervention and better patient outcomes.

BE: Early detection of BE is essential for improving both clinical outcomes and cost-effectiveness. BE affects about 1% of the global population and typically progresses through stages of low-grade and high-grade dysplasia before developing into EAC. Therefore, endoscopic surveillance is essential, as more than 90% of dysplastic lesions are detected either during the initial endoscopy or within six months, suggesting that many lesions are already present but may be initially overlooked[27]. Nevertheless, neoplasia in BE can be missed in up to 33% of cases[26]. The early detection of BE and associated dysplastic changes may reduce cancer-related mortality. Recent advances in AI have significantly improved the endoscopic detection and characterization of BE-associated neoplasia[28]. Table 1 provides a summary of studies evaluating AI models for the detection and characterization of BE and associated neoplasia.

Table 1 Summary of artificial intelligence studies on the detection and characterization of Barrett’s epithelium and associated neoplasia[8,29,30,33,34,37,39-42,45].

Ref.	Purpose	Design	Training set	Test set	IEE	AI model	Performance
de Groof et al[39]	Real-time Det + Ch of early BE neoplasia	Prospective study	1247 i	297 i	WLI	Hybrid: ResNet-UNet	Acc: 89%; Se: 90%; Sp: 88%
de Groof et al[40]	Real-time Det + Ch of early BE neoplasia	Prospective pilot study	1544 i	144 i	WLI	Hybrid: ResNet-UNet	Acc: 90%; Se 91%; Sp: 89%
Hashimoto et al[41]	Distinguish dysplastic from nondysplastic BE	Prospective pilot study	1374 i	458 i	WLI; NBI	Inception-ResNet-v2	Acc: 95%; Se: 96%; Sp: 94%
Ali et al[33]	BE risk stratification	Prospective pilot study	10000 i	194 v	WLI; NBI	ResNet-50; DeepLabv3 +	Acc: 98%
Hussein et al[45]	Det and delineation of BE dysplasia	Retrospective study	148936 v	276 v	WLI; i-scan	ResNet-101	Se: 91%; Sp: 79%; AUC: 0.93
Abdelrahim et al[37]	Det and localization of BE neoplasia	Prospective multicenter trial	1090171 i + v	471 i; 75 v	WLI	VGG16; SegNet	I: Acc/Se/Sp: 95%; v: Acc: 92%; Se: 95%; Sp: 91%
Tsai et al[30]	Det of BE	Retrospective study	771 i	160 i	NBI	EfficientNetV2B2	Acc: 94%; Se: 94%; Sp: 94%
Xin et al[8]	Det of early BE neoplasia	Retrospective multicenter	14046 i	400 i; 188 v		EfficientNet-Lite1; MobileNetV2; DeepLabv3 +	I: Se: 88%; Sp: 90%; v: Se: 79%; Sp: 94%
Meinikheim et al[34]	AI impact on nonexp BE neoplasia Det	Multicenter RCT	51273 i	96 v	WLI; NBI; Chromo	DeepLabv3; ResNet-50	Se: 70% to 78% w/AI; Sp: 67% to 73% w/AI
Jukema et al[42]	AI vs exp/nonexp in BE neoplasia Det	Prospective	3468 i	161 i; 161 v	NBI	EfficientNet-Lite1	Se: 84% to 96% w/AI; Sp: 90% to 98% w/AI
Jong et al[29]	Ensure AI reliability in BE neoplasia Det	Retrospective real word experience	1102 i; 12011 v	117 i	WLI; NBI	Hybrid: ResNet-50-vision transformer	Se: 85% (high-quality I); 62% (mod); 47% (low)

Det: Detection; Ch: Characterization; BE: Barrett’s epithelium; AI: Artificial intelligence; exp: Expert endoscopists; nonexp: Nonexpert endoscopists; RCT: Randomized controlled trial; i: Images; v: Videos; WLI: White light imaging; NBI: Narrow-band imaging; IEE: Image-enhanced endoscopy; Chromo: Chromoendoscopy with acetic acid and indigo carmine; ResNet: Residual network; UNet: U-shaped network; Deeplab: Deep convolutional neural networks for semantic image segmentation; SegNet: Segmentation network; VGG: Visual geometry group; EfficientNet: Efficient convolutional neural network; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; AUC: Area under the curve; mod: Moderate; w: With.

Multiple CADe systems-trained on extensive multicenter image and video datasets have demonstrated high accuracy, sensitivity, and real-time performance[29]. These systems often utilize white light imaging (WLI), and some of them rely on narrow-band imaging (NBI)[30,31] or linked-color imaging (LCI)[32] to enhance visualization and improve lesion detection. Traditional endoscopic detection of BE is challenging, with expert endoscopists demonstrating only 50% sensitivity and low interobserver agreement. AI systems help address these limitations by significantly reducing interobserver variability and improving overall diagnostic consistency. Most of these models were evaluated using static endoscopic images; however, few studies also incorporated video datasets to better reflect real-world clinical settings. Notably, detection performance remained robust across challenging test sets and live settings, including short segment BE[32,33], although image quality and training data diversity influenced outcomes. For instance, some DL models achieved sensitivities exceeding 90% and consistently outperformed nonexpert endoscopists, with some matching and in certain cases surpassing expert endoscopists[30,34-37]. A meta-analysis confirmed diagnostic performance of AI in detecting early BE neoplasia with pooled sensitivity and specificity around 90% and 84%, respectively[38]. In several studies, AI showed higher sensitivity than endoscopists and enabled fast, accurate lesion localization, supporting its integration into routine screening and surveillance[37,39,40].

AI-based CADx systems, particularly those trained on NBI, have demonstrated high standalone sensitivity and specificity in distinguishing dysplastic from non-dysplastic BE[31,41]. Moreover, some AI models have shown the ability to differentiate between low-grade and high-grade dysplasia[39]. In addition to their diagnostic accuracy, CADx systems also boost the confidence and performance of general endoscopists, enabling them to reach expert-level accuracy[42,43]. Other systems using magnifying endoscopy (ME) with i-scan or CNNs trained on NBI zoom videos achieved high area under the curve (AUC) and real-time frame processing speeds, enabling accurate dysplasia classification and optimizing biopsy site selection[41,44,45]. Tools based on WLI were also explored, with some showing lower performance compared to NBI and LCI, reinforcing the value of optical enhancement for AI-based characterization[31,32].

Early ESCC: ESCC is equally challenging to diagnose due to its extremely subtle endoscopic appearance and the absence of specific symptoms in the initial stages[46]. Enhanced imaging techniques, including Lugol dye chromoendoscopy and NBI, are used to improve lesion detection and margin visualization[28]. Population-based screening is typically recommended in high-risk geographic regions and for individuals with known risk factors, including a history of head and neck tumors[47]. Nevertheless, despite advancements in endoscopic imaging, the majority of ESCC cases are still diagnosed at advanced stages, which portends a poor prognosis[28]. DL models have shown the potential to improve early cancer detection and lesion characterization, predict invasion depth, and delineate lesion margins with remarkable accuracy, all contributing to effective treatment planning[28,48]. Table 2 provides a summary of studies evaluating AI models for the detection and characterization of early ESCC and precancerous lesions.

Table 2 Summary of artificial intelligence studies on the detection and characterization of premalignant and early esophageal squamous cell carcinoma[49-54,66,67,70].

Ref.	Purpose	Design	Training set	Test set	IEE	AI model	Performance
Horie et al[67]	Det of early ESCC	Retrospective	8428 i	1118 i	WLI; NBI	SSD	Acc/Se: 98%; NPV: 95%; PPV: 40%
Ohmori et al[70]	Det of early ESCC	Retrospective	11283 non-ME i; 11279 ME i	523 non-ME i; 204 ME i	WLI; NBI; BLI	SSD	Non-ME Acc: 79%; Se: 95%; Sp: 69%; ME: Acc: 77%; Se: 98%; Sp: 56%
Fukuda et al[66]	AI vs exp in early ESCC Det and Ch	Retrospective	17274 i	144 v	NBI; BLI	SSD	Det: Acc: 63%; Se: 91%; Sp: 51%; Ch: Acc: 88%; Se: 86%; Sp: 89%
Yang et al[52]	AI vs exp in Det of early ESCC	Retrospective	22994 i	222 i; 104 v	WLI; NBI; BLI	YOLOv3; ResNet-v2	Acc: 99%; Se: 100%; Sp: 99%
Yuan et al[51]	Prediction of ID in early ESCC	Retrospective multicenter	7094 i	1589 i	NBI	HRNet; OCRNet	Acc: 91%; ID prediction: 74%
Tani et al[50]	Real-time Det of early ESCC	Prospective single center trial	25048 i	237 i	WLI; NBI	ResNet-101	Acc: 81%; Se: 68%; Sp: 83%
Li et al[49]	Det of early ESCC and precancerous lesions	RCT	26543 i	3117 i	WLI; NBI	ENDOANGEL-ELD	Acc: 98%; Se: 90%; Sp: 98%
Ma et al[54]	Det + optical biopsy for early ESCC	Prospective	25056 i	2442 i; 187 v	WLI; pCLE	iCLE inception-ResNet-v2	Acc: 98%; Se: 95%; Sp: 99%
Aoyama et al[53]	AI vs exp/nonexp in early ESCC Det	Prospective	280 v	115 v	NBI	YOLOv3	Acc: 77%; Se: 76%; Sp: 79%

Det: Detection; Ch: Characterization; ESCC: Esophageal squamous cell carcinoma; AI: Artificial intelligence; ID: Invasion depth; IEE: Image-enhanced endoscopy; exp: Expert endoscopists; nonexp: Nonexpert endoscopists; RCT: Randomized controlled trial; i: Images; v: Videos; ME: Magnified endoscopy; WLI: White light imaging; NBI: Narrow-band imaging; BLI: Blue-laser imaging; pCLE: Probe-based confocal laser endomicroscopy; SSD: Single shot multi-box detector; YOLO: You only look once; ResNet: Residual network; HRNet: High-resolution network; OCRNet: Object-contextual representations network; ELD: Esophageal lesion detection; iCLE: Integrated confocal laser endomicroscopy; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; NPV: Negative predictive value; PPV: Positive predictive value.

AI systems have significantly improved the detection of superficial ESCC and precancerous lesions, while also reducing inter-operator variability. Several randomized controlled trials (RCTs) using CNNs[49-51] have demonstrated increased detection rates, reduced miss rates and improved diagnostic confidence among novice and mid-level endoscopists. Further, several models are performing on par with[52] or exceeding experts endoscopists[53-56]. Meta-analyses[57,58] further confirmed the high diagnostic accuracy of AI-assisted endoscopy for early ESCC, with pooled sensitivities and specificities exceeding 90%, consistently outperforming endoscopists in early detection of sub centimetric lesions. Retrospective studies[59-62] utilized you only look once (YOLO)-based architectures trained on large datasets of WLI, NBI, LCI and blue-laser imaging (BLI). CAD systems and other DL models consistently achieved high sensitivity, specificity and accuracy in identifying early ESCC[59,63,64]. Additionally, the integration of hyperspectral imaging with AI further enhanced detection performance, particularly in WLI settings[60,65].

Beyond detection, AI has shown strong potential in the characterization and staging of esophageal neoplasia. DL models can predict invasion depth, classify vascular patterns and enhance optical biopsy interpretation, ultimately reducing the need for unnecessary tissue sampling[54]. For example, they can reliably differentiate between esophageal squamous dysplasia and nondysplastic mucosa, as well as distinguish superficial from advanced cancer with high diagnostic accuracy[60,66,67]. Several modern and sophisticated DL models such as clustering-constrained attention multiple instance learning[68], efficient net-vision transformer hybrids[64], and single shot multi-box detector (SSD) architectures[69], have been trained on histologic slides, ME-NBI images and videos to predict mucosal, submucosal, and deeper invasion. These models can reveal distorted or irregular vascular patterns in invasive ESCC. They consistently achieved high accuracy and outperformed or matched experienced endoscopists[70-72]. For example, the interpretable AI-based invasion depth prediction system significantly improved endoscopists’ accuracy in assessing invasion depth and was validated across multiple centers[73]. Similarly, AI tools trained to classify intrapapillary capillary loops (IPCLs) subtypes using ME enhanced junior endoscopists’ ability to assess invasion depth[74]. The Japanese IPCLs classification is frequently used to guide AI training for depth assessment, helping to differentiate intramucosal from submucosally invasive ESCC and aiding in the selection of lesions suitable for endoscopic resection[75].

AI in early gastric neoplasia

Gastric premalignant conditions (GPMC) include atrophic gastritis (AG), gastric intestinal metaplasia (GIM), dysplasia, and certain gastric epithelial polyps such as adenomas, some hyperplastic polyps, as well as type 1 neuroendocrine tumors[47]. Early-stage gastric cancer (EGC) is defined as a carcinoma confined to the mucosa or submucosa, regardless of the presence or absence of lymph node metastasis[47,76]. The risk of progression from GMPC varies based on established risk factors, including age, Helicobacter pylori infection, anatomic extent and histologic severity of AG/GIM, GIM subtype (complete vs incomplete), family history of gastric cancer, smoking and dietary habits[76]. Advanced-stage gastric cancer has a median survival of less than one year and a five-year survival rate below 20%[77]. In contrast, EGC carries a favorable prognosis, with five-year survival exceeding 90% when detected and treated promptly. Hence the critical importance of early detection and surveillance[28,76]. High-quality upper endoscopy is strongly recommended for identifying GPMC and EGC, with an emphasis on adequate mucosal visualization, careful inspection and detailed anatomical mapping[47,78]. AI may enhance the early detection and characterization of GPMC and EGC, facilitating timely intervention, guiding appropriate surveillance strategies, and ultimately improving patient outcomes.

GMPC: Early detection of GPMC is essential for optimizing clinical outcomes. Helicobacter pylori testing, treatment, and eradication confirmation are recommended in all individuals with GPMC[47]. Routine gastric biopsies are not recommended unless there is suspicion of GPMC[47]. Although AG, GIM and dysplasia are detectable using WLI with or without NBI or LCI, they are frequently missed due to limited endoscopist familiarity, endoscopist fatigue, lesion variability and blind spots. AI tools have shown promise in detecting GPMC in a well-visualized stomach, though current evidence remains insufficient to support routine clinical use[76,79]. By providing real-time assistance, DL models can reduce miss rates, improve lesion detection and enhance lesion characterization[80]. Table 3 provides a summary of studies evaluating AI models for the detection and characterization of GPMC.

Table 3 Summary of artificial intelligence studies on the detection and characterization of gastric premalignant conditions[81,83-90,92,94-100].

Ref.	Purpose	Design	Training set	Test set	IEE	AI model	Performance
Zhang et al[92]	Real-time Det of gastric polyps	RCT	708 i	50 i	WLI	SSD-GPNet	mAP: 90%; PDR by > 10%
Zhang et al[99]	Det of AG	Retrospective	3829 i	1641 i	WLI; i-scan	CNN-CAG	Acc: 94%; Se: 95%; Sp: 94%
Xu et al[85]	AI vs exp/nonexp in Det of GPMC	Retrospective multicenter	5198 i	1052 i; 98 v	ME-NBI; ME-BLI	ENDOANGEL	AG: Acc: 86% (i); 88% (v); GIM: Acc: 86% (i); 90% (v)
Lin et al[90]	Det of AG/GIM	Retrospective multicenter	2193 i	273 i	WLI	TResNet	AG: Acc: 96%; AUC: 0.98; Se: 96%; Sp: 96%; GIM: Acc: 98%; AUC: 0.99; Se: 98%; Sp: 97%
Watanabe et al[100]	Det of gastric indefinite for dysplasia lesions	Retrospective	2961 i	248 i	WLI	SSD + miR148a DNA methylation	AUC: 0.93 (exp) > 0.83 (AI + miR148a) > 0.59 (trainees)
Kodaka et al[95]	AG Det and OLGA staging in patients w/Helicobacter pylori infection	Retrospective	11497 i	7724 i	WLI	ResNet-50	AUC: 0.75 (AI + Kyoto) > 0.67 (AI + OLGA) > 0.66 (AI alone)
Zhao and Chi[98]	Severity classification of AGAI vs endoscopists	Prospective	2922 i	268 v	NBI	UNet	(Increase) DR of mod AG (16% vs 8%) and severe AG (7% vs 3%); (decrease) unnecessary Bx
Zhao et al[89]	Det of AGAI vs endoscopists	Prospective case-control	4175 i	676 v	NBI	UNet	Acc: 91% vs 72%; Se: 84% vs 63%; Sp: 97% vs 82%; AUC: 0.91 vs 0.74
Yang et al[81]	Det of AG/GIM	Retrospective	21420 i	5355 i	WLI; LCI	SE-ResNet	AG: Acc: 97%; Se: 99%; Sp: 95%; GIM: Acc: 99%; Se: 99%; Sp: 99%
Li et al[96]	Severity classification of GIM	Retrospective	837 i	278 i	NBI; LCI	CDCN	Acc: 84%
Shi et al[88]	Det of AG	Retrospective	6216 i	600 i; 118 v	WLI	GAM-EfficientNet	i: Acc: 94%; Se: 93%; Sp: 94%; v: Acc: 92%; Se: 96%; Sp: 89%
Iwaya et al[87]	Det of GIM and OLGIM staging	Retrospective	5753 i	1150 i	HE slides	ResNet-50	Se: 98%; Sp: 95%; OLGIM stage III/IV classified in 18%
Fang et al[86]	AI vs pathologists in Det and grading of AG/GIM	Prospective multicenter	1745 i	545 i	Pathology slides	GasMIL	(Increase) pathologists’ performance (AUC: 0.95 vs 0.88)
Tao et al[97]	AG Det and risk stratification vs exp	Retrospective	5856 i	869 i; 119 v	WLI	UNet ++ ResNet-50 ENDOANGEL	(Increase) Se vs exp i: 93% vs 77%; v: 95% vs 86%
Niu et al[94]	GIM grading and OLGIM staging	Retrospective multicenter	470 i	333 i	ME-NBI	Faster R-CNN	Pred high-risk stage Acc: 84%
Zou et al[83]	Det of Helicobacter pylori infection AI vs endoscopists	Multicenter RCT	7377 i	2080 i	WLI	EfficientNet	Acc: 93% vs 76%; Se: 92% vs 79%; Sp: 93% vs 75%
Xu et al[84]	AI vs exp/nonexp in Det of AG/GIM	Single center RCT	NA	1968 v	WLI	ENDOANGEL	(Increase) DR of AG: 23% vs 17%; (Increase) DR of GIM: 14% vs 9%

AI: Artificial intelligence; IEE: Image-enhanced endoscopy; Det: Detection; exp: Expert endoscopists; nonexp: Nonexpert endoscopists; RCT: Randomized controlled trial; i: Images; v: Videos; AG: Atrophic gastritis; GPMC: Gastric premalignant conditions; GIM: Gastric intestinal metaplasia; OLGA: Operative link on atrophic gastritis assessment; AGAI: Atrophic gastritis artificial intelligence; OLGIM: Operative link for gastric intestinal metaplasia assessment; w: With; WLI: White light imaging; NBI: Narrow-band imaging; BLI: Blue-laser imaging; ME: Magnified endoscopy; HE: Hematoxylin-eosin; SSD: Single shot multi-box detector; GPNet: Gated pyramid network; CNN: Convolutional neural networks; CAG: Chronic atrophic gastritis; TResNet: Transformer-residual network; ResNet: Residual network; UNet: U-shaped network; R-CNN: Region-based convolutional neural networks; mAP: Mean average precision; PDR: Polyp detection rate; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; AUC: Area under the curve; DR: Detection rate; Bx: Biopsies; mod: Moderate; LCI: Linked-color imaging; CDCN: Color-guided deformable convolution network; Pred: Prediction; NA: Not applicable.

Numerous studies, RCTs and meta-analyses have demonstrated the effectiveness of AI in detecting GPMC[81-83]. Multiple AI models, particularly CNNs, have shown high accuracy in detecting Helicobacter pylori infection from upper GI endoscopic images, significantly improving diagnostic precision and potentially reducing biopsies requirements and healthcare costs. Several DL models[84-88] were trained on large datasets from diverse imaging modalities (WLI, LCI, BLI, NBI, and histologic images) to detect AG and GIM. They consistently achieved high diagnostic accuracy, sensitivity, and specificity, often surpassing both expert and nonexpert endoscopists[89-91]. Additionally, AI tools like SSD-GPNet[92] showed high speed and accuracy in detecting gastric polyps, particularly small ones. However, another study has shown that the added benefit of AI in polyp detection may be limited compared to the performance of both junior and senior endoscopists[93].

AI is also used to characterize and stage GPMC, particularly in the context of cancer risk stratification. Several DL models were developed to predict standardized histological classifications such as operative link on GIM assessment and operative link on AG assessment, with performance matching or exceeding that of expert pathologists[86,94,95]. They offered highly accurate interpretation of histopathologic and endoscopic features of GIM and AG, enhancing consistency in diagnosis and enabling non-invasive risk stratification and severity grading without the need for biopsy[87,96-99]. Other studies integrated AI with molecular markers such as miR-148a methylation[100]. This combined histologic and molecular approach improves diagnostic accuracy for indefinite for dysplasia gastric lesions, especially among nonexpert endoscopists and pathologists.

EGC: EGC often presents subtly, with a miss rate of around 10%, particularly among less experienced endoscopists or during low-quality exams[78]. Endoscopic resection has become central to EGC staging and treatment, yet 25%-40% of endoscopic submucosal dissections in expert centers target incurable lesions, underscoring limitations in current selection methods[78]. AI has shown strong potential in improving EGC detection, reducing variability across skill levels and enhancing lesion characterization[28,58]. Table 4 summarizes studies evaluating AI models for the detection and characterization of EGC.

Table 4 Summary of artificial intelligence studies on the detection and characterization of early gastric cancer[101-108,110,112-117,119,122].

Ref.	Purpose	Design	Training set	Test set	IEE	AI model	Performance
Zhu et al[122]	ID prediction of EGC	Retrospective	790 i	203 i	WLI	ResNet-50	Acc: 89%; Se: 76%; Sp: 96%; AUC: 0.94
Horiuchi et al[112]	AI vs exp in EGC Det	Retrospective	2570 i	174 v	ME-NBI	GoogLeNet	Acc: 85%; Se: 95%; Sp: 71%; AUC: 0.87
Wu et al[108]	Det of EGC	Multicenter RCT	NA	1050 v	WLI	ENDOANGEL	Acc: 85%; Se: 100%; Sp: 84%
Wu et al[107]	Det of EGC	RCT	NA	1812 v	WLI	ENDOANGEL-LD	(Decrease) miss rate (AI 6% vs endoscopists 27% RR = 0.22)
Wu et al[117]	AI vs exp in ID and DS of EGC	Prospective multicenter	1131 i	100 v	ME-NBI	ENDOANGEL	ID: Acc: 79% vs 64%; DS: Acc: 71% vs 64%
Wu et al[116]	Real time Det of EGC	Prospective single center trial	9824 i	2010 v	WLI	ENDOANGEL-LD	Acc: 92%; Se: 92%; Sp: 92%; PPV: 25%; NPV: 100%
Ueyama et al[103]	Det of EGC	Retrospective	5574 i	2300 i	ME-NBI	ResNet-50	Acc: 99%; Se: 99%; Sp: 98%
He et al[115]	Det of EGC	Retrospective multicenter	4667 i	4702 i; 187 v	ME-NBI	ENDOANGEL-ME	Acc: 90%; Se: 93%; Sp: 94%
Li et al[110]	AI vs exp in EGC Det	Retrospective	1630 i	267 i; 77 v	ME-NBI	ENDOANGEL-LA	i: Acc: 89%; Se: 86%; Sp: 92%; v: Acc: 87%; Se: 84%; Sp: 88%
Tang et al[114]	Det of EGC	Retrospective multicenter	13151 i	1577 i; 20 v	NBI	YOLOv3	Acc: 93%; AUC: 0.95
Jin et al[113]	Det of EGC AI vs exp	Prospective	5708 i	1425 i; 10 v	WLI; NBI	Mask R-CNN	Acc: 90%; Se: 91%; Sp: 89%
Gong et al[106]	Real-time Det and ID prediction of EGC	RCT	5017 i	2524 v	WLI	CDSS	DR: 96%; ID Acc: 86%; lesion class: Acc: 82%
Lee et al[101]	EGC pathological Ch	Retrospective	4336 i; 153 v	436 i; 89 v	WLI	ENAD CAD-G	Acc: UH: 90%; SMI: 88%; LVI: 88%; LNM: 93%
Chang et al[119]	Classification of EGC	Retrospective real-world data	21918 i	6785 i; 296 v	WLI	ENAD CAD-G	i: Acc: EGC: 82%; dysplasia: 88%; v: Acc: EGC: 88%; dysplasia: 91%
Zhao et al[102]	Det of EGC LCI vs WLI	Retrospective	9021 i	116 v	WLI; LCI	CADe	(Increase) Se: LCI: 94% vs WLI: 79%; Sp: 93% in both
Lee et al[101]	Det of EGC	Retrospective	30000 i	500 i	WLI	CADe (ALPHAON®)	Acc: 88%; Se: 93%; Sp: 87%; AUC: 0.96
Soong et al[105]	Raman spectroscopy for EGC risk strat	RCT	NA	25 v	WLI	SPECTRA IMDx™	Acc: 100%; Se: 80%; Sp: 92%
Feng et al[104]	AI vs exp/nonexp in EGC Det	Prospective	12000 i	1289 i; 130 v	WLI	DCNN	Se: 97%; Sp: 89%; AUC: 0.93

EGC: Early-stage gastric cancer; ID: Invasion depth; AI: Artificial intelligence; IEE: Image-enhanced endoscopy; Det: Detection; exp: Expert endoscopists; nonexp: Nonexpert endoscopists; RCT: Randomized controlled trial; i: Images; v: Videos; DS: Differentiation status; Ch: Characterization; LCI: Linked-color imaging; WLI: White light imaging; strat: Stratification; NA: Not applicable; NBI: Narrow-band imaging; ME: Magnified endoscopy; ResNet: Residual network; GoogLeNet: Google network; LD: Lesion detection; LA: Logical anthropomorphic diagnostic system; YOLO: You only look once; R-CNN: Region-based convolutional neural networks; CDSS: Clinical decision support system; ENAD: Endoscopy neural assisted detection; CAD-G: Computer-aided detection for gastric lesions; CADe: Computer aided detection; DCNN: Deep convolutional neural network; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; AUC: Area under the curve; RR: Relative risk; NPV: Negative predictive value; PPV: Positive predictive value; UH: Undifferentiated histology; SMI: Submucosal invasion; LVI: Lympho vascular invasion; LNM: Lymph node metastasis.

AI systems have significantly improved the detection rates of EGC in both retrospective[101-103] and prospective real-time settings[104-106], with RCTs demonstrating reduced miss rates and blind spots, as well as increased inspection time, compared to routine endoscopy[107,108]. Further, endoscopists assisted by CNNs can outperform nonexpert endoscopists in identifying EGC, but their performance remains comparable to expert ones[105,109-112]. In contrast, YOLOv3-based and mask region-based CNNs have shown superior accuracy and sensitivity, surpassing both junior and senior endoscopists[113,114]. Similarly, real-time AI models like ENDOANGEL-ME achieved near-perfect sensitivity in prospective cohorts[115-117], often surpassing experts in accuracy and reducing miss rates during WLI. Nevertheless, DL models trained on LCI or NBI outperformed those based on WLI alone[102]. A meta-analysis[118] confirmed AI’s high accuracy in detecting EGC, with best results achieved using ME-NBI, further validating its real-world utility.

Additionally, CNN-based models have shown strong performance in EGC characterization to support decision-making in endoscopic resection. They classify lesions while assessing key features such as invasion depth, histologic differentiation and lymphovascular involvement[119,120]. For instance, a meta-analysis evaluating invasion depth prediction, reported high pooled sensitivity, specificity and AUC[111]. Moreover, in a prospective real-time human-machine competition, AI outperformed expert endoscopists in predicting the differentiation status and invasion depth of EGC[117]. However, in retrospective studies[121,122], it enabled novice endoscopists to perform at expert levels in differentiating intramucosal from advanced gastric cancer. Similarly, a video-based classifier outperformed static image models in depth prediction[123]. Finally, advancing beyond standard DL models, explainable AI tools like the double-check support system have integrated lesion classification with real-time image quality assessment, improved interobserver agreement, reduced diagnostic time and provided greater consistency in guiding endoscopic resection decisions[124].

Despite advancements in AI for detecting and characterizing early cancers and precancerous lesions of the upper GI tract, widespread clinical integration remains limited by several challenges. These include the reliance on large, high-quality annotated datasets, retrospective study designs, limited generalizability beyond academic centers, and variability in histopathologic reference standards. Real-time implementation also faces technical and logistical barriers, particularly in community settings. While video-based studies suggest AI can outperform endoscopists, especially in diagnostic accuracy and lesion characterization, large-scale, prospective validation in real world settings is essential to improve AI reliability in routine practice. As AI continues to evolve, its potential to enhance detection, characterization, and reduce variability in clinical practice underscores the importance of integrating these tools into routine endoscopic workflows to improve patients’ outcomes.

AI IN THE DETECTION AND CHARACTERIZATION OF PREMALIGNANT AND EARLY MALIGNANT LESIONS OF THE SMALL BOWEL

Small bowel (SB) tumors represent up to 6% and 3% of all GI neoplasms and malignancies, respectively. The incidence of SB tumors has been increasing over the last few years[125]. The majority of SB neoplasms are malignant and fall into four categories: Adenocarcinomas, neuroendocrine tumors, lymphomas and sarcomas[125]. Premalignant lesions in the SB comprise different types of polyps (adenomatous, hamartomatous and serrated) which could be sporadic or part of a hereditary tumor/polyposis syndrome, or associated with immune mediated inflammatory conditions such as Crohn’s or celiac disease[126]. It is worth mentioning that a considerable number of patients with SB tumors have metastatic disease at diagnosis. Late presentations are ascribed to the nonspecific and variable symptoms, the unreachable site by conventional endoscopy, and the subtle changes in appearance in comparison to the surrounding mucosa in visible lesions[127]. There are different staging systems for different SB tumors, and no consensus on the precise definition of an early SB malignancy has been issued in the literature.

Endoscopic evaluation of the SB had formerly been a longstanding challenge for the gastroenterologist, due to the complex anatomy of the SB and the limited direct visualization achieved with traditional esophagogastroduodenoscopy and colonoscopy. The introduction of video capsule endoscopy (VCE) and DAE has rendered the investigation of suspected SB lesions more feasible. VCE allows real-time complete visualization of the entire intestinal mucosal surface yielding thousands of images. DAE serves as a diagnostic as well as endotherapeutic tool[128]. However, this great stride in SB endoscopy carries multiple challenges alongside. Interpreting the large bulk of images in VCE is time consuming and physicians spend up to 120 minutes reading and reporting a full CE recording. On the other hand, DAE is technically demanding and requires high level of experience and skills to overcome the risk of oversight and distraction, and hence reduce lesion miss rate[129,130].

Multiple studies using AI for the development of CAD to help physicians overcome the hurdles associated with the increased utilization of VCE and DAE have been conducted. The main focus of AI implementation is to alleviate the burden on endoscopists through the reduction of reading time, enhanced real time diagnosis and detection of hard-to-detect lesions, and decrease miss rate[131,132]. Table 5 summarizes the studies.

Table 5 Summary of artificial intelligence studies on the detection of polyps and protruding lesions by video capsule endoscopy or small bowel endoscopy[135,138-140,213-221].

Ref.	Endoscopy type	Study design	Training set	Test set	AI model	Performance
Barbosa et al[135]	VCE	Retrospective	104 tumors, 100 normal	700 tumors, 2300 normal	CNN	Se: 94%; Sp: 93%
Li and Meng[213]	VCE	Retrospective	540 tumors, 540 normal	60 tumors, 60 normal	ML (SVM)	Acc: 84%; Se: 82%; Sp: 8%
Constantinescu et al[138]	VCE	Prospective	54 videos	90 images (32 polyps, 58 normal)	ANN	Acc: 98%; Se: 94%; Sp: 91%
Liu et al[214]	VCE	Retrospective	1800 (105 patients, 89 videos)	89 videos (15 tumors, 74 normal)	ML (SVM)	Se: 98%; Sp: 97%
Faghih Dinevari et al[215]	VCE	Retrospective	300 tumors, 300 normal	100 tumors, 100 normal	ML (SVM)	Acc: 94%; Se: 94%; Sp: 93%
Yuan and Meng[216]	VCE	Retrospective	1000 polyps, 3000 normal	200 tumors, 600 normal	DL	Acc: 98%
Ding et al[217]	VCE	Retrospective	158, 235	113, 268, 334	CNN	Se: PPA: 99.9%; PLA: 99.9%; Sp: PPA and PLA: 100%
Saito et al[218]	VCE	Retrospective	30, 584	17507 (7507 protruding lesions)	CNN	Se: 91%; Sp: 80%
Xie et al[219]	VCE	Retrospective	148, 357, 922	146, 956, 145	CNN	Se: 99%
Cardoso et al[139]	Enteroscopy	Retrospective	6340	507 protruding lesions, 1078 normal	CNN	Acc: 97%; Se: 97%; Sp: 97%
Inoue et al[220]	EGD	Retrospective	531	1080	CNN	Se: 95%; Sp: 87%
Ding et al[221]	VCE	Retrospective	280, 426	240 videos	CNN	Acc: 81%; Se: 96%; Sp: 98%
Zhu et al[140]	DBE	Retrospective	8222	3148	CNN	Det model: Se: 92%; Sp: 93%; Class model: Se: 80%-93%; Acc: 86%

AI: Artificial intelligence; VCE: Video capsule endoscopy; EGD: Esophagogastroduodenoscopy; DBE: Double balloon enteroscopy; CNN: Convolutional neural network; ML: Machine learning; SVM: Support vector machine; ANN: Artificial neural network; DL: Deep learning; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; PPA: Per patient analysis; PLA: Per lesion analysis.

Recently, a systematic review and metanalysis[123] including 12 studies on the use of CAD of SB tumors and polyps during VCE showed a pooled sensitivity and specificity of 89% and 91%, respectively. Another study showed that AI assisted VCE reading outperformed expert gastroenterologists in terms of accuracy and sensitivity and decreased the mean reading time by up to 12-fold[133,134]. The majority of the studies detected protruding lesions defined as any elevation above the surface of the mucosa. This includes polyps, tumors and submucosal lesions. On the other hand, data on the role of AI in DAE is scarce and only few studies have been conducted.

Barbosa et al[135] were the first to propose a computer aided system using CNN for the detection of SB tumors called discrete wavelet transform. A novelty in this system was the characterization of texture alteration by statistical modelling of texture descriptors at different scales and colors. It initially showed a sensitivity of 99% and a specificity of 97%. Few years later, the system was tested in real life setting showing a sensitivity of 94% and specificity of 93%[136]. This change in performance metrics can be explained by overfitting training datasets where the models perform well during training but less well when applied to real world clinical settings[137].

Results from other retrospective studies showed comparable sensitivity and specificity (Table 5). AI scores surpassed those of experienced gastroenterologists predominantly in the sensitivity of detecting protruding lesions/polyps and reading time. For trainees, AI assistance increased the overall performance and accuracy to levels on par with experts.

A prospective study was performed[138] using an ANN based system to detect SB polyps during VCE. Feature extraction included primarily shape, but also color and texture. There was no statistically significant difference between AI and experienced physicians’ readings regarding sensitivity and specificity. Experienced physicians had better positive and negative predictive values (NPV).

Regarding DAE, Cardoso et al[139] designed a deep CNN algorithm for the detection of protruding enteric lesions that included epithelial tumors and subepithelial lesions encountered during single or double balloon enteroscopy (DBE). The sensitivity, specificity and overall accuracy of the model reached 97% each. The positive and NPVs were 95% and 99%, respectively.

Zhu et al[140] built a deep CNN system called ENDOANGEL-DBE for SB lesions in DBE. It is composed of detection and classification models. The model classified lesions into 4 categories: Diverticula, erosions/ulcers, angioectasia and protruding lesions. The sensitivity for the detection of protruding lesions was 93%. The total accuracy of ENDOANGEL-DBE was superior to endoscopists who recorded 77% (P < 0.01). The overall performance of ENDOANGEL-DBE was comparable to the experts in lesion classification (85 % and 81%; P = 0.253).

In conclusion, AI assisted SB endoscopy was superior in terms of accuracy and sensitivity for the detection of polyps and protruding lesions (including tumors) even when compared to expert endoscopists. Another significant advantage is VCE reading time reduction, a tedious time-consuming task for the gastroenterologist. Among all the studies investigating the detection of polyps and protruding lesions during VCE, there hasn’t been a clear histopathological classification of these lesions. This significant limitation makes the extent through which AI contributes to the detection of premalignant lesions in the SB unclear. The clinical relevance, effect on morbidity, mortality and cost effectiveness of these advancements are yet to be investigated with long term studies.

Future research might address the role of AI assistance in corelating the anatomical localization during DAE with findings on VCE, paving the way for endoscopic management of these lesions.

AI IN THE DETECTION AND CHARACTERIZATION OF PREMALIGNANT AND EARLY MALIGNANT LESIONS OF THE COLON

Colorectal cancer (CRC) is the third most common cancer worldwide and ranks second in cancer related mortality[141]. Colonoscopy is the gold standard for screening and prevention of CRC as it allows resection of precancerous lesions[142]. Premalignant lesions may be missed during colonoscopy; miss rates for adenomas, serrated lesions and advanced adenomas were 26%, 27% and 9%, respectively in a large meta-analysis[143]. This was shown to account for more than 50% of interval cancers, which represent 10% of all CRC[144,145]. Furthermore, lateral spreading tumors (LSTs) have a high malignant potential and can be elusive and easily missed during colonoscopy. LST non-granular type, including sessile serrated lesions (SSLs), are the most difficult to detect and pose the risk of post-colonoscopy interval cancer[146]. A 1% increase in the detection of right-sided serrated polyps resulted in a substantial 7% reduction in the risk of interval CRC[147]. Adenoma detection rate (ADR) is recognized as the most reliable quality indicator in screening colonoscopy, and a previous study revealed a 3% decrease in the risk of CRC for every 1% increase in ADR[148].

AI can assist in colonoscopy by improving the detection of premalignant lesions, enhancing the characterization of premalignant and early malignant lesions and by decreasing inter-endoscopists’ variability. CADe is the most developed AI system for identifying polyps.

CADe in colonoscopy

Effect of CADe assisted colonoscopy on ADR: Multiple prospective and retrospective studies evaluated the use of CADe in colonoscopy (Table 6). Maroulis et al[145] described an innovative detection system, colorectal lesions detector, that aids in detecting polyps by combining feature extraction and classification algorithms. The system’s accuracy in lesion detection exceeded 95%. An intelligent model that incorporates color-texture analysis methods and support vector machine’s algorithms into a sound pattern recognition framework could analyze low-quality videos with an accuracy of more than 94%[149].

Table 6 Summary of retrospective and prospective studies on computer aided detection for detecting colorectal polyps[150,222-229].

Ref.	Study design	Number	Country	Outcomes
Park et al[222]	Retrospective ex vivo	562 images	United States	Se: 86%; Sp: 85%
Billah et al[223]	Retrospective ex vivo	14000 images	Bangladesh	Se: 99%; Sp: 99%
Lequan et al[224]	Retrospective ex vivo	38 videos	China	Se: 71%; PPV: 88%
Zhang et al[225]	Retrospective ex vivo	2442 images	China	Se: 98%; PPV: 99%; Acc: 86%
Misawa et al[226]	Retrospective ex vivo	546 videos	Japan	Per-frame Se: 90%; Sp: 63%; Acc: 76%; Per-polyp Se: 94%
Urban et al[150]	Retrospective ex vivo	63559 images, 40 videos	United States	Se: 90%; Acc: 96%
Yamada et al[227]	Retrospective ex vivo	144823 video images	Japan	Se: 97%; Sp: 99%
Klare et al[228]	Prospective in vivo	55 colonoscopies	Germany	Per-polyp Se: 75%; ADR: 29%; PDR: 51% (31% and 56% in endoscopist)
Ozawa et al[229]	Ex vivo	23495 images	Japan	Se: 92%; PPV: 93%; Acc: 85%

Acc: Accuracy; Se: Sensitivity; Sp: Specificity; PPV: Positive predictive value; ADR: Adenoma detection rate; PDR: Polyp detection rate.

The first real-time CNNs based model was tested on more than 8000 hand-labeled images with more than 4000 polyps and showed an accuracy of more than 96% for polyp detection. It outperformed expert endoscopists with relatively low false positive rate (around 5%)[150]. Other studies had variable sensitivity and lower accuracy than that study (Table 6).

Multiple RCTs have assessed the performance of CADe compared to standard colonoscopy (Table 7). CADe improved ADR in most studies but had no effect in others when compared to standard colonoscopy. Initial trials showed that CADe improved ADR compared to standard colonoscopy when baseline ADR was rather low[151]. These findings were confirmed by a group with higher baseline ADR who used GI genius, the first Food and Drug Administration-approved CADe system[152]. To note that ADR was significantly higher for diminutive (≤ 5 mm) and small size (6 mm-9 mm) adenomas. The same group showed that AI also improves ADR as well as adenoma per colonoscopy (APC) performed by non-expert endoscopists[153]. More recent RCTs in the East demonstrated high baseline ADR in control arms that showed further increase with CADe systems[154,155].

Table 7 Summary of randomized controlled trials on computer aided detection for the detection of colorectal polyps[151-159,166,175-178,230-235].

Ref.	Study design	Number of patients	Country	Outcomes (AI vs control)
Wang et al[151]	Single center RCT	1058	China	ADR: 29% vs 20%
Wang et al[156]	Single center RCT (double blind)	962	China	ADR: 34% vs 28%
Gong et al[157]	Single center RCT	704	China	ADR: 16% vs 8%
Su et al[158]	Single center RCT	659	China	ADR: 29% vs 17%
Liu et al[159]	RCT	1026	China	ADR: 39% vs 23%
Wang et al[175]	Tandem RCT	369	China	AMR: 14% vs 40%
Repici et al[152]	Multicenter RCT	685	Italy	ADR: 55% vs 40%
Kamba et al[176]	Tandem RCT	358	Japan	AMR: 14% vs 37%
Glissen Brown et al[177]	Multicenter tandem RCT	223	United States	AMR: 20% vs 31%
Repici et al[153]	RCT	660	Italy, Switzerland	ADR: 53% vs 45%
Wallace et al[178]	Tandem RCT	240	Italy, United Kingdom, United States	AMR: 16% vs 32%
Shaukat et al[166]	RCT	1359	United States	APC: 1 vs 0.8
Xu et al[230]	Multicenter RCT	3059	China	ADR: 40% vs 32%
Gimeno-García et al[231]	Single center RCT	370	Spain	ADR: 55% vs 41%
Nakashima et al[154]	Single center RCT	415	Japan	ADR: 59% vs 48%
Karsenti et al[232]	Single center RCT	2592	France	ADR: 38% vs 34%
Lau et al[155]	Single center RCT	766	China	ADR: 58% vs 45%
Maas et al[233]	Multicenter RCT	916	Germany, Netherlands, United States, Israel	ADR: AI 37% vs 30%
Thiruvengadam et al[234]	Single center RCT	1100	United States	ADR: 43% vs 34%
Park et al[235]	Multicenter RCT	805	Korea	ADR: 35% vs 28%

AI: Artificial intelligence; ADR: Adenoma detection rate; AMR: Adenoma miss rate; APC: Adenoma per colonoscopy; RCT: Randomized controlled trial.

Interestingly, a double-blind RCT using a sham system (CADe-DB trial) showed that CADe significantly increased ADR[156]. Other RCTs showed significant increase in ADR compared to controls[157-159]. In addition, ENDOANGEL model significantly increased the detection of advanced adenomas (> 10 mm)[157].

An analysis of RCTs assessing deep CNN-based CADe in real-time colonoscopy and including a total of 4962 patients showed a higher pooled ADR in the CADe group, but the pooled rates of relative risk for advanced ADR and sessile serrated ADR were similar to controls[160]. Furthermore, a meta-analysis that studied 10 different CADe systems showed an increase in ADR of 24% but not for advanced adenomas and SSLs. It also showed higher rates of unnecessary removal of non-neoplastic polyps[161]. Another meta-analysis showed an increase in ADR (45% vs 37%) and a mild increase in the detection of advanced colorectal neoplasia (12.7% vs 11.5%)[162].

A network meta-analysis showed that AI increased ADR compared with both mucosal visualization tools and chromoendoscopy, but there was no increase in the detection of SSLs[163]. Another network meta-analysis showed improvement in ADR with AI compared to other endoscopic interventions aimed at increasing ADR (such as distal attachment devices, dye-based/virtual chromoendoscopy, water-based techniques, and balloon-assisted devices)[164]. Finally, a recent RCT comparing LCI colonoscopy to LCA (LCI + AI) colonoscopy found higher ADR and APC in LCA group, particularly for small and diminutive polyps[165].

By contrast, several recent RCTs and a retrospective study failed to demonstrate that AI-assisted colonoscopy improves ADR compared to standard colonoscopy[166-170]. Possible explanations for this include performance of colonoscopies by experienced endoscopists, a higher baseline ADR and a shorter procedure time in the AI group. Moreover, the use of AI in populations with low risk of adenoma can minimize an increase in ADR. Also adding AI in bowel cancer screening programs where expert colonoscopists usually use distal attachment devices may not leave room for improvement in ADR. As such, results may differ between geographic areas due to differences in populations risk and endoscopists behaviors. Another retrospective well-designed pragmatic implementation trial also showed negative results for CADe during a 3-month trial[171]. Quality of mucosal exposure by endoscopists or their behavior regarding CADe could have altered the results in this real-world trial. A recent RCT from Kuwait, one of the few from the Middle East, found a non-significant increase in the ADR (relative risk = 1.26, 0.80-2.00) in the CADe group[172]. The sample size was too small to build solid conclusions. Finally, a recent large United States multicenter RCT revealed a 17% higher APC detection rate in AI-assisted colonoscopy compared with conventional high-definition colonoscopy. Nevertheless, no significant differences were found in adenoma, advanced adenoma and SSL detection rates[173]. The exclusion of trainees and less experienced endoscopists might have contributed to the latter findings.

It’s important to mention the impact of Hawthorne effect which implies that endoscopists might improve their performance when they know they are being monitored, thus reducing the effect of AI. Also, the absence of uniformity in AI systems between different studies may alter ADR results. Lastly, excessive false alarms can desensitize endoscopists who might ignore AI. Further large scale, multicenter and multinational real-world studies are needed to compare the effect of combining CADe with mucosal exposure techniques on ADR in less experienced endoscopists.

In summary, CADe has shown encouraging results in controlled trials, mainly in the detection of diminutive and small polyps. Its role in a real-world setting remains uncertain. The adenoma miss rate (AMR) can be more sensitive than the ADR to assess differences in lesion detectability even among endoscopists with a high ADR[174]. Multiple tandem RCTs comparing CADe with standard colonoscopy showed a significantly lower AMR in the CADe group[175-178], and this was significant only for diminutive and small polyps. Some of these RCTs also showed a significant decrease in SSLs miss rate[176,177]. A systematic review and meta-analysis of 14 RCTs suggested that CADe assisted colonoscopy was associated with a 65% reduction in AMR, a 78% reduction in the SSL miss rate, a 52% increase in ADR and a 93% increase in the number of adenomas > 10 mm[179].

Effect of CADe assisted colonoscopy on the detection of sessile serrated adenomas and LST: Ahmad et al[180] showed that their AI system was able to detect 80% of “subtle lesions”, outperforming experienced and less experienced endoscopists who detected only 37% and 11% of them, respectively. Lin et al[181] recently studied the detection performance of a deep neural network algorithm on LSTs. The accuracy of the model in identifying LSTs exceeded 99%, significantly better than that of novice endoscopists and similar to that of expert ones.

Zhou et al[182] noticed that the per-frame sensitivity for non-granular LSTs and small sessile serrated adenomas/polyps should be further improved. A recent 3-arm RCT showed that AI could increase the SSL and advanced adenoma detection rates when compared with standard high-definition colonoscopy. Further, the combination of endo cuff with AI further improves SSLs detection rate compared to AI alone[183].

In summary, most trials did not show an increase in the detection of advanced adenomas and SSLs. In recent studies, CADe seems promising for the detection of subtle and advanced lesions, but data is still limited, and further studies are needed.

CADx in colonoscopy

While the real-time prediction of histology was based on virtual chromoendoscopy that enhances mucosal and vascular patterns, there has been growing interest in developing CADx systems to better characterize colonic polyps and differentiate them into neoplastic and non-neoplastic. In this context, the American Society for Gastrointestinal Endoscopy (ASGE) published standards for the performance thresholds required for any endoscopic technology to support a resect and discard (90% or greater agreement with histopathology for post-polypectomy surveillance intervals) and diagnose and leave (90% or greater NPV for adenomatous histology) strategy for diminutive polyps[184]. To note that these thresholds apply only when the endoscopic technology is used with “high confidence”. Modern CADx models utilize CNN with standard NBI or WLI and with high performance (Tables 8 and 9).

Table 8 Summary of retrospective studies on computer-aided diagnosis for characterization of colorectal polyps[185,188-192,195,196,229,236-238].

Ref.	Country	AI system	Number	WLI/IEE	Primary outcomes	Results
Tischendorf et al[236]	Germany	SVM	209 polyps	Magnifying NBI	Adenomas vs non adenomas	Se: 90%; Sp: 70%; Acc: 85%
Takemura et al[237]	Japan	SVM	1519 polyps, 371 images	Magnifying NBI	Neoplastic vs non neoplastic	Se: 97%; Sp: 97%; Acc: 97%
Misawa et al[195]	Japan	EndoBRAIN	1179 images	Endocytoscopy, NBI	Prediction of histology	Se: 84%; Sp: 97%; Acc: 90%; PPV: 98%; NPV: 82%
Komeda et al[188]	Japan	CNN	1800 images, 10 videos	WLI	Adenomas vs non adenomas	Acc: 75%
Byrne et al[185]	Canada	DCNN	117000 images, 388 videos	NBI	Adenomas vs non adenomas (DP)	Se: 98%; Sp: 83%; PPV: 90%; NPV: 97%; Acc: 94%
Chen et al[238]	China	DCNN	2441 images	NBI	Neoplastic vs hyperplastic (DP)	Se: 96%; Sp: 78%; PPV: 89%; NPV: 90%; Acc: 90%
Sánchez-Montes et al[189]	Spain	ML	225 polyps	HD WLI	Adenomas vs non adenomas (DP)	Se: 92%; NPV: 91%; Acc: 90%; DP: NPV: 96%; Acc: 87%
Song et al[190]	Korea	ENAD	12480 images, 451 polyps	NBI	Adenomas vs SSLs	Se: 82% vs 84%; Sp: 93% vs 88%; Acc: 81% vs 82%
Kudo et al[196]	Japan	EndoBRAIN-EYE	69142 images	Endocytoscopy, NBI	Neoplastic vs non neoplastic	Se: 96%; Sp: 94%; PPV: 96%; NPV: 94%; Acc: 96%
Zachariah et al[192]	United States	CAD-EYE	5912 images	WLI, NBI	Diminutive adenomas vs SSLs/hyperplastic polyps	Se: 96%; Sp: 90%; PPV: 94%; NPV: 93%; Acc: 94%
Jin et al[191]	Korea	DCNN	2450 polyps	NBI	Adenomas vs non adenomas (DP)	Se: 83%; Sp: 91%; Acc: 87%
Ozawa et al[229]	Japan	DCNN (SSD)	27508 images	WLI, NBI	Adenomas vs non adenomas (DP)	WLI: NPV: 85%; NBI: NPV: 91%

AI: Artificial intelligence; IEE: Image-enhanced endoscopy; WLI: White light imaging; CNN: Convolutional neural network; SVM: Support vector machine; DCNN: Deep convolutional neural network; ML: Machine learning; ENAD: Endoscopy neural assisted detection; CAD: Computer assisted device; SSD: Single shot multi-box detector; NBI: Narrow band imaging; HD: High definition; DP: Diminutive polyps; SSL: Sessile serrated lesion; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; PPV: Positive predictive value; NPV: Negative predictive value.

Table 9 Summary of prospective studies on computer-aided diagnosis for the characterization of colorectal polyps[194,200,201,208,209,239-243].

Ref.	Country	AI system	Number	WLI/IEE	Primary outcomes	Results
Gross et al[239]	Germany	SVM	434 small polyps (< 10 mm)	Magnifying NBI	Adenomas vs non adenomas (small polyps)	Se: 95%; Sp: 90%; PPV: 93%; NPV: 92%; Acc: 93%
Kominami et al[240]	Japan	SVM	1262 polyps, 118 images	Magnifying NBI	Adenomas vs non adenomas (small polyps)	Se: 93%; Sp: 95%; PPV: 95%; NPV: 93%; Acc: 97%
Mori et al[194]	Japan	EndoBRAIN	61952 images, 466 polyps	Endocytoscopy, NBI	Adenomas vs non adenomas (DP)	Se: 95%; Sp: 92%; PPV: 96%; NPV: 96%; Acc: 98%
Barua et al[208]	Norway, United Kingdom, Japan	EndoBRAIN	892 polyps	WLI, NBI, magnification	Neoplastic vs non neoplastic polyps (DP)	Se: 90%; Sp: 86%
Hassan et al[242]	Italy	GI genius	544 polyps	NBI, BLI	Adenomas vs non adenomas (DP)	NPV: 98%; Se: 82%; Sp: 93%; Acc: 92%
Minegishi et al[200]	Japan	EndoBRAIN	395 polyps	NBI	Adenomas vs non adenomas (DP)	NPV: 94%; Se: 94%; Sp: 63%; Acc: 86%
Rondonotti et al[209]	Italy	CAD EYE	596 polyps	WLI	Adenomas vs non adenomas (DP)	NPV: 91%; Se: 89%; Sp: 88%; Acc: 88%
Li et al[241]	Singapore (multicenter)	CAD EYE	661 polyps	WLI	Neoplastic vs non neoplastic polyps	Se: 62%; Acc: 72%
Hassan et al[242]	Italy	CAD EYE; GI genius (head-to-head comparison)	319 polyps	WLI, BLI	Adenomas vs non adenomas (DP)	NPV: 97% vs 98%; Se: 82% vs 86%; Sp: 92% vs 94%
Houwen et al[201]	Netherlands (multicenter) + Spain (1 center)	POLAR	423 polyps	NBI	Neoplastic (adenomas and SSLs) vs non neoplastic	Se: 89%; Sp: 38%; Acc: 79%
Rex et al[243]	United States (multicenter)	GI genius	2695 polyps	WLI, NBI	Adenomas vs non adenomas	Se: 91%; Sp: 65%

IEE: Image-enhanced endoscopy; SVM: Support vector machine; CAD: Computer assisted device; GI: Gastrointestinal; WLI: White light imaging; NBI: Narrow-band imaging; BLI: Blue-laser imaging; DP: Diminutive polyps; SSL: Sessile serrated lesion; Acc: Accuracy; Se: Sensitivity; Sp: Specificity; PPV: Positive predictive value; NPV: Negative predictive value.

Effect of CADx on the characterization of polyps into neoplastic vs non-neoplastic: A deep CNN model was evaluated on 106 diminutive polyps. Its accuracy was 94%, the sensitivity for identification of adenomas was 98%, the specificity was 83%, the NPV was 97%, and the positive predictive value (PPV) was 90%. This was the first system to exceed the ASGE threshold requirements[185]. The findings were corroborated by two other studies[186,187]. However, other studies testing CADx with white light failed to demonstrate high accuracy[188,189].

Some studies on CADx assessed its performance depending on the level of experience of endoscopists. A DL CADx model was developed to classify lesions into 3 histological groups: Serrated polyp, benign adenoma/mucosal or superficial submucosal cancer, and deep submucosal cancer. The overall diagnostic accuracy of the model was higher than that of trainees and comparable to experts in NBI. The performance of the trainees improved with CADx assistance[190]. These results were corroborated by another study showing that the overall accuracy of novice endoscopists increased with CADx assistance[191], and by another study that used a CNN-based optical pathology model with both NBI and white light[192].

In another study, a CAD software was developed using densely connected CNNs based on the modified Sano classification. The CAD software was trained with NBI images and tested with separate sets of images using NBI and BLI. The CAD software had an AUC of 94.3% for the internal set and 84.5% and 90.3% for the external sets (NBI and BLI, respectively). This model is unique since it achieved AUCs comparable with experts and similar results with NBI and BLI, although BLI was not part of the training set[193].

In addition, many authors were interested in combining AI with endocytoscopy (EC) for polyp characterization. CAD-EC systems could be helpful for untrained endoscopists. The model achieved high sensitivity and accuracy in diagnosing neoplastic changes, reaching those of expert endoscopists and significantly outperforming those of trainees[194]. Furthermore, EC exceeded the proposed thresholds required for “diagnose and leave’’ strategy and had better outcomes than trainees and experts[186,195,196].

Two recent meta-analyses aimed to explore the benefits and harms of CADx assistance in the optical diagnosis of colorectal polyps[197,198]. The outcome measures for benefit and for harm with use of CADx were the proportion of polyps predicted to be nonneoplastic that could be left in place and the proportion of neoplastic polyps that would be left in situ due to an incorrect diagnosis. In the first meta-analysis, CADx did not show harm or benefit in the “diagnose and leave’’ strategy. The second meta-analysis did not show significant difference for the proportion of polyps that would not be removed, but the proportion of incorrectly predicted neoplastic polyps was lower with the CADx-assisted strategy than with the CADx-unassisted strategy.

It’s worth mentioning that to date there’s one RCT about CADx, comparing autonomous AI to AI-assisted human (AI-H) optical diagnosis of colorectal polyps. Autonomous AI had significantly higher agreement with pathology-based surveillance intervals compared to AI-H (91.5% vs 82.1%, P = 0.016)[199].

Few studies evaluated CADx in the optical diagnosis of SSLs. Minegishi et al[200] prospectively evaluated the performance of an NBI-CAD system in a real-world setting. Its sensitivity for neoplastic lesions, including SSL, was 93%, while the specificity was 61.5% and NPV was 94%. Concordance rates with American, European, and Japanese guidelines were 90%, 93%, and 98%, respectively. A subgroup-analysis including polyps > 10 mm showed sensitivity and specificity for SSLs of 81% and 62%, respectively. A multicenter prospective study in Spain and The Netherlands showed that their CADx system had a low performance in distinguishing neoplastic (adenomas and SSLs) from non-neoplastic lesions (hyperplastic polyps)[201]. In conclusion, most CADx systems classify polyps as either adenomatous or non-adenomatous but still cannot accurately differentiate SSLs from hyperplastic polyps[202].

CADx in the characterization of early CRC: CADx was also investigated in the prediction of early CRC depth, mainly in Japan. Takeda et al[203] developed an EC coupled CADx system that showed an accuracy of 94% in diagnosing invasive cancer, which increased to 99% in high-confidence diagnosis group. A CADx system with magnifying NBI diagnosed invasive cancers with 84% sensitivity, 83% specificity, 53% PPV, 96% NPV, and 83% accuracy[204]. Another model that used white light endoscopy only for the prediction of depth of invasion of CRC had an accuracy of 90%, similar to that in experts, and it assisted trainees to reach experts’ level[205].

In summary, CADe may increase the detection of diminutive polyps and non-advanced adenomas[206,207], and it remains unclear if AI can increase detection of subtle lesions and advanced adenomas such as SSLs and LSTs. Therefore, the exact role of CADe in decreasing the incidence of interval CRC and its related mortality is yet to be demonstrated. Further, CADx systems did not clearly show benefits over endoscopists’ optical diagnosis. For the leave-in situ strategy, CADx did not increase the proportion of non-neoplastic polyps identified for which polypectomy could be avoided. Furthermore, it slightly decreased the misclassification of neoplastic polyps as non-neoplastic. CADx may be beneficial for less experienced endoscopists in optical histology diagnosis. Nevertheless, for expert endoscopists, it increased the proportion of polyps where the diagnosis was made with high confidence[208,209].

Man-AI collaboration is crucial and endoscopists trust their CADe system more when it generates less false-positive results, avoiding unnecessary polypectomies. However, determining a false-positive with CADe model is relatively less challenging than identifying a misdiagnosis with CADx, especially when it comes to novice endoscopists[210,211]. Also, endoscopists must ensure optimal mucosal exposure to the AI system to function effectively. More large-scale RCTs and real-world prospective studies are needed and should focus on patients’ outcomes rather than just polyp detection.

In conclusion, AI is a promising technology that can revolutionize colonoscopy. RCTs and meta-analyses have demonstrated statistically significant benefits with AI regarding ADR, APC, AMR and SSLs miss rate, although these outcomes have been inconsistent. Moreover, AI might be useful in real time optical diagnosis of diminutive colorectal polyps. Clinical practice guidelines on CADe-assisted colonoscopy, published in April 2025 by the American Gastroenterological Association[212] don’t recommend for or against the use of CADe systems in colonoscopy. They state that CADe systems predominantly increase the detection of low-risk polyps, which may result in more frequent and costly follow-up colonoscopies with uncertain benefits in preventing cancer.

AI IN CLINICAL PRACTICE

AI provides several advantages in the detection and characterization of premalignant and early malignant lesions of the GI tract. Those include improved lesion detection particularly for trainees, enhanced optical diagnosis and disease pattern recognition, improved decision making, reading time-reduction and ability to spend more time with patients. The anticipated benefit of implementing these advancements is to eventually aid in cancer prevention and consequently reduce the cost of healthcare due to decreased treatment expenses. This end result is achieved when increased detection through CADe and surveillance endoscopies are complemented with CADx. This approach allows for less unnecessary biopsies, a resect and discard or a leave in situ strategy when appropriate, or definitive treatment allocation as indicated. Despite these benefits, the growing integration of AI technologies in the medical field is accompanied by multiple challenges including data standardization, overfitting and generalizability of AI models, as well as a multitude of legal and ethical issues. Furthermore, cost effectiveness of AI models, and physicians’ attitudes towards adopting such technology are potential hurdles. When it comes to AI implementation in gastroenterology training programs, the real question is not about using AI or not, but rather about how and when to use it. AI gives trainees an opportunity and a new modality of real-time teaching assistance and an instant feedback system. It has shown promising results so far. On the other hand, overreliance on AI, cognitive overload and distractibility are concerns that should be addressed when AI-assistance is applied in a fellowship program. Some studies proposed a model for AI implementation process and timing during fellowship training. Fellowship programs should ensure that trainees are aware of how AI functions, the ethics of using AI, as well as the potential harm. They should provide feedback on their performance with frequent reassessment. We believe that for AI to augment medical practice sustainably, different parties including medical professionals, policymakers, AI researchers and stakeholders should be involved in the construction and implementation process. AI can reshape the medical landscape when it is positioned along a suitable infrastructure primed for this change with adequate training and resources.

FUTURE RESEARCH

Future research should focus on direct comparisons between CADe and CADx systems on one hand, and expert endoscopists on the other hand, particularly in real-time settings. This could lead to better understanding of their complementary roles and limitations. Outcome-based studies evaluating the long-term clinical impact of AI-assisted endoscopy on cancer detection rates, patient survival, and cost-effectiveness are also essential, although they require extended follow-up and large-scale multicenter collaboration. Additionally, the integration of AI across various imaging modalities, the development of standardized performance benchmarks, and the validation of AI tools in diverse patient populations and clinical environments remain critical areas for future investigation. Exploring AI’s potential in training and decision support may also help bridge the gap between novice and expert performance, enhancing overall endoscopic practice.

CONCLUSION

The field of gastroenterology and specifically GI endoscopy provided an appealing foundation for AI-assisted image analysis research. The evidence from these studies varies between the upper GI tract, SB and colon. Evidence supporting AI-assisted detection and characterization of premalignant and early cancerous lesions in the esophagus and stomach is the most consistent and promising. For the colon, AI was particularly superior in the detection of diminutive lesions. The long-term outcomes from this increase in sensitivity is yet to be explored. On the other hand, there is insufficient evidence on the role of AI-assisted SB endoscopy and further high-quality studies are required.