Artificial intelligence-assisted endoscopy in the detection of early gastrointestinal cancer: Progress, challenges, and future directions

Abstract

Gastrointestinal (GI) cancers are a leading cause of cancer-related death, and early diagnosis is crucial for improving patient outcomes. Traditional endoscopy, while essential, depends on the skill of endoscopists and is prone to errors. Recent advancements in artificial intelligence (AI), particularly deep learning with convolutional neural networks, have shown promise in enhancing the early detection of GI cancers. This review highlights the role of AI-assisted endoscopic technologies in the detection, localization, and diagnosis of GI cancers across various sites, including the oral cavity, pharynx, esophagus, stomach, small intestine, colon, and anal canal. AI-powered systems, such as computer-aided detection and diagnosis, have significantly improved adenoma detection rates and lesion characterization, aiding clinical decision-making. Integrating AI with advanced endoscopic techniques like narrow-band imaging, magnifying endoscopy, and capsule endoscopy has enhanced diagnostic accuracy. Despite these advances, challenges remain, including model generalization, data quality, and the need for efficient human-AI collaboration. Regulatory approval, legal concerns, and integration into clinical workflows also pose barriers to widespread adoption. Future developments in multimodal data fusion, edge computing, and AI-augmented reality are expected to improve the precision and accessibility of AI-assisted endoscopy for early GI cancer screening.

Key Words: Artificial intelligence; Gastrointestinal endoscopy; Early cancer detection; Deep learning; Computer-aided diagnosis

Core Tip: Artificial intelligence (AI)-assisted endoscopy technologies have significantly advanced early detection and diagnosis of gastrointestinal cancers, enhancing adenoma detection rates and improving clinical outcomes. Recent studies demonstrate that AI, through deep learning models, can effectively identify small lesions, reduce missed diagnoses, and assist in clinical decision-making across various gastrointestinal regions, including the esophagus, stomach, and colon. However, challenges such as data quality, model generalization, and physician-AI collaboration remain. Overcoming these issues will ensure AI’s broader clinical integration, making it a vital tool in precision medicine and early cancer screening.

INTRODUCTION

Gastrointestinal (GI) cancers remain a major cause of cancer-related mortality worldwide, and early diagnosis is critical for effective treatment[1]. GI malignancies including those of the stomach, small intestine, esophagus, rectum, and colon pose a serious threat to global health. According to the World Health Organization, oral cancer alone caused approximately 10.3 million deaths globally in 2020[2]. Early detection is therefore essential for improving patient survival outcomes. Unfortunately, because early-stage GI cancers often lack specific symptoms, they are frequently diagnosed at advanced stages, underscoring the importance of early screening and timely diagnosis as key strategies in cancer prevention and control.

Endoscopy, as a direct visualization method for the GI tract, plays an indispensable role in early cancer detection. For instance, colonoscopy has demonstrated high sensitivity in colorectal cancer screening. Conventional white-light endoscopy (WLE) remains the standard diagnostic technique, but its accuracy depends heavily on the endoscopist’s expertise and subjective judgment. This reliance introduces risks of missed diagnoses, particularly for small or subtle lesions, and can be exacerbated by operator fatigue and procedural duration. For example, the diagnostic sensitivity of endoscopy for atrophic gastritis has been reported to be only 42%[3]. Endoscopic ultrasonography is an effective diagnostic modality for gastric subepithelial tumors but differentiating between GI stromal tumors and benign mesenchymal lesions remains challenging[4].

In recent years, artificial intelligence (AI) has achieved major breakthroughs in medical imaging, particularly in image processing and pattern recognition, providing new opportunities to enhance diagnostic accuracy, speed, and automation. Deep learning, especially convolutional neural networks (CNNs), has been widely applied to medical image analysis[5]. CNNs can automatically extract hierarchical visual features from images, significantly improving computer-aided diagnosis (CADx) capabilities. AI-based analysis accelerates interpretation, reduces human bias, and provides consistent diagnostic outputs, with particular advantages in detecting complex or minute lesions.

METHODOLOGY

A comprehensive literature search was conducted to identify studies related to the application of AI-assisted endoscopy in the early detection of GI cancers. The search was performed in databases such as PubMed, IEEE Xplore, and Google Scholar, using keywords like “AI-assisted endoscopy”, “early gastrointestinal cancer detection”, “deep learning in endoscopy”, and “gastrointestinal imaging”, combined with appropriate Boolean operators. The search was limited to articles published between 2010 and 2025, and only studies published in English were included.

The inclusion criteria were as follows: First, original research was prioritized, focusing on AI applications in GI endoscopy, particularly studies involving the early detection of cancers such as gastric cancer, colorectal cancer, esophageal cancer, small bowel diseases, and oral cancers. Second, the studies had to involve human participants. Third, the studies must provide detailed methodologies of AI algorithms and their clinical applications or validation data, particularly those focusing on real-time diagnosis, computer-aided detection (CADe), and CADx. Additionally, high-quality review articles were also included, provided they offered a comprehensive summary of the integration of AI with endoscopic technologies, thereby providing background and framework for further research in the field.

Exclusion criteria included studies that did not focus on AI applications in endoscopy and those lacking clinical validation or performance evaluation. Studies that did not provide sufficient clinical data or experimental evidence to support the practical application of AI algorithms were excluded. Animal studies, in vitro studies, and papers that primarily involved theoretical analysis were also excluded.

The process of literature identification and selection followed a systematic approach. After initial retrieval, studies were screened for relevance based on their subject matter, and full-text reviews were conducted to confirm compliance with the inclusion criteria. Ultimately, studies that significantly contributed to the understanding of AI’s role in early GI cancer detection were selected, ensuring that this review is both academically rigorous and effectively supports the discussions.

TECHNICAL FOUNDATIONS

Principles and modalities of GI endoscopy

WLE remains the most commonly used method in GI examination. It operates by illuminating mucosal tissue with visible light and capturing reflected signals to produce detailed surface images. While WLE enables direct visualization of mucosal structure and color changes, its sensitivity for early lesions is limited due to minimal contrast between abnormal and surrounding normal tissues. To overcome these shortcomings, magnifying endoscopy (ME) enhances visualization by providing optical magnification of up to several dozen times. When combined with chromoendoscopy techniques using dyes such as indigo carmine or methylene blue ME allows detailed observation of mucosal microstructures and vascular patterns, thereby improving the detection of minute lesions[6]. The principle of chromoendoscopy lies in the selective absorption of dyes by different tissue components, which enhances contrast and facilitates distinction between neoplastic and non-neoplastic mucosa. In comparison, narrow-band imaging (NBI) utilizes specific narrow-band wavelengths (415 nm and 540 nm) to enhance contrast in superficial vascular networks and mucosal surface patterns. NBI has proven particularly effective in the detection of early neoplastic changes[7], and multiple clinical studies have demonstrated its superior sensitivity and specificity for early esophageal squamous cell carcinoma and gastric cancer relative to conventional WLE. Beyond these modalities, blue laser imaging and laser endoscopy represent emerging technologies that refine wavelength emission and illumination methods to provide higher contrast and resolution. These optical innovations offer improved visualization of mucosal microarchitecture and serve as rich data sources for the development of AI-assisted systems (Figure 1).

Figure 1  Principles of gastrointestinal endoscopy modalities.

Principles and algorithms of AI-assisted endoscopy

Computer vision technologies form the foundation of AI-assisted endoscopy. The core function of these systems is to simulate human visual perception extracting diagnostic features from two-dimensional endoscopic images and classifying lesions.

Early CADe systems were limited by manually engineered features such as color histograms and texture descriptors, resulting in poor generalization across different clinical settings. The emergence of deep learning, and particularly CNNs, has overcome these limitations. CNNs can automatically learn hierarchical features from raw image data, eliminating the need for manual feature design and dramatically improving detection performance. Networks such as visual geometry group network, residual network, and dense convolutional network have shown excellent results in early detection of gastric and colorectal cancers, achieving diagnostic accuracy comparable to expert endoscopists (Table 1). From a practical perspective, the workflow of an AI-assisted endoscopy system typically comprises four key stages. First, image acquisition is performed, during which the endoscopic device captures real-time static images or dynamic video sequences of the GI tract. Second, in the preprocessing stage, the system enhances image quality through operations such as denoising, color normalization, and distortion correction. Third, during feature extraction, CNNs automatically identify and learn multi-level features from low-level textures to high-level semantic representations. Finally, classification and output are achieved using a Softmax or other classifiers to generate lesion probability scores, thereby providing clinicians with precise, data-driven diagnostic support.

Table 1 Explanation of deep convolutional neural network architectures.

Architecture name	Core innovation/structural features	Relevance in medical field
VGGNet	Adopts a concise structure of “stacked small convolutional kernels (3 × 3) + pooling layers”, enhancing feature extraction capability by increasing network depth	A classic model for basic feature extraction in medical images, suitable for preliminary lesion detection and medical image classification (e.g., X-ray disease screening), laying the foundation for subsequent architectures in medical AI
ResNet	Introduces “residual connections” (cross-layer feature transmission) to solve the gradient vanishing problem in deep network training, enabling the construction of ultra-deep networks	Significantly improves feature extraction accuracy for complex medical images, applicable to pathological section analysis and 3D medical image segmentation (e.g., tumor boundary extraction), serving as a core architecture for disease diagnosis models
DenseNet	Employs “dense connections” (direct feature sharing across all layers) to enhance feature propagation efficiency and reduce parameter redundancy	Excels in fine-grained analysis of medical images, such as micro-lesion recognition and multi-modal medical image fusion (e.g., combining CT and MRI images), demonstrating distinct advantages in precision medical diagnosis

VGGNet: Visual geometry group network; ResNet: Residual network; DenseNet: Dense convolutional network; AI: Artificial intelligence; 3D: Three dimensional; CT: Computed tomography; MRI: Magnetic resonance imaging.

EARLY GI CANCER AI PROGRESS

Technological advances

The technological advancements in AI-assisted endoscopic diagnosis are mainly reflected across several dimensions, including image preprocessing and enhancement, CADe systems, and CADx systems. Through deep learning models, these technologies have significantly improved the efficiency and accuracy of endoscopic diagnosis. In terms of image preprocessing and enhancement, deep learning models effectively address the limitations of traditional endoscopic image quality. For instance, image denoising and dehazing models can remove artifacts caused by uneven illumination or mucus interference while preserving lesion details[8]. Furthermore, AI-driven super-resolution reconstruction techniques can transform low-resolution endoscopic videos into high-resolution images, making subtle mucosal textures and lesion boundaries clearer. This not only optimizes the clinician’s visual experience but also provides high-quality input data for subsequent lesion detection, thereby improving overall diagnostic sensitivity[9]. CADe systems represent one of the most mature applications of AI in endoscopy. Utilizing CNNs and transformer-based models, CADe systems analyze real-time endoscopic video streams to automatically identify and highlight suspicious areas such as polyps, adenomas, and early cancers. These suspicious regions are typically displayed using bounding boxes or heatmaps to alert the endoscopist. This real-time prompting mechanism helps physicians maintain focus and reduces the risk of missed diagnoses caused by fatigue or inattention particularly effective in detecting small or flat lesions. Clinical studies confirm CADe systems significantly boost adenoma detection rate (ADR), supporting community hospital and less experienced endoscopists. Specifically, AI-assisted colonoscopy increases ADR from 35% to 43%, with multiple studies validating its efficacy[10]. Once a lesion is detected, the CADx system further enables lesion characterization and risk assessment. Deep learning-based CADx models can not only distinguish between histological types of polyps such as adenomatous vs hyperplastic but also predict the depth of invasion in esophageal or gastric cancers. Some studies have even explored AI-assisted grading of precancerous lesion risk, allowing endoscopists to make real-time decisions on resection or surveillance strategies during examination. The core value of CADx lies in its decision-support capability, transforming AI from a simple “detection tool” into a “diagnostic assistant” that provides clinicians with actionable insights throughout the diagnostic process[11].

Clinical applications in the GI tract

Early detection of oral lesions: Oral cancer, particularly oral squamous cell carcinoma (OSCC), is a major component of GI tumors, with great clinical significance for early screening and prevention. OSCC is the most common subtype of oral cancer, and its global incidence continues to rise[6]. Due to the anatomical complexity of the oral cavity, lesions often remain hidden in regions such as the base of the tongue, buccal mucosa, and palate. Conventional visual examination has limited sensitivity for early detection, leading to a considerable proportion of patients being diagnosed at middle or advanced stages, severely impacting prognosis and survival rates[12].

Recent advancements in AI technologies have opened new possibilities for the early diagnosis and screening of oral cancer[7,13]. Deep learning-based oral endoscopic image recognition systems have matured substantially in recent years. These systems can automatically identify suspicious lesion areas under both WLE and NBI modes, significantly improving the sensitivity and specificity of early cancer detection by clinicians. For example, CNNs have demonstrated outstanding performance in analyzing oral histopathological and endoscopic images, effectively identifying OSCC and its precancerous lesions, such as oral submucous fibrosis and leukoplakia[14-16]. NBI technology, which enhances contrast in mucosal and submucosal blood vessels using narrow-band light (400-550 nm), allows benign, precancerous, and malignant lesions to be distinguished more clearly[17,18].

A recent systematic review and meta-analysis indicated that NBI has high diagnostic accuracy in evaluating the malignant transformation of oral potentially malignant disorders, particularly when applying the intrapapillary capillary loop classification system[19]. NBI-guided surgical resection of oral cancer has also been reported to reduce local recurrence rates and improve patient survival[20]. AI applications in optical imaging technologies like optical coherence tomography have shown substantial potential for identifying malignant entities[21-23]. AI assists clinicians in screening and diagnosing oral cancer, especially in differentiating between normal mucosa, precancerous lesions, and carcinomas with high accuracy[24].

Most studies in this section are retrospective and observational, with relatively small sample sizes. The use of different AI algorithms and imaging devices across studies introduces some variability in results. While the majority of studies report positive findings, the heterogeneity in datasets and equipment limits the generalizability of the conclusions. Further large-scale, multicenter studies are needed to validate these findings and address these limitations.

Early detection of pharyngeal lesions: The pharynx, as the intersection of the digestive and respiratory tracts, holds significant clinical importance in the early screening of GI tumors[25]. Pharyngeal cancers, particularly laryngeal and hypopharyngeal carcinomas, often present with nonspecific early symptoms such as mild sore throat, foreign-body sensation, or hoarseness that can easily be overlooked by both patients and clinicians. Consequently, most cases are diagnosed at advanced stages, severely impacting prognosis and survival outcomes[26].

In recent years, AI has gained growing attention in the early diagnosis of pharyngeal cancers. Researchers have integrated acoustic signals and endoscopic image data to develop multimodal diagnostic models capable of capturing correlations between vocal changes and mucosal abnormalities[27]. Additionally, deep learning-based video stream analysis has introduced new approaches for real-time lesion recognition, enabling automatic segmentation and localization of suspicious regions during dynamic examinations[28].

AI-assisted diagnostic systems that integrate multimodal data including laryngoscopic images, NBI, white-light images, endoscopic biopsy images, and patient voice characteristics have demonstrated high diagnostic performance for detecting laryngopharyngeal cancers. One such system achieved 90.0% sensitivity, 92.3% specificity, and 91.7% overall accuracy[25]. AI models have shown remarkable results, including achieving 93.3% sensitivity for detecting hypopharyngeal cancer[25]. AI applications in this field are advancing, and large-scale datasets will be essential to improving early detection of pharyngeal cancers[29-31].

The research on AI in pharyngeal cancer detection is still in its early stages, with many studies being pilot projects. These studies tend to have small sample sizes and limited follow-up. While AI models show high diagnostic performance, the lack of standardized imaging techniques and equipment diversity poses challenges to comparing results across studies. Larger, more diverse studies are required to confirm the generalizability of these findings.

Early detection of esophageal cancer: Esophageal cancer, including both esophageal squamous cell carcinoma and adenocarcinoma, remains a major global health burden with high incidence and mortality rates[32]. Early diagnosis is crucial to improving patient prognosis, but early-stage esophageal lesions often present with subtle and nonspecific endoscopic findings, such as mild mucosal irregularities, faint erythema, or slight discoloration, that are easily overlooked under conventional WLE[32].

AI-based recognition technologies utilizing NBI and ME have advanced rapidly in recent years. AI systems can automatically analyze microvascular morphology and mucosal surface patterns in endoscopic images, assisting clinicians in accurately identifying early neoplastic lesions[33,34]. Deep learning models like YOLOv5 and RetinaNet have been employed to develop algorithms that combine white-light and NBI data, achieving accurate early diagnosis of esophageal cancer[35].

AI systems have significantly improved diagnostic accuracy for early esophageal cancer and precancerous lesions, especially in Barrett’s esophagus and esophageal adenocarcinoma[36]. Integration of AI with endoscopic ultrasonography has also enhanced the accuracy of predicting tumor invasion depth, which aids in clinical staging and treatment decisions[37]. AI-assisted multi-omics analysis, encompassing genomics, proteomics, and radiomics, helps in personalized treatment planning[30,31]. Countries such as Japan and Germany have already developed and clinically validated AI-assisted systems for early esophageal cancer detection[33].

The studies on AI in esophageal cancer detection are mostly of medium quality, with several studies being observational and lacking randomization. Sample sizes are generally moderate, and there is some heterogeneity in the AI models used. Although AI has shown strong potential in detecting early lesions, further validation with larger and more diverse datasets is needed, particularly for the integration of multi-omics data.

Early detection of gastric cancer: Gastric cancer remains a significant global health challenge, particularly in East Asia, due to its high incidence and mortality rates[38]. Early detection and timely treatment are critical for improving patient outcomes. However, traditional gastroscopy faces significant challenges due to the stomach’s complex anatomy, wide examination field, and variations in operator experience[39].

AI, particularly deep learning algorithms like CNNs, has shown exceptional performance in real-time analysis of gastroscopic videos. AI can automate feature extraction and lesion identification, assisting physicians in detecting suspicious areas more efficiently[40,41]. Additionally, AI models help with lesion characterization, distinguishing malignant, adenomatous, and inflammatory changes[42].

AI-assisted gastroscopy has been shown to achieve significantly higher sensitivity for early gastric cancer detection compared to general endoscopists[43]. In large-scale clinical trials, AI systems not only improved diagnostic accuracy but also reduced lesion recognition time, highlighting their potential in clinical practice[44]. AI applications also aid in optimizing biopsy strategies, further enhancing diagnostic performance[45].

AI-assisted gastroscopy studies mostly rely on retrospective analyses and single-center data. The sample sizes are often small, which may limit the robustness of the findings. There is also variability in the AI models used across studies. While the studies generally show improvements in diagnostic accuracy, the lack of standardized protocols and equipment differences across centers requires more rigorous testing in multi-center trials to confirm these results.

Early detection of small intestinal lesions: Small bowel malignancies, although relatively rare, hold significant clinical importance, especially for evaluating small intestinal polyps, vascular malformations, Crohn’s disease, and small bowel bleeding. Traditional imaging methods have limitations due to the length and anatomical location of the small intestine[46,47].

AI has demonstrated substantial potential in analyzing contrast-enhanced (CE) images. Deep learning algorithms rapidly process large CE datasets to automatically detect and localize lesions, such as bleeding sites, ulcers, polyps, and tumors[48]. This automation reduces physician workload and enhances detection accuracy[49]. Additionally, AI models are increasingly being applied to assess small intestinal motility and evaluate conditions like Crohn’s disease[49,50].

AI-assisted capsule endoscopy systems have been proven to perform at levels comparable to, or even surpassing, experienced clinicians in detecting small bowel lesions[51]. AI tools for inflammatory bowel disease have already improved the detection rates of polypoid and non-polypoid dysplasia[52]. AI’s integration with capsule endoscopy provides new avenues for overcoming barriers in early small bowel malignancy screening[53].

The majority of studies on AI in small bowel cancer detection use CE data, with moderate sample sizes. The effectiveness of AI varies depending on the algorithms and imaging technologies used. While AI has demonstrated high accuracy in detecting lesions, the studies often rely on retrospective data, and more prospective studies with larger sample sizes are needed to validate the performance of AI-assisted CE systems in real-world settings.

Early detection of colorectal lesions: Colorectal cancer is the second leading cause of cancer-related deaths worldwide, and early identification of precancerous lesions is crucial to reducing colorectal cancer incidence and mortality. However, conventional colonoscopy faces challenges in detecting flat adenomas, which are often overlooked by endoscopists[54,55].

AI-driven CADe systems have been developed to analyze real-time colonoscopy video feeds, automatically identifying suspicious lesions and providing visual cues to endoscopists. These systems significantly improve ADR, particularly for small (< 5 mm) adenomas and serrated lesions[10,56].

AI-assisted colonoscopy has been shown to increase ADR from approximately 35% to 43%, with multiple studies confirming its effectiveness in enhancing adenoma detection[10]. Additionally, AI models help with lesion characterization, distinguishing adenomatous from non-adenomatous lesions, and aid in selecting appropriate resection techniques[42]. AI has also proven valuable in optimizing biopsy strategies, increasing positive detection rates and minimizing unnecessary procedures[52].

AI-based CADe systems for colorectal cancer are well-supported by clinical studies, although many studies still rely on retrospective data. The studies show a significant improvement in ADR, but sample sizes are often small, and there is variability in the quality of colonoscopy equipment used. The evidence is promising, but further large-scale, multi-center studies are necessary to fully assess the potential of AI in routine clinical practice.

Early detection of anal canal lesions: Diseases of the anal canal, particularly anal squamous cell carcinoma, are often misdiagnosed as hemorrhoids, leading to delayed treatment and poorer prognoses[57]. The diagnostic accuracy of traditional anoscopy is highly dependent on the operator’s experience and is challenged by the narrow lumen and complex anatomy of the anal canal[58].

AI applications in high-resolution anoscopy (HRA) have demonstrated potential for improving the accuracy of early detection of anal canal lesions. Deep learning models, such as CNNs, are particularly effective for lesion detection and classification[58].

AI-based predictive models have integrated imaging features with human papillomavirus infection status to enhance the accuracy of anal canal cancer risk prediction[59]. AI models have shown promising results in differentiating between high-grade squamous intraepithelial lesions and other conditions, significantly improving early detection and reducing misdiagnosis[59].

The research on AI in anal canal cancer detection is limited, with most studies focusing on HRA data. Sample sizes are often small, and there is considerable variability in imaging techniques and AI algorithms used. Despite positive results, the evidence is still preliminary, and more extensive clinical trials are needed to confirm the diagnostic accuracy of AI models in detecting anal canal lesions, particularly in high-risk populations.

CLINICAL PRACTICE AND VALIDATION STUDIES

With the rapid advancement of AI-assisted endoscopic technologies, an increasing number of studies are promoting the transition of AI systems from laboratory algorithm validation to real-world clinical practice. Accordingly, clinical research and validation efforts now focus on evaluating the diagnostic efficacy, clinical feasibility, operational convenience, and impact on patient outcomes of AI systems during actual endoscopic examinations.

In terms of clinical trial design and methodology, existing studies can generally be categorized into two types. The first type is retrospective studies, which utilize previously collected endoscopic images or video data to validate AI algorithms, primarily assessing their sensitivity, specificity, and consistency with pathological gold standards. Although such studies can rapidly accumulate evidence, they may not fully reflect the complexity and variability of real-world endoscopic procedures.

The second type is prospective randomized controlled trials (RCTs), in which patients are randomly assigned to receive either AI-assisted endoscopy or conventional endoscopic examination. The diagnostic efficacy of AI is then verified by comparing parameters such as ADR, early cancer detection rate, and procedure time between the two groups. For instance, numerous multicenter RCTs have confirmed that AI can significantly improve ADR in colorectal screening.

Across different GI regions, AI has demonstrated strong clinical performance. In the esophagus, AI systems have shown outstanding results in detecting Barrett’s esophagus and early esophageal squamous cell carcinoma, with diagnostic accuracies comparable to expert endoscopists and markedly superior to those of general practitioners. In the stomach, AI assists in identifying subtle mucosal color and structural abnormalities, and clinical validation studies have shown that AI can shorten the learning curve and help less-experienced endoscopists improve their early cancer detection rates. In the colorectum, clinical validation is the most robust and comprehensive to date numerous RCTs have demonstrated that AI significantly enhances ADRs, particularly for flat and sub-5 mm lesions, thus establishing a solid clinical evidence base for integrating AI into routine colorectal cancer screening.

In real-world clinical workflows, the value of AI extends beyond merely improving lesion detection rates it also positively influences operator behavior and decision-making. On one hand, AI can highlight suspicious regions in real time within endoscopic video streams, prompting operators to slow down and inspect more carefully. On the other hand, AI alert systems can effectively reduce missed lesions when endoscopists experience fatigue or lapses in attention. Furthermore, some AI systems are now capable of predicting the benign or malignant nature of detected lesions, providing decision support regarding whether immediate biopsy or treatment is warranted.

From the perspective of clinical validation outcomes and ongoing challenges, the results are highly encouraging. Multiple prospective studies and RCTs have demonstrated that AI-assisted endoscopy can significantly improve detection rates of both early cancers and precancerous lesions, while also reducing missed diagnoses and shortening the endoscopist learning curve, thereby contributing substantially to enhanced diagnostic accuracy and procedural efficiency in clinical endoscopy practice. Following the review of clinical validation evidence for AI-assisted endoscopy across various GI segments, this section provides a comparative summary of representative systems (Table 2) to better illustrate their technical characteristics and validation status. The table systematically categorizes these AI tools across different developmental stages from research to clinical deployment based on target anatomical sites, core functions, key performance metrics, and validation maturity.

Table 2 Summary of comparisons of major artificial intelligence-assisted endoscopy systems.

System name	Target site and function	Key performance metrics	Validation status and characteristics
Deep learning-based endoscopy systems	Esophagus, stomach: Early cancer detection and diagnosis	Sensitivity for early gastric cancer > 90%, specificity > 80%	Mostly in clinical research phase: Validation often involves single-center or retrospective studies; demonstrates potential to match or surpass human experts in specific tasks
		Detection accuracy for early esophageal cancer comparable to expert endoscopists
AI-assisted capsule endoscopy systems	Small bowel: Automatic detection of ulcers, bleeding, polyps, etc.	Sensitivity for small bowel lesions > 95%, specificity > 90%	Validated by multicenter prospective studies: Some systems have received regulatory approval and are in clinical use; aims to address the inefficiency of analyzing large CE image volumes
		Significantly increases reading speed, reducing physician workload by > 70%
CADe colonoscopy systems	Colorectum: Real-time polyp detection (CADe)	Increases adenoma detection rate by an absolute 5%-10%	Some systems approved by FDA, CE, NMPA: Supported by the highest level of evidence (multicenter RCTs); integrated into commercial endoscopy platforms; value is pronounced in community practice settings
		Particularly effective for detecting small polyps (< 5 mm) and flat adenomas
CADx colonoscopy systems	Colorectum: Real-time polyp characterization (CADx)	Accuracy for optical diagnosis of adenomatous polyps > 90%	Some features approved and commercialized: Integrated with CADe systems; aims to provide “see-and-diagnose” capability, reducing unnecessary polypectomies and screening costs
		Enables reliable “diagnose-and-leave” or “resect-and-discard” strategies with > 90% confidence
AI-assisted laryngoscopy/pharyngeal diagnosis systems	Pharynx, larynx: Early cancer detection	Sensitivity for laryngopharyngeal cancer 90%-93%, specificity > 92%	Primarily in prospective research or pilot project phase: Sample sizes are relatively small, but shows great promise for multimodal AI in complex anatomical sites
		Capable of multimodal analysis integrating voice signals and images
AI-assisted high-resolution anoscopy	Anal canal: Detection of HSIL	Shows high accuracy in differentiating HSIL from other conditions	Research is very preliminary and exploratory: Limited sample sizes; a promising tool for screening specific high-risk populations but requires further validation
		Can integrate HPV status for risk prediction

AI: Artificial intelligence; CE: Contrast-enhanced; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; HSIL: High-grade squamous intraepithelial lesion; HPV: Human papillomavirus; FDA: Food and Drug Administration; NMPA: National Medical Products Administration; RCTs: Randomized controlled trials.

AI DETECTION: CORE CHALLENGES

Core challenges in AI detection practice

AI-assisted endoscopic technologies have achieved remarkable progress in the screening of early GI cancers, yet several significant challenges remain. First, there are considerable performance variations among different AI systems, which hinder their global standardization and generalizability.

Data quality and annotation represent another major challenge. AI models rely heavily on high-quality annotated datasets; however, the complexity of endoscopic imagery often leads to inconsistent or inaccurate labeling, directly affecting model training performance. The subjectivity of annotation further exacerbates fluctuations in data quality, reducing reproducibility and reliability across studies.

Additionally, the generalization capability of AI models remains insufficient. Models trained on data from a specific hospital, device, or population often perform inconsistently in different clinical settings, particularly in resource-limited regions. This highlights the urgent need for cross-regional, multicenter datasets to improve model robustness and stability under diverse conditions.

The collaborative integration of AI and physicians also poses an ongoing challenge. Although AI can assist diagnosis, enhancing both efficiency and accuracy, achieving an optimal balance between AI’s automated analysis and clinicians’ experiential judgment remains unresolved. Overreliance on AI may risk diminishing physician autonomy and critical thinking.

Regulatory approval pathways are another critical hurdle for AI-assisted endoscopy. In many regions, AI technologies must undergo rigorous validation and approval processes to meet regulatory standards, which can be time-consuming and costly. Moreover, differences in regulatory frameworks across countries can delay the global adoption of AI systems. The regulatory landscape must evolve to create clear, standardized pathways for AI in medical devices to facilitate their faster and more widespread implementation.

Medical legal considerations also pose challenges for AI integration in clinical practice. The use of AI for diagnostic support raises important legal questions regarding liability and accountability. In the event of a misdiagnosis or treatment error, it is unclear whether the physician, the AI system developer, or the healthcare institution is responsible. Clear guidelines and regulations are needed to address these concerns and ensure that AI applications are legally and ethically sound in clinical environments.

Explainability and interpretability of AI models is an ongoing issue. Deep learning models, while powerful, are often criticized for their “black-box” nature, meaning their decision-making processes are not transparent or easily understood by clinicians. This lack of explainability can reduce physician trust in AI systems, particularly when critical decisions are involved. There is an urgent need for AI systems that not only perform well but also provide clinicians with clear, interpretable explanations of how conclusions were reached, ensuring that physicians can confidently incorporate AI into their decision-making process.

Clinical workflow integration and implementation barriers also pose significant challenges. AI-assisted endoscopy must seamlessly integrate into existing clinical workflows, ensuring that the technology enhances efficiency without disrupting established practices. For instance, AI models should be designed to work with the range of endoscopic equipment already in use and should be able to process data in real-time without causing delays in patient care. Additionally, the introduction of AI tools may require training and support for clinicians to ensure they are able to effectively use the technology. Implementation also involves overcoming financial and logistical challenges, particularly in resource-limited settings where access to advanced technologies may be restricted.

In summary, AI-assisted endoscopy for early GI cancer screening faces multiple challenges, including data quality, model generalization, device compatibility, regulatory approval, legal and ethical considerations, explainability, clinical workflow integration, and human AI collaboration frameworks. Addressing these issues will be essential for enabling AI to play a more integral role in clinical endoscopy, ultimately providing more precise and reliable support for the early detection and diagnosis of GI cancers.

Strategic recommendations for addressing challenges

Enhancing data quality and annotation consistency: To overcome the data quality and annotation challenges in AI-assisted endoscopy, a multi-faceted approach is needed. First, the establishment of standardized annotation guidelines is essential to reduce subjectivity and ensure consistent data labeling across institutions. This could involve collaboration between international expert panels to define common protocols for annotating endoscopic images. In addition, leveraging advanced AI-based pre-annotation tools, followed by expert validation, can significantly streamline the process while minimizing human errors. Moreover, creating large, diverse annotated datasets through multi-center collaborations will help improve the robustness of AI models, ensuring their ability to generalize across various clinical settings and populations. Examples such as the publicly available GI-pathology dataset and other large-scale collaborations can serve as models for future endeavors.

Fostering multi-center collaborations for model generalization: To address the issue of model generalization, particularly in resource-limited settings, it is crucial to promote large-scale multi-center data sharing and collaboration. This would enable AI models to be trained on more diverse datasets, improving their accuracy and applicability across different geographic regions and healthcare systems. The establishment of international research consortia dedicated to AI in medical imaging, similar to efforts like the Radiological Society of North America’s AI initiatives, would accelerate the creation of robust datasets that can better handle variations in patient populations, imaging devices, and clinical workflows. By incorporating data from a wide range of hospitals, regions, and demographics, these models can be more readily adapted to local conditions, ensuring that AI tools can provide equitable healthcare solutions across the globe.

Advancing explainability and human-AI collaboration: To improve the trust and adoption of AI systems in clinical practice, efforts should be directed at enhancing model explainability and fostering better collaboration between AI systems and clinicians. AI models should not only provide accurate diagnostic support but also offer transparent reasoning for their predictions, thereby enabling clinicians to understand and interpret AI-generated recommendations. Developing “explainable AI” frameworks that clarify how the AI arrives at specific conclusions would help bridge the gap between machine output and human decision-making. Furthermore, the integration of AI should not be viewed as a replacement for clinicians, but as a tool to augment their expertise. Efforts should focus on creating workflows that enable AI to assist without undermining physician autonomy. This could involve designing user-friendly interfaces.

AI ENDOSCOPY: FUTURE DIRECTIONS

Multimodal data fusion is a key direction for the future development of AI-assisted endoscopy technologies. The goal is to integrate endoscopic images, pathological images, genomic data, and clinical information to enable precise screening of early GI cancers. AI-driven liquid biopsy techniques, utilizing biomarkers such as circulating tumor DNA, circulating tumor cells, and exosomal RNA, offer non-invasive diagnostic options while emphasizing AI’s role in processing complex data and enhancing detection sensitivity. This cross-modal data integration significantly improves the accuracy and sensitivity of AI in GI endoscopic screening and supports the development of personalized treatment plans.

Real-time intelligent assistance is another important focus for the future of AI-assisted endoscopy. AI models have already been deployed for real-time processing on edge computing devices, enabling the instant detection of gastric neoplasms during endoscopic procedures with high classification accuracy and low inference latency. This real-time feedback system reduces dependency on cloud servers, ensuring data privacy and security, and improving diagnostic efficiency[60]. Future development will focus on algorithm optimization to enhance recognition speed and accuracy, as well as expanding real-time diagnostic capabilities to more GI diseases.

The integration of AI with augmented reality (AR) technology will revolutionize endoscopic procedures. AR navigation systems have been explored in GI endoscopy and laparoscopy, where AR elements are overlaid on real-time endoscopic images to provide structural localization and lesion guidance, enhancing the intuitiveness and safety of the procedure[61]. Although the use of AR in GI endoscopy is still in its early stages, its combination with AI holds the potential to become the core of the next-generation intelligent endoscopic navigation systems. Future research will focus on developing more precise AR navigation algorithms and utilizing AI for real-time image analysis to offer more comprehensive procedural guidance to physicians.

In the area of personalized and precision medicine, AI-assisted endoscopy technologies are expected to integrate patient genomic data, family history, and other clinical indicators to create customized screening and treatment plans for each patient, thereby improving diagnostic accuracy and reducing unnecessary tests and interventions. In pathological diagnostics, multimodal visual-language AI-assisted systems, combining pathological images with natural language interaction, have been developed to enhance diagnostic efficiency and educational outcomes. Future directions will involve developing more sophisticated AI models that analyze multiple data types to provide personalized treatment recommendations for clinicians.

In broader medical fields, multimodal AI has become a growing trend. A significant body of research has explored the application of deep learning in integrating multiple medical modality data, showing that multi-source data integration can greatly improve overall model performance. These studies provide important insights for the development of future AI-assisted endoscopy systems.

CONCLUSION

This review has summarized the latest advancements of AI in the field of early GI cancer detection, spanning from the oral cavity to the pharynx, esophagus, stomach, small intestine, and colon, significantly advancing the progress of early screening, diagnosis, and precision treatment of GI cancers. The focus has been on the potential of AI technologies to enhance the accuracy of early cancer screening, reduce missed diagnoses, and optimize clinical decision-making. In recent years, several prospective, multicenter clinical studies have demonstrated that AI-assisted endoscopic systems based on deep learning can significantly improve ADR and the ability to identify small lesions. The introduction of AI not only enhances screening efficiency but also reduces the risk of missed or incorrect diagnoses in real-world clinical settings. However, AI-assisted endoscopy still faces several constraints in its widespread clinical application. Firstly, data quality and annotation consistency remain prominent issues. Most current AI models are trained on images collected from single-center and homogeneous devices, and their generalization ability may significantly decrease when applied across different regions and devices. Secondly, the performance variability of AI systems across different populations and disease spectra requires further evidence, which will have a key impact on technical standardization and regulatory approval. Additionally, the collaborative model between doctors and AI is still not fully matured. While AI can provide lesion annotations and preliminary diagnostic suggestions during real-time endoscopy, physician experience and comprehensive judgment remain irreplaceable in interpreting complex lesions and formulating treatment strategies. Balancing efficiency improvement with clinical decision safety and accuracy will be a crucial topic for future development. Looking ahead, the development of AI in GI early cancer screening will focus on three key directions: First, further enhancing model generalization by introducing large-scale datasets from multiple centers, devices, and demographics to reduce performance variability. Second, integrating multimodal information, including endoscopic images, histopathological data, genomic and microbiome data, to achieve more precise personalized risk assessments and screening strategies (despite potential uncertainties). Third, accelerating the clinical translation from laboratory validation to real-world applications, ensuring rapid deployment of AI systems in diverse healthcare systems. Predictions suggest that with the continued advancement of standardized data sharing, algorithm optimization, and clinical evidence-based research, AI-assisted endoscopy may become an increasingly important tool for early GI cancer screening in the coming years. In conclusion, despite the existing technical and clinical challenges, the ongoing evolution of AI, coupled with multidisciplinary integration, will drive the comprehensive realization of precision medicine and personalized screening in the field of GI early cancer detection.