Opportunities and challenges of artificial intelligence-assisted endoscopy and high-quality data for esophageal squamous cell carcinoma

Abstract

This review comprehensively summarized the potential of artificial intelligence (AI) in the management of esophageal cancer. It highlighted the significance of AI-assisted endoscopy in Japan where endoscopy is central to both screening and diagnosis. For the clinical adaptation of AI, several challenges remain for its effective translation. The establishment of high-quality clinical databases, such as the National Clinical Database and Japan Endoscopy Database in Japan, which covers almost all cases of esophageal cancer, is essential for validating multimodal AI models. This requires rigorous external validation using diverse datasets, including those from different endoscope manufacturers and image qualities. Furthermore, endoscopists’ skills significantly affect diagnostic accuracy, suggesting that AI should serve as a supportive tool rather than a replacement. Addressing these challenges, along with country-specific legal and ethical considerations, will facilitate the successful integration of multimodal AI into the management of esophageal cancer, particularly in endoscopic diagnosis, and contribute to improved patient outcomes. Although this review focused on Japan as a case study, the challenges and solutions described are broadly applicable to other high-incidence regions.

Key Words: Artificial intelligence; Esophageal cancer; Endoscopy; Deep learning; National database; Clinical translation; Multimodal artificial intelligence

Core Tip: This review detailed how artificial intelligence (AI) mitigates operator dependence in the endoscopic diagnosis of esophageal squamous cell carcinoma by comparing the sensitivity and specificity of innovative deep learning models with those of expert endoscopists. This further highlights the use of large-scale repositories, such as the National Clinical Database and Japan Endoscopy Database, for robust AI training and validation. Multimodal AI using big databases proposes a multi-institutional or multi-vendor AI strategy in Japan. Finally, we outlined future directions for real-time endoscopic support and the integration of clinical outcomes into next-generation AI-driven endoscopy.

INTRODUCTION

Esophageal squamous cell carcinoma (ESCC) is often detected at an advanced stage because of its absence of symptoms. Although the 5-year survival rate for early stage-ESCC exceeds 85%, for advanced disease it is < 20%[1].

Screening endoscopy plays a central role in the early detection of ESCC, particularly in high-risk populations, such as Japanese males with a history of smoking or alcohol consumption[2-4]. A multicenter study reported that linked color imaging, an advanced image-enhanced endoscopy mode, improved the detection rate of upper gastrointestinal neoplasms, highlighting its importance in surveillance[5]. Furthermore, novel image-enhanced technologies, such as texture and color enhancement imaging, are expected to improve detection rates by enhancing the visibility of early esophageal cancer through the optimization of mucosal structure and tone[6-8]. Narrow band imaging (NBI) plays a pivotal role in the detailed examination of the detected lesions. Combining NBI with magnifying observation to analyze the microvascular patterns of the mucosal surface, based on the classification by the Japan Esophageal Society, is an established method for accurately predicting the extent and invasion depth of superficial squamous cell carcinoma[9-11]. Thus, recent endoscopic technologies have significantly contributed to the early diagnosis of esophageal cancer by enhancing both the detection capability in screening and the qualitative diagnostic performance in detailed examinations[12].

In recent years artificial intelligence (AI) techniques, particularly deep learning-based image analysis, have advanced rapidly and stimulated vigorous medical research. The applications of AI in endoscopic diagnosis are expanding and are pivotal for the early detection of esophageal cancer. Therefore, this review aimed to summarize the current status, challenges, and prospects of AI-assisted endoscopy in the management of esophageal cancer. AI offers a practical solution to overcome the long-standing challenges of operator-dependent diagnostic accuracy and the difficulty in disseminating expert-level skills, thereby promising a paradigm shift in ESCC management.

CURRENT STATUS OF AI-ASSISTED ENDOSCOPY IN THE MANAGEMENT OF ESOPHAGEAL CANCER

To clarify the current status of AI in the management of ESCC, we conducted a comprehensive search on PubMed for studies published between 2019 and 2025. The results are classified and summarized in three tables. Table 1 details key clinical studies on diagnostic performance while Table 2 focuses on the technical and algorithmic aspects of AI models. Table 3 provides an overview of recent review articles and meta-analyses.

Table 1 Summary of clinical studies on the diagnosis of esophageal squamous cell carcinoma using artificial intelligence.

Ref.	Year	Primary outcomes	Secondary outcomes	Patients studied	Type of data	Study design
Zhao et al[65]	2019	Feasibility of CAD-based IPCL classification	Observer agreement and pixel-level/Lesion-level accuracy	219 patients (30 with inflammation, 24 with low-grade intraepithelial neoplasia, 165 with early esophageal cancer); 185 lesions selected for analysis	NBI-ME images with histological findings	Retrospective study
Fukuda et al[18]	2020	NBI-based CAD system for ESCC diagnosis	Observer agreement, AI speed, and detection time	Training: 1544 pathological SCC lesions and 458 non-cancer/normal tissues; 354 video clips for testing	NBI/BLI endoscopic images and videos with pathology assessment	Evaluation study using retrospectively collected data
Shimamoto et al[66]	2020	AI system for invasion depth estimation in superficial ESCC	AI vs expert endoscopists on same validation videos	909 patients for training dataset; 102 videos of superficial ESCC cases used for validation	WLI/NBI/BLI endoscopic media with invasion depth pathology	Retrospective training; independent video-based validation
Shiroma et al[67]	2021	AI detection capability of T1 ESCC in EGD videos	Real-time AI-assisted detection: Performance and comparison with endoscopists	8428 images (training set); 144 EGD videos (validation set); 40 patients (validation set)	Retrospective EGD data: WLI/NBI videos and images	Retrospective study
Ikenoyama et al[68]	2021	LVL prediction from non-stained images and cancer risk assessment	Comparison of AI performance with experienced endoscopists	595 patients (6634 training images); 72 patients (667 independent validation images)	WLI/NBI endoscopic images and clinical data (ESCC, HNSCC, Lugol staining)	Retrospective study
Li et al[69]	2021	Early ESCC CAD system and comparison with WLI-based models	Reduced missed diagnoses and unnecessary biopsies	Training: 235 cases (abnormal NM-NBI images), 412 cases (normal images); Validation: 284 cases (202 abnormal, 82 normal)	NM-NBI/WLI images, patient/Lesion data, and histology	4-phase observational retrospective study with endoscopist assessment
Waki et al[70]	2021	AI performance in ESCC detection under simulated oversight	AI vs endoscopists: Sensitivity gain, specificity loss with AI support	Training: 1376 superficial ESCC cases (17336 images); 196 non-cancerous cases (2916 images); Testing: 52 superficial ESCC cases (1459 images); 47 non-cancerous lesions (1168 images)	NBI/BLI/WLI images and videos of superficial ESCC, benign lesions, and normal esophagus	DL-based AI system: Retrospective development with partial prospective validation
Meng et al[71]	2022	CAD performance metrics (AUROC, accuracy, sensitivity, specificity)	CAD vs endoscopists: Diagnostic performance and predictive values	837 patients (training); 323 patients (test)	Non-magnified WLI/NBI images and histology-confirmed SESCC/HGIN data	Retrospective diagnostic accuracy
Tajiri et al[72]	2022	AI-based ESCC vs non-ESCC classification under simulated conditions	AI vs endoscopists: Subgroup accuracy by pathology and lesion size	Training: 1433 superficial ESCC cases (25048 images); 410 non-cancerous esophageal lesions (8557 images); Testing: 123 superficial ESCC cases (3370 images); 107 non-cancerous lesions (2075 images)	ME/non-ME endoscopic images and videos (WLI, NBI, BLI)	DL-AI system development in simulated setting (retrospective/prospective mixed design)
Feng et al[16]	2023	AI diagnostic accuracy for superficial ESCC detection	AI diagnostic support and cancer feature recognition	1283 patients (training); 319 (internal validation); 905 (external validation); total 2507 patients, 9663 images	WLI images (Olympus/Fujifilm) and SESCC-confirmed clinical data	Retrospective diagnostic study (partly prospectively registered)
Yuan et al[22]	2023	Real-time detection and delineation of early ESCC using a novel AI system	Accurate classification of image modalities	Not available	WLI/NBI videos and Lugol-stained ME images with histology-confirmed clinical data	Pilot study (video demonstration)
Wang et al[73]	2023	IPCL-based early ESCC localization and classification	Performance metrics and model comparison	246 patients (2887 ME-BLI images); 81 patients (493 ME-NBI images)	Magnified ME-BLI/NBI images and histology-confirmed ESCC data	Pilot retrospective image collection study
Wang et al[74]	2024	YOLO-HSI integration for early ESCC detection	HSI vs RGB models: Classification of normal, dysplasia, and ESCC	16 patients (7 with esophageal SCC, 9 with squamous dysplasia, 10 healthy controls); training dataset: 1836 images	WLI/NBI images converted to HSI with pathology evaluation	Not clearly described; development and evaluation study
Nakao et al[75]	2024	AI support for non-experts in ESCC detection	ESCC detection rate, observation time, and adverse events	320 patients included in analysis (AI group: 152; control group: 168)	WLI/NBI/Lugol endoscopic media with histopathology and clinical data	Prospective, single-center, exploratory randomized controlled trial
Aoyama et al[20]	2025	AI model for superficial ESCC detection from endoscopic videos	Subgroup analysis and comparison with endoscopist performance	Training data: 280 cases (140 with lesions, 140 without); Test data: 115 cases (57 with lesions, 58 without)	NBI endoscopic videos with histopathological diagnosis	Prospective data, retrospective analysis design
Ma et al[76]	2025	Diagnostic performance for ESN classification	iCLE vs experts and non-experts: Diagnostic performance and support evaluation	2803 patients (for iCLE training/validation); 226 patients (image test); additional patients for video recognition testing (prospective, total N not shown)	pCLE video/image data with histopathological gold standard	Prospective diagnostic study

AI: Artificial intelligence; AUROC: Area under the receiver operating characteristic; BLI: Blue light imaging; CAD: Computer-aided diagnosis; DL: Deep learning; EGD: Esophagogastroduodenoscopy; ESCC: Esophageal squamous cell carcinoma; ESN: Esophageal squamous neoplasm; HGIN: High grade intraepithelial neoplasia; HNSCC: Head and neck squamous cell carcinoma; HSI: Hyperspectral imaging; iCLE: Integrated confocal laser endomicroscopy; IPCL: Intra-papillary capillary loop; ME: Magnifying endoscopy; NBI: Narrow-band imaging; NM: Non-magnifying; pCLE: Probe-based confocal laser endomicroscopy; RGB: Red, green, blue; SCC: Squamous cell carcinoma; SESCC: Superficial esophageal squamous cell carcinoma; WLI: White light endoscopy; YOLO: You only look once.

Table 2 Summary of technical or algorithmic studies on esophageal squamous cell carcinoma using artificial intelligence.

Ref.	Year	Primary outcomes	Secondary outcomes	Patients studied	Type of data	Study design
Everson et al[24]	2019	Abnormal IPCL classification metrics and prediction time	CNN feature visualization, eCAM analysis, and future AI insights	17 individuals (10 with ESCN, 7 with normal esophageal squamous epithelium)	HD ME-NBI videos/images (.png) with matched histopathology	Proof-of-concept study
Zhao et al[65]	2019	Feasibility of CAD-based IPCL classification	Observer agreement and pixel-level/Lesion-level accuracy	219 patients (30 with inflammation, 24 with low-grade intraepithelial neoplasia, 165 with early esophageal cancer); 185 lesions selected for analysis	NBI-ME images with histological findings	Retrospective study
Horie et al[77]	2019	CNN-based AI for esophageal cancer detection	Detection of small/superficial lesions and predictive values	Training: 8428 images from 384 patients; Test: 47 patients with 49 cancer lesions	WLI/NBI endoscopic images with histopathology	Retrospective study
Ohmori et al[17]	2020	Image analysis system for ESCC detection and classification	Comparison of diagnostic performance between AI system and expert endoscopists	135 patients for validation; training data included 804 histologically confirmed superficial esophageal SCC lesions	WLI/NBI images (with/without magnification) and histological data	Retrospective training; prospective external validation
Guo et al[78]	2020	Real-time AI for early ESCC and precancerous lesion detection	Frame-based/Lesion-based sensitivity/specificity and normal video evaluation	Training data: 191 cases of precancerous/early ESCC; 358 cases of non-cancerous lesions; validation data: 100 endoscopic videos from 41 patients with ESCC and 30 normal controls	NBI images/videos with histological confirmation	Development study with retrospective training and delayed validation dataset
Tang et al[79]	2021	Diagnostic performance of the DCNN model	Evaluation of DCNN support and agreement with endoscopists	1078 patients (for training and cross-validation); 243 patients (for independent internal and external validation)	WLI endoscopic images and retrospective clinical data	Multicenter diagnostic retrospective study
Uema et al[80]	2021	CNN system for microvessel classification in superficial ESCC	CNN vs endoscopists in microvessel classification, agreement, and diagnostic support	336 patients with 393 SESCC lesions; 1777 training images; 617 validation images	Trimmed ME-NBI images of SESCC and clinical data	Retrospective single-center study
Liu et al[44]	2022	Development of AI model to detect and delineate early ESCC under WLI endoscopy	Detection and delineation performance	1239 patients (13083 images for training/testing); 262 patients (1479 internal test images); 96 patients (648 external test images)	WLI images with confirmed early ESCC clinical data	Retrospective study
Yuan et al[21]	2022	AI-based IPCL subtype prediction for early ESCC	AI validation, diagnostic support, and comparison with endoscopists	685 patients (training/validation); 176 patients (ER validation dataset)	ME-NBI images and confirmed precancerous/superficial ESCC data	Retrospective multicenter study
Zhao et al[81]	2022	Effectiveness of AI-DEN system for early EC diagnosis	Diagnostic accuracy and speed comparison	300 patients suspected of having esophageal cancer; Training group: 200 patients (148 with early cancer, 52 with benign disease); Test group: 100 patients (92 with early cancer, 8 with benign disease)	NBI-DEN images with pathology results	Retrospective study
Tani et al[31]	2023	Diagnostic accuracy of AI system vs endoscopists	Diagnostic accuracy and safety outcomes	388 patients (registered); 380 patients (underwent endoscopy); 237 lesions (evaluated)	Endoscopic images (WLI, NBI/BLI) and real-time clinical data	Single-center prospective single-arm non-inferiority trial
Zhang et al[45]	2023	Interpretable AI-IDPS for ESCC invasion depth prediction	AI-IDPS vs DL models and endoscopists; trust in AI predictions	581 patients for training; validation using 196 images and 33 continuously collected videos	ME-NBI images/videos with pathology reports and endoscopist feedback	Multicenter retrospective study with crossover validation by endoscopists
Wang et al[74]	2024	YOLO-HSI integration for early ESCC detection	HSI vs RGB models: Classification of normal, dysplasia, and ESCC	16 patients (7 with esophageal SCC, 9 with squamous dysplasia, 10 healthy controls); training dataset: 1836 images	WLI/NBI images converted to HSI with pathology evaluation	Not clearly described; development and evaluation study
Nakao et al[75]	2024	AI support for non-experts in ESCC detection	ESCC detection rate, observation time, and adverse events	320 patients included in analysis (AI group: 152; control group: 168)	WLI/NBI/Lugol endoscopic media with histopathology and clinical data	Prospective, single-center, exploratory randomized controlled trial
Aoyama et al[20]	2025	AI model for superficial ESCC detection from endoscopic videos	Subgroup analysis and comparison with endoscopist performance	Training data: 280 cases (140 with lesions, 140 without); Test data: 115 cases (57 with lesions, 58 without)	NBI endoscopic videos with histopathological diagnosis	Prospective data, retrospective analysis design
Ma et al[76]	2025	Diagnostic performance for ESN classification	iCLE vs experts and non-experts: Diagnostic performance and support evaluation	2803 patients (for iCLE training/validation); 226 patients (image test); additional patients for video recognition testing (prospective, total N not shown)	pCLE video/image data with histopathological gold standard	Prospective diagnostic study

AI: Artificial intelligence; BLI: Blue light imaging; CAD: Computer-aided diagnosis; CNN: Convolutional neural network; DCNN: Deep convolutional neural network; DEN: Digestive endoscopy; DL: Deep learning; EC: Esophageal cancer; eCAM: Explicit class activation maps; ER: Endoscopic resection; ESCC: Esophageal squamous cell carcinoma; ESCN: Early squamous cell neoplasia; ESN: Esophageal squamous neoplasm; HD: High-definition; HSI: Hyperspectral imaging; iCLE: Integrated confocal laser endomicroscopy; IDPS: Invasion depth prediction system; IPCL: Intra-papillary capillary loop; ME: Magnifying endoscopy; NBI: Narrow-band imaging; pCLE: Probe-based confocal laser endomicroscopy; RGB: Red, green, blue; SCC: Squamous cell carcinoma; SESCC: Superficial esophageal squamous cell carcinoma; WLI: White light endoscopy; YOLO: You only look once.

Table 3 Summary of reviews and meta-analysis on esophageal squamous cell carcinoma using artificial intelligence.

Ref.	Year	Primary outcomes	Secondary outcomes	Patients studied	Type of data	Study design
Zhang et al[63]	2020	Not applicable	Not applicable	Not applicable	Not applicable	Mini-review
Syed et al[82]	2020	Summary of literature using DL to identify esophageal tumors	DL/CNN overview, current limitations, and future directions	Not applicable (not focused on individual study data; 21 relevant studies reviewed)	Endoscopic images with histology-confirmed clinical data	Mentored review
Huang et al[83]	2020	Not applicable	Not applicable	Not applicable	Not applicable	Mini-review
Lazăr et al[84]	2020	ML model comparison for endoscopic esophageal lesion assessment	AI in care quality, workflow efficiency, and physician support	Not applicable (review of multiple studies)	Review of AI in endoscopic assessment of esophageal disease	Review
Namikawa et al[85]	2020	Review of studies using CNNs in gastrointestinal endoscopy	AI in GI polyp/cancer detection and capsule endoscopy	Not applicable	Not applicable	Review article
Zhang et al[1]	2021	Diagnostic accuracy of AI-assisted models	AI vs endoscopists: Pooled accuracy and subgroup performance	Not applicable	Endoscopic media and histology-confirmed clinical data	Meta-analysis
Liu et al[44]	2021	Not applicable	Not applicable	Not applicable	Not applicable	Mini-review
Ma et al[13]	2022	CNN-AI for early EC detection from endoscopy	Heterogeneity and bias analysis (I2, meta-regression, Deeks’ plot)	Meta-analysis of 7 studies; number of patients and images varied across studies	Meta-analysis of WLE/NBI-based studies with histological confirmation	Meta-analysis
Nagao et al[64]	2022	AI in upper GI: Current status, clinical integration, and future outlook	AI in H. pylori infection diagnosis, gastric anatomy, and upper GI cancer detection	Not applicable (review of multiple studies)	CNN-based AI for diagnosis of GI and pharyngeal cancers	Review
Tokat et al[86]	2022	Overview of the applicability of AI technology in upper gastrointestinal endoscopy	Clinical AI challenges and applications in GI oncology	Not applicable	WLE/NBI/ME-NBI images with clinical and histological data	Review article
Guidozzi et al[14]	2023	AI in endoscopic diagnosis of esophageal cancer	AI vs endoscopists, tumor types, and bias risk (QUADAS-2)	1590 patients for ESCC diagnosis (14 studies); 478 patients for EAC diagnosis (9 studies)	English studies using endoscopic images/videos	Systematic review and meta-analysis
Pan et al[87]	2023	AI in early ESCC endoscopic diagnosis: Current status	Discussion of limitations and future prospects of AI in ESCC diagnosis	Not applicable (review of multiple studies)	Review of DL models for ESCC diagnosis	Review
Tao et al[32]	2024	AI vs experts: Accuracy in early EC and depth diagnosis	AI vs endoscopists: Performance, invasion depth diagnosis, and bias assessment	Not applicable	Meta-analysis of 19 studies with heterogeneous sample sizes	Meta-analysis
Kikuchi et al[88]	2024	Recent studies on endoscopic AI for GI tumors	Discussion on the future outlook of AI in gastroenterology	Not applicable	Published studies on DL-based endoscopic AI	Review
Yan et al[33]	2025	Not applicable	Not applicable	Not applicable	Not applicable	Review article
Shahzil et al[89]	2025	Lesion color contrast, visibility, and GI detection rate	Lesion visibility and analysis of serrated lesions and adenomas	16634 individuals (data integrated from 17 studies)	Clinical and endoscopic image data	Systematic review and meta-analysis

AI: Artificial intelligence; CNN: Convolutional neural network; DL: Deep learning; EAC: Esophageal adenocarcinoma; EC: Esophageal cancer; ESCC: Esophageal squamous cell carcinoma; GI: Gastrointestinal; H. pylori: Helicobacter pylori; ME: Magnifying endoscopy; ML: Machine learning; NBI: Narrow-band imaging; WLE: White light endoscopy.

Diagnostic performance using AI-assisted endoscopy for ESCC

Recent meta-analyses collectively demonstrated high diagnostic accuracy of AI-assisted endoscopy for early ESCC detection (Table 3). A meta-analysis of convolutional neural network systems reported a pooled sensitivity of 0.90 [95% confidence interval (CI): 0.82-0.94], specificity of 0.91 (95%CI: 0.79-0.96), and area under the curve of 0.95 (95%CI: 0.93-0.97) for early ESCC detection[13]. Another meta-analysis reported a pooled sensitivity of 91.2% (84.3%-95.2%) and specificity of 80.0% (64.3%-89.9%)[14]. These robust findings are supported by numerous individual clinical studies (Table 1).

An AI system combined white light imaging (WLI) and NBI images in an internal validation set with a sensitivity of 96.64%, specificity of 95.35%, and accuracy of 91.75%; for early ESCC. The accuracy, sensitivity, and specificity of early ESCC detection were 85.7%, 92.6%, and 80.0%, respectively[15,16]. For non-magnified NBI/blue-laser imaging, the same system yielded a sensitivity of 100%, specificity of 63%, and accuracy of 77%; for WLI the corresponding values were 90%, 76%, and 81%, respectively[17]. In a video-based evaluation, expert endoscopists achieved a sensitivity of 79%, specificity of 72%, and accuracy of 75%, whereas the AI model attained a sensitivity of 91%, specificity of 51%, and accuracy of 63% (a higher sensitivity but lower specificity and overall accuracy)[18]. These data suggest that AI can extract pixel-level patterns imperceptible to the human eye, maximizing diagnostic yield from various imaging modalities. Given that the interobserver agreement among endoscopists for early ESCC is approximately 70%, AI can serve as a valuable tool to enhance diagnostic consistency[19].

Improvement in endoscopic skill using AI

Beyond its standalone capabilities AI shows great potential for enhancing the skills of endoscopists. With AI assistance the sensitivity and accuracy of general endoscopists increased from 57.4% to 66.5% and from 68.6% to 75.9%, respectively. Among experts the sensitivity increased from 59.1% to 70.0%, and the accuracy increased from 72.1% to 79.3%[20]. AI has markedly enhanced the ability of less experienced endoscopists to classify intrapapillary capillary loop patterns and predict invasion depth[21] (Table 2). AI systems detect and precisely delineate very small flat lesions (approximately 3 mm) that are often missed visually regardless of WLI/NBI use, magnification, or iodine staining[22,23]. The mean processing time to detect and outline a lesion on a single image was 17 milliseconds for AI and 92 seconds for an expert endoscopist. Video diagnosis can be performed at 26-37 milliseconds per frame, and some systems can scan an entire cancer video within 1 second, making it sufficiently fast for real-time use. Several solutions have been integrated into endoscopic processors that do not require additional monitoring[16,18,22,24].

Challenges for the clinical application of AI

Despite these promising results several challenges must be addressed for successful clinical integration. The black-box nature of deep learning models is a significant barrier to their clinical adoption in endoscopy as it limits transparency and clinician trust[25]. Explainable AI (XAI) techniques are being actively developed to address this challenge by providing interpretable justifications for AI outputs[26]. In endoscopic diagnosis methods such as gradient-weighted class activation mapping generate heat maps that visually indicate the image regions that influence the AI decision, allowing clinicians to critically assess the focus and reasoning of the model[27]. Studies have demonstrated that XAI not only improves diagnostic support but also enhances the trust and acceptance of endoscopists. The ENDOANGEL-ED system, which incorporates explainability, was shown to significantly increase both diagnostic accuracy and clinician confidence compared with traditional deep learning models while also improving endoscopist performance in multireader studies[27]. Beyond diagnostic support XAI has educational value in visualizing expert lesion recognition patterns, which can be leveraged for training and skill development[28,29].

Thus, the integration of XAI is foundational for building clinician trust, meeting regulatory and ethical requirements, and facilitating the safe and effective incorporation of AI into routine endoscopic practice[28,30]. Some studies have failed to demonstrate the non-inferiority of AI systems compared with human endoscopists[31]. Furthermore, the creation of AI models requires large, high-quality datasets; insufficient volume, poor quality, and sampling bias remain major obstacles[1]. False positives during AI diagnosis are an additional concern[32]. Future priorities include large multicenter validation studies, optimization of algorithms for real-time video processing, and proactive resolution of ethical and legal issues from a patient-centered perspective[33].

BIG DATA-DRIVEN AI STRATEGIES PIONEERED BY NATIONAL CLINICAL DATABASE AND THE JAPAN ENDOSCOPY DATABASE

Introduction of national registries in Japan

The National Clinical Database (NCD) and Japan Endoscopy Database (JED) are pivotal national registries in Japan that provide a foundation for quality improvement, research, and policies in surgical and endoscopic medicine[34,35]. Established in 2010 and closely linked to the surgical board certification system, the NCD is a nationwide web-based registry that captures > 95% of surgical procedures in Japan[34,36]. It enables comprehensive tracking of surgical outcomes, risk adjustment, and institutional benchmarking[37,38]. By providing feedback to participating hospitals and facilitating national and international comparisons, the NCD drives continuous quality improvement and evidence-based policies in Japanese surgical care[34,36,39]. Its vast dataset has also illuminated key clinical trends, such as an aging patient population and a shift towards minimally invasive surgery[37,40,41].

The JED, initiated by the Japan Gastroenterological Endoscopy Society, is a prospective project aimed at becoming one of the world’s largest endoscopy databases[35]. It standardizes terminology and data collection for all endoscopic procedures, allowing for robust real-world assessment of diagnostic and therapeutic techniques, complication rates, and procedural trends, thereby supporting competency evaluation, adverse event monitoring, and high-impact research[42,43].

Role of national databases in AI research

Building robust and generalizable AI models requires large-scale heterogeneous datasets to mitigate data bias, which is a challenge often faced by single-institution studies (Figure 1)[44,45]. Japan’s national registries, the NCD and JED, provide a decisive solution because of their scale, quality, and standardized structure. JED provides an unparalleled platform for the development of AI in endoscopic diagnosis[35,42,43]. It collects endoscopic images and videos under standardized protocols and a wealth of associated metadata, including histopathological diagnoses, specific endoscope models used, and experience level of the endoscopists[35]. This structured dataset is invaluable in all phases of AI development. High-quality images linked to definitive pathological results ensure the quality of both model training and internal validation. Comprehensive surgical and pathological data from the NCD are ideal resources for creating and validating AI-powered risk models and outcome-prediction tools for perioperative management[36,39,46].

Figure 1 Flowchart of artificial intelligence development process using databases. AI: Artificial intelligence.

The goal of these registries is to enable robust external validation. By reflecting real-world clinical heterogeneity through multicenter, multivendor, and multioperator data, these studies allow for a true assessment of the robustness and generalizability of multimodal AI[47]. This provides Japanese researchers with a significant competitive advantage in developing and deploying clinically reliable and scalable multimodal AI applications globally[48,49].

Future potential of national registries in Japan

Despite their immense potential leveraging large-scale databases such as the NCD and JED for routine AI integration presents several key challenges. A primary hurdle lies in the data infrastructure: Incomplete data interoperability across different hospital systems complicates harmonization, whereas issues of data quality and completeness, such as missing entries, can undermine the reliability of AI models[35,42]. Furthermore, ensuring real-world generalizability is critical as models trained on national data must perform robustly across heterogeneous clinical practices and patient populations without overfitting[48-51]. The successful clinical adoption of these tools also depends on their seamless integration into existing workflows, a process often hindered by concerns over the interpretability of black-box models and clinician trust[52]. Finally, navigating the complex regulatory, ethical, and governance landscapes, including data privacy and liability, presents a significant barrier that requires ongoing collaboration among clinicians, regulators, and industry[48,50,53]. Systematically addressing these multifaceted challenges is essential to unlock the full potential of national databases for developing clinically impactful AI technologies. The goal is to create multimodal AI models that are not only accurate but also fair, robust, and seamlessly integrable into diverse clinical environments.

ETHICAL AND LEGAL CONSIDERATIONS

Ethical and legal concerns hinder the development of AI. The black-box nature of AI and the uncertainty regarding the legal responsibility for diagnostic errors are critical barriers to social acceptance and full incorporation into screening programs. The clinical integration of AI raises complex legal questions regarding liability in the event of diagnostic errors[54]. In the Japanese legal context, current regulations do not explicitly define the legal status or liability of AI-assisted diagnostic tools. Although the prevailing view holds that physicians retain their ultimate responsibility under the Medical Practitioners Act, the practical application of this principle in AI-supported settings remains ambiguous, particularly in cases of misdiagnosis or omission of relevant information. Although the principles that AI serves as a supportive tool and that the ultimate diagnostic responsibility rests with the clinician are widely accepted, its application in specific clinical scenarios is not straightforward[55].

Consider a false-positive case in which the AI flags a lesion, but the clinician, based on expert judgment, correctly identifies it as benign, thus avoiding an unnecessary biopsy. The current consensus in the medical literature emphasizes that robust, externally validated clinical evaluations, including prospective studies and randomized trials, when feasible, are essential before AI tools can be integrated into practice[56]. While the clinician’s decision is medically sound, a record of overruling an AI alert could potentially increase litigation risk if a different lesion is missed. Conversely, in a false-negative scenario both the AI and clinician fail to detect a lesion[57]. In such cases if an AI system is proven to have a performance level equivalent to or exceeding that of an average specialist, a new legal landscape may emerge in which a clinician’s failure to consult the AI or their overreliance[57] on it could be considered a breach of the standard of care[57]. The rapid use of AI models in clinical settings has outpaced the establishment of standardized evaluation criteria, leading to variability in performance assessments and challenges in ensuring safety, efficacy, and equity across diverse patient populations[58].

Ethical and legal challenges extend beyond liability. For instance, the issue of incidental findings in which AI flags a suspicious lesion in an adjacent area outside the primary focus of examination, such as the pharynx, during esophageal screening, presents a new ethical dilemma[59]. Clear guidelines regarding the extent of a clinician’s obligation to investigate incidental and out-of-specialty findings have not yet been established. Although no high-profile legal cases have emerged in Japan regarding incidental AI findings, international reports, particularly in radiology, have raised concerns that the failure to investigate AI-flagged abnormalities, even in adjacent or non-target areas, could lead to legal liability. For instance, in the United Kingdom and United States, legal commentaries have highlighted that overlooking AI-generated alerts may be considered a breach of the standard of care, especially when such alerts pertain to clinically significant but incidental findings[57-59].

Furthermore, ethical frameworks emphasize the need for transparent data stewardship, clear governance, and mechanisms for patient empowerment and consent management[60]. Legal frameworks in Japan, such as the Next Generation Medical Infrastructure Act, provide a basis for the secondary use of medical data but face challenges in practical implementation, including underutilization and regulatory complexity[61]. Establishing protocols for the quality management of updated AI models and ensuring transparent communication of any changes to the clinical community are equally important. Building societal consensus and regulatory frameworks to address these issues is the key to sustainable development and safe implementation of AI technology[1,44,45].

PROSPECTS OF AI APPROACHES FOR EARLY ESCC

Recent advancements in image-enhanced endoscopy have led to a multimodal AI approach that integrates complementary information from diverse sources (Figure 2). By simultaneously inputting images of the same lesion captured in multiple modes, such as WLI, NBI, and linked color imaging, and training AI to analyze these features cohesively, such systems may achieve a level of robustness and diagnostic accuracy that surpass that of any single modality AI. This multimodal strategy has the potential to become the standard treatment for AI-assisted endoscopic diagnosis. However, most current multimodal AI systems remain in the experimental or proof-of-concept phase and have not yet been implemented in routine clinical workflows.

Figure 2 Multimodal artificial intelligence information integration scheme. AI: Artificial intelligence; WLI: White light imaging; NBI: Narrow band imaging; LCI: Linked color imaging.

Beyond lesion detection and border delineation, multimodal AI, which integrates pathological, clinical, and genomic data, is attracting attention as a next-generation platform for personalized medicine, enabling depth-of-invasion prediction, treatment-response assessment, and therapeutic decision-making. Ideally, AI should serve as an adjunct, elevating the diagnostic ability of less-experienced physicians to the expert level, thereby standardizing care. To this end AI should not serve as a replacement for clinicians but as a tool for intelligence augmentation, enhancing their inherent skills.

Recent studies, including those involving generative AI, have consistently shown that human-AI collaboration yields superior outcomes in tasks ranging from diagnosis to text generation, surpassing the performance of either humans or AI working alone[55,62]. This collaborative paradigm is poised to contribute significantly to endoscopic diagnosis by improving accuracy, standardizing care, and reducing physicians’ cognitive loads[20,23,63,64]. AI implementation must be accompanied by continuous updates to the legal framework, insurance policies, and physician-training programs. Only through a holistic ecosystem spanning technical robustness, legal clarity, and societal consensus can AI truly augment clinical care in a safe and sustainable manner.

CONCLUSION

Advances in AI have potential benefits not only for the early detection of esophageal cancer but also for highly specialized tasks such as selecting treatment strategies and evaluating therapeutic efficacy. Multimodal AI and AI-based educational systems are expected to become future standards of care, further advancing personalized healthcare. As AI technology continues to evolve, exclusive dependency on it should be avoided. Continuous refinement of clinical expertise, knowledge, and ethical standards is indispensable. Given the inherent uncertainty of medicine, we must remain vigilant and adaptable.

However, limitations such as dataset bias, regulatory uncertainty, and the lack of long-term outcome evidence remain significant barriers to clinical implementation. Addressing these challenges will require sustained interdisciplinary collaboration among clinicians, data scientists, engineers, and ethicists. Such collaborations are essential to ensure the safe, effective, and ethically sound integration of AI into real-world clinical practice.

Ultimately, the goal of integrating AI is not to replace clinicians but to enhance their capabilities, standardize the quality of care, and disseminate expertise. This review provided a roadmap for clinicians and researchers to address opportunities and challenges, ensuring that AI is leveraged to achieve meaningful improvements in the care of patients with ESCC.

ACKNOWLEDGEMENTS

We thank the reviewers for their comments, which have helped us improve the manuscript. Open Evidence (https://openevidence.com), Gemini 2.5pro (https://gemini.google.com/), and GPT-4o and o3-pro (https://chat.openai.com/) were used for the search and draft editing, respectively.