Artificial Intelligence for Esophageal Precancerous Lesions and Esophageal Cancer

Abstract

Esophageal cancer is associated with a relatively poor prognosis. Early detection and diagnosis are important for the survival and prognosis of patients with esophageal cancer. Recently, the development of artificial intelligence (AI) and its application in clinical medicine has led to remarkable progress in various endoscopic fields, including the detection of Barrett’s esophagus and esophageal cancer. Compared to human errors induced by fatigue and impairment in diagnostic precision, progression in the field of AI, including deep learning and convolutional neural networks, has resulted in improved diagnostic accuracy. Subtle microvascular changes in lesions, visual disturbances, and fatigue, as well as varying endoscopic expertise depending on the operator, are the major causes of missing rates in detecting cancerous lesions. Specifically, the rate of missed diagnoses of esophageal squamous cell carcinoma reaches 4%–17%. In contrast, deep learning-based AI assists in the diagnosis of esophageal squamous cell carcinoma, evaluation of invasion depth, microvascular classification, and delineation of the margin of the lesion. In this review article, we investigated the development of the current AI model introduced for esophageal precancerous lesions, including Barrett’s and esophageal cancer.

INTRODUCTION

Esophageal cancer (EC) ranks seventh highest amongst malignancy occurrences worldwide, with 604000 new cases of incidence each year [1]. It has the sixth highest mortality rate, accounting for 544000 deaths per year [1]. The two most common histologic subtypes of EC, squamous cell carcinoma and adenocarcinoma, exhibit different geographic manifestations. Apart from the rapidly increasing global incidence rate of adenocarcinoma, squamous cell carcinoma has a high incidence rate in Eastern Asia [2,3]. With poor prognosis and high mortality rates, the early detection of EC is important for favorable disease outcomes.

As a precancerous condition, Barrett’s esophagus (BE) is characterized by the replacement of the natural squamous epithelium with intestinal differentiation of specialized epithelium. The development of BE is closely related to gastroesophageal reflux disease (GERD); however, its progression to cancerous conditions is independent of the presence of GERD.

Missed diagnosis with conventional white-light endoscopy (WLE) for early EC that is demonstrated as interval cancer is commonly encountered in clinical field [4,5]. Chromoendoscopy using Lugol’s solution is a standard screening tool for detecting early EC, but it has the limitations of causing chest discomfort for patients and prolongation of endoscopic procedure time for the endoscopist [6,7]. Progressive endoscopic imaging modalities including narrow band imaging (NBI) and magnification endoscopy have been applied in combination with WLE to enhance the endoscopic detection rate of early malignant lesion features in intrapapillary capillary loops (IPCLs) [8].

However, endoscopic detection of early EC is not easy as it depends on the endoscopists’ subjective decision and endoscopic technical expertise in detecting the delicate and fine microvascular change of esophageal mucosa [9].

Specifically, the rate of missed diagnosis for esophageal squamous cell carcinoma (ESCC) is 4%–17% [9].

Recently, artificial intelligence (AI) has been introduced into various medical fields to support and assist in diagnosis. This presents a promising possibility for meeting the requirements for endoscopic detection and diagnosis, including early EC [10,11].

Machine learning and deep learning are key factors of AI [12]. Among these, convolutional neural networks (CNNs), including ResNets, are most commonly used for image classification tasks, whereas object-detection frameworks such as You Only Look Once (YOLO) are optimized for real-time lesion localization. Encoder–decoder architectures with U-Net are frequently applied on segmentation and margin delineation, and more recent models such as Vision Transformers (ViTs) are currently being introduced. CNN-based models extract hierarchical features from images through multilayered convolutional operations, whereas deep learning detects convoluted structures from datasets using a backpropagation algorithm [13]. These architectures enable AI systems to function as supportive assistants by analyzing large-scale endoscopic and histopathological imaging datasets.

In this review article, we investigated the development of the current AI model introduced for esophageal precancerous lesions, including Barrett’s and EC.

APPLICATION OF AI IN DETECTING BE NEOPLASIA

The detection and identification of subtle structural changes in esophageal adenocarcinoma (EAC) associated with BE, which can be easily missed by human vision, are assisted by the application of AI. Improvement in the neoplasia detection rate (NDR) has contributed to the early diagnosis of EAC and advanced prognosis. A reduction of 3.5% in Barrett’s neoplasia was achieved by a 1% increase in the NDR [14]. There has been a shift from focusing on endoscopic still images to real-time systems for evaluating the model performance of AI, resulting in an abrupt progression in the capacity of AI computing systems. Most BE detection systems have deep CNN backbones for feature extraction in combination with detection heads or segmentation networks for lesion localization. Current algorithms are generated independently through manipulation by the operator, allowing the integrated output of AI to be presented as a region of interest surrounded by bounding boxes or heatmaps that act as active assistants during the endoscopic procedure. The goal of the final development and implementation of AI in the endoscopic clinical field is to remove and overcome the gap between preliminary validation results and real-world outcomes. Table 1 summarizes the current studies with AI application in BE and early Barrett’s neoplasia, including the imaging modality, AI architecture, study design, and diagnostic performance.

Table 1.

Summary of studies on AI application for Barrett’s esophagus and early Barrett’s neoplasia

Study	Design	Modality	Target lesion	AI task	AI architecture	Key performance	Validation
de Groof et al. (2020) [15]	Prospective	WLE (images)	Early BE neoplasia	CADe	hybrid ResNet/U-Net	Per-level accuracy/sensitivity/specificity 90%/91%/89%	External
	Pilot study
Fockens et al. (2023) [16]	Prospective	WLE, NBI (images and videos)	Early BE neoplasia	CADe	EfficientNet-Lite1 encoder, MobileNetV2 DeepLabV3+ decoder	All-comers (image) sensitivity/specificity: 90%/70%	External
	Multicenter					All-comers (video) sensitivity/specificity: 97%/85%
						Benchmarking (image) sensitivity/specificity: 90%/80%
						Benchmarking (video) sensitivity/specificity: 91%/82%
Hussein et al. (2022) [17]	Prospective	WLE with i-scan	Early BE neoplasia	CADe	ResNet101, FCNResNet50	Per-image AUC 0.93/sensitivity 91%	Internal
Takeda et al. (2024) [18]	Retrospective	WLE, LCI	Short-segment BE	CADe	Vision Transformer (ViT-B/16)	Accuracy/sensitivity/specificity	Internal
						WLI: 84.1%/89.6%/75.5%
						LCI: 90.5%/90.1%/91.1%

AI, artificial intelligence; BE, Barrett’s esophagus; WLE, white-light endoscopy; CADe, computer-aided detection; NBI, narrow band imaging; LCI, linked color imaging; AUC, area under the receiver operator curve.

de Groof et al. [15] evaluated preliminary diagnostic accuracy of deep learning–based computer-aided detection (CADe) systems trained on high-resolution WLE images detecting early Barrett’s neoplasia. It had per-level accuracy of 90% and demonstrated significance of evaluating a CADe system in live endoscopic procedures [15].

Fockens et al. [16] studied the development and validation of CADe system training using a dataset of 14046 images from 15 centers. It presented results of sensitivity and specificity for all-comers image test set with 90% and 70% and all-comers video test set with 90% and 70%.

Hussein et al. [17] developed CNN-based systems using highdefinition-WLE and 119 i-scan-based videos, including dysplastic and non-dysplastic BE (NDBE). It presented a per-image sensitivity in the test set of 91%, specificity of 79%, and area under the receiver operator curve of 93% to detect dysplasia [17].

More recently, Takeda et al. [18] applied a Vision Transformer (Vit-B/16-384) to diagnose short-segment BE (SSBE) with WLE and linked color imaging (LCI). It has a similar specialist accuracy, with results of 84.1% for white-light imaging (WLI) and 90.5 for LCI.

BE HISTOPATHOLOGIC DIAGNOSIS WITH DEEP LEARNING APPLICATION

Similar to the application of the CNN model for the classification of pathological images, machine learning, including supervised learning, is essential for pathologists in the diagnostic process. CNN-based architectures such as ResNet, Inception, and custom deep feature extractors are commonly used and often combined with weakly supervised or unsupervised learning strategies to reduce the annotation burden. In this sequence, a large amount of endoscopic data is processed to develop diagnostic models that are independent of labeling. Table 2 presents a summary of outcomes from current studies of BE histopathologic diagnosis using deep learning applications.

Table 2.

Pathology-based AI studies for Barrett’s dysplasia and esophageal adenocarcinoma

Study	Design	Modality	Target lesion	AI task	AI architecture	Key performance	Validation
Tomita et al. (2019) [19]	Retrospective	WSI	BE-no-dysplasia, BE-with-dysplasia, ADC	CADe	ResNet18	Accuracy BE-no-/BE-with-dysplasia/ADC 85%/89%/88%	Internal
Faghani et al. (2022) [20]	Retrospective	WSI	BE-no-dysplasia, BE-with-LGD, BE-with-HGD	CADe	ResNet101, YOLOv5	Sensitivity, specificity	Internal
						BE-no-dysplasia/with-LGD/HGD 94.4%, 96.2%/81.8%, 97.3%/94.7%, 90.2%

AI, artificial intelligence; WSI, whole-slide images; ADC, adenocarcinoma; LGD, low-grade dysplasia; HGD, high-grade dysplasia; CADe, computer-aided detection.

For the histological diagnostic confirmation of BE and associated neoplasia, deep learning is highly applicable, as supported by current studies.

Tomita et al. [19] evaluated a novel deep learning model that utilizes tissue-level annotations with high resolution histological image analysis for detection of BE and EAC. The proposed model presented classification accuracy with 0.85 (95% confidence interval [CI], 0.81–0.90) for the BE-no-dysplasia class, 0.89 (95% CI, 0.84–0.92) for the BE-with-dysplasia class, and 0.88 (95% CI, 0.84–0.92) for the adenocarcinoma class [19]. It also presented a mean accuracy of 0.83 (95% CI, 0.80–0.86) and marginal outperformance compared with the sliding window approach on the same testing set of esophagus tissue with 4 classes [19].

Faghani et al. [20] developed a deep learning model to predict dysplasia grades using whole-slide imaging. There were 542 patients’ slides with NDBE, low-grade dysplasia (LGD), and high-grade dysplasia (HGD) included in the study [20]. LGD has 81.3% sensitivity and 100% specificity, and NDBE as well as HGD reached >90% [20]. The overall positive predictive value and sensitivity metric was 0.91 for NDBE, 0.90 for LGD, and 1.0 for HGD [20]. It demonstrated potential for improving diagnosis of BE dysplasia as well as endoscopic treatment application [20].

IMAGE-BASED AI IN ESOPHAGEAL SQUAMOUS CELL CANCER

WLI based AI models

Conventional WLI endoscopy is commonly applied for detecting ESCC but it has limitation of detecting and delineate margins of early ESCC.

Despite these limitations, Liu et al. [21] developed a model to detect and delineate margins of early ESCC using WLI endoscopy. A total of 13083 WLI images from 1239 patients were used to train and test the AI model, which demonstrated an accuracy of 85.7% for detecting lesions and 93.4% for delineating margins [21].

In another study, Baik et al. [22] evaluated an AI algorithm detection model developed based on YOLOv5 and the RetinaNet architecture with WBI and non-magnified NBI. It demonstrated an outcome with a precision of 98.4% and sensitivity of 91.3% in the NBI dataset of RetinaNet, and a precision of 93.7% and sensitivity of 89.9% in the WLI dataset of the YOLOv5 architecture [22].

In a recent study, Shiroma et al. [23] developed video-based WLI AI systems to improve simultaneous real-time detection using a single-shot multibox detector deep neural network architecture. They used 144 esophagogastroduodenoscopy (EGD) videos for the two validation sets, followed by the evaluation of 40 patients’ 80 EGD videos including 20 patients with superficial ESCC and 20 with non-ESCC. With real-time AI assistance, the sensitivities of the endoscopists were significantly improved compared with those without AI support (p<0.05).

Although there are limitations in assessing the depth of invasion and histological differentiation, video-based WLI AI systems have expertise in screening and reducing the missing rate.

NBI and blue-light imaging non-magnifying image-enhanced endoscopy-based AI models

Since subtle vascular characterization is utilized in the early detection of ESCC, advancements in NBI and blue-light imaging (BLI) in electronic chromoendoscopic technology have allowed the accurate identification of early neoplastic changes. Therefore, NBI- and BLI-based trained AI models generally demonstrate a higher specificity than WLI-based systems, particularly for ESCC detection.

Yuan et al. [24] developed a superficial ESCC detecting AI system with non-magnified NBI images from 4 hospitals followed by training and validating with 10047 still images and 140 videos from 1112 patients and 1183 lesions. Internal and external tests presented accuracy of detecting lesions as 92.4% and 89.9%, respectively, with delineating extents accuracy 88.9% and 87.0%, respectively [24].

NBI and BLI magnifying endoscopy based AI models

Magnified NBI or BLI with high-resolution AI models are suitable for analyzing detailed microvascularity and depth of invasion according to the Japanese Esophageal Society (JES) classification, particularly IPCL morphology. Current IPCL-based AI studies have demonstrated the ability to differentiate intramucosal (T1a) from submucosally invasive (T1b) ESCC with high accuracy. Yuan et al. [25] reported development of an AI-assisted diagnosis system to make assessment on microvascular morphology of lesions with retrospectively collected magnifying endoscopic (ME) NBI from three hospitals. A total of 7094 ME-NBI images from 685 patients were applied for training and validation and the AI system that was developed resulted in improving the diagnostic ability of endoscopists for the IPCL classification as well as superficial ESCC [25].

Table 3 summarizes the endoscopic AI studies on ESCC, including the modality, AI architecture, and validation level.

Table 3.

Endoscopic AI studies of ESCC

Study	Design	Modality	Target lesion	AI task	AI architecture	Key performance	Validation
Lin et al. (2022) [21]	Retrospective	WLI (images)	ESCC	CADe	DCNN	Detecting lesions accuracy of 85.7%	External
						Delineating margins accuracy 93.4%
Baik et al. (2025) [22]	Retrospective	WLI, non-magnified NBI (images)	ESCC	CADe	YOLOv5, RetinaNet	NBI dataset precision/sensitivity 98.4%/91.3%	External
						WLI dataset precision/sensitivity 93.7%/89.9%
Shiroma et al. (2021) [23]	Retrospective	WLI (video)	ESCC	CADe	Single Shot MultiBox Detector	Sensitivity 85%	Internal
Yuan et al. (2023) [24]	Prospective	Non-magnified NBI (video)	ESCC	CADe	YOLACT	Detecting lesions accuracy of 92.4%	External
						Delineating margins accuracy 88.9%
Yuan et al. (2022) [25]	Retrospective	ME-NBI (images)	ESCC	CADx	HRNet+OCR (DCNN)	Accuracy 91.3%	External

AI, artificial intelligence; ESCC, esophageal squamous cell carcinoma; WLI, white-light imaging; NBI, narrow band imaging; CADe, computer-aided detection; ME, magnifying endoscopic.

Beyond invasive carcinomas, esophageal squamous dysplasia represents a clinically significant premalignant condition and a critical target for early detection strategies. Low- and high-grade squamous dysplasia are associated with a substantially increased risk of progression to ESCC, and their identification directly influences surveillance intervals, endoscopic therapy decisions, and patient prognosis. However, the endoscopic diagnosis of squamous dysplasia is particularly challenging, as dysplastic lesions often present with subtle color changes or minimal surface irregularities that may be indistinguishable from inflammatory or reactive mucosal changes.

However, in the current literature, dysplastic lesions are often grouped together with early ESCC or included as part of a broader neoplastic category, making it difficult to draw firm conclusions regarding the performance of AI for dysplasia alone. Moreover, most studies were retrospective and image-based, with limited prospective validation and data on the real-time detection of squamous dysplasia during real-world endoscopy.

LIMITATIONS OF CURRENT AI APPLICATIONS

Although the promising and beneficial expectations of the current studies outcomes provide positive feedback, there are several crucial limitations that should be considered in future studies.

First, there is a lack of large-scale multicenter randomized trials in the field of published AI EC. Most current studies are retrospective or single-center investigations, and image datasets are more common. Future follow-up study scales and methodologies must be considered to improve the reliability of the data applications and clinical outcomes.

Second, the general application of AI systems in the current environments of specific, selected, and distinct study fields is limited. Selected endoscope models are applied, and highly expert optimized centers are involved in training, raising questions on providing standard outcomes for the general population in the real world.

Third, controlling the variability of the label definition and ground-truth assignment is another issue. Since the estimated diagnostic accuracy is affected by the annotation strategy (still images versus videos), reference standards (biopsy versus resection pathology), and the use of consensus labeling, it is important to control for heterogeneity across studies.

Finally, the potential unintended consequences of deskilling behaviors of endoscopists affected by AI CAD systems is concerning. Although results from upper gastrointestinal endoscopy are limited, evidence from AI applications in colonoscopy raises concerns about prolonged reliance on CAD systems, which may alter endoscopist behavior resulting in reduced diagnostic ability when using non-AI detection methods.

AI APPLICATION ON ENDOSCOPIC QUALITY CONTROL AND SURVEILLANCE

Considering the current limitations of AI applications in the real world, the introduction of quality-control-based AI systems may contribute to standardizing examination quality and reducing operator-dependent variability. Quality-controlled AI systems focus on the adequacy of liminal distension, complete mucosal visualization with minimal blind spots, ensuring systematic inspection, and appropriate inspection time. Therefore, quality-controlled AI systems offer immediate feedback to reduce missed esophageal neoplasia in routine clinical practice. Notably, current commercially available systems provide detection assistance (CADe) and procedure monitoring, rather than diagnosis (CADx).

The ENDOANGEL system is a commercially available AI system approved by the National Medical Products Administration (NMPA) of China that provides quality control monitoring and detection assistance with blind-spot detection, inspection time monitoring, and real-time alerts. In a multicenter study, the ENDOANGEL system demonstrated reductions in blind spots (mean 5.38 [standard deviation (SD) 4.32] vs. 9.82 [SD 4.98]; p<0.001) and longer inspection times (5.40 [SD 3.82] min vs. 4.38 [SD 3.91] min; p<0.001) demonstrating the potential improvement in quality control [26].

The ENDO-AID system is another commercially available CAD system that integrates the Olympus EVIS X1 endoscopic system. It has enhanced visualization and detection assistance with CADe, CADx and AI-assisted imaging functions [27]. A recent retrospective study including 2000 patients presented significant improvement of mean polyp per patient score in the AI group (1.26) compared with the control group (0.85) (p<0.001) [27].

The WISENSE system is a real-time quality-controlled platform demonstrating a reduced blind spot rate in EGD procedures in comparison to the control group (5.86% vs. 22.46%, p<0.001) [28].

Therefore, future research focusing on the clinical endpoints of missing rates, interval cancer rates, and evaluation of unnecessary biopsies or procedures would advance the field of study of comparing AI-augmented quality-controlled, lesion-detection CADe, and a combination of quality-control with CADe to overcome the limitations of real-world AI applications.

CONCLUSION

AI has rapidly advanced as a supportive tool for the detection and characterization of esophageal precancerous lesions and EC with the development of various AI architectures, including ResNet-based CNNs, YOLO-based real-time detectors, U-Net segmentation models, and ViTs. Current evidence demonstrates that AI models improve the detection of Barrett’s neoplasia and early ESCC, assist in margin delineation, and support invasion-depth assessment using microvascular pattern analysis. However, there are limitations that require the development of quality-control-focused AI systems that standardize endoscopic examinations and reduce blind spots for the practical clinical endpoints of achieving advancement in missing rates, interval cancer incidence, and endoscopist performance.