Stomach mirrors A1 revolution: 150,000 video data sets remodel the digestive early-sifting pattern

We are faced with a group of thought-provoking data: China has about 480,000 new stomach cancers each year, accounting for more than 40 per cent of new cases of stomach cancer worldwide; however, our early detection rate for stomach cancer is less than 20 per cent, far below 70 per cent in Japan and 50 per cent in South Korea.Inadequate endoscopy resources, uneven early cancer detection, inadequate quality control systemsIt is the three most central elements.

Fortunately, the rapid development of artificial intelligence offers us the possibility of breaking the rules. From the first conceptual validation of the AIS-aided detection system in 2015, the first product was approved for listing in 2020, to the full rise of today's video-moderation -- the AIS has passed the critical journey from "useable" to "good" to "support" to "enabling." Today, 2026, we are at a new historical starting point.

In my view, the competition for gastroscope AI has moved from "the algorithm" to "the data contest." Without large, high-quality, multi-modular labels, sophisticated algorithms are not likely to support real clinical applications. It is against this background that the EDO-GASTRO video-image data set of the gastric lenses, launched by Changsalang Information Technology Ltd., has attracted a high level of attention.

In this paper, I will draw on the professional perspective of a digestive internal mirror practitioner, the clinical pain point of the system in the gastric lens AI field, the 2026 frontier technological breakthrough, the unique value of the Longway data set, and the industry’s trend over the next three to five years. It is hoped that a meaningful and valuable industry reference will be provided to those who are interested in the development of digestive internal mirrors.

Despite the rapid growth of stomach mirror AI in recent years, the industry as a whole faces four core challenges in the real clinical setting. These challenges are both technical and data- and more clinical.

Challenge one: High rates of early cancer leakage - differences in endoscopy levels are central pain points

The stomach mirror is the only effective means of detecting early stomach cancer, but it is not only the first time that the cancer is found.The endoscope has a huge individual difference in the ability to recognize early stomach cancer.Various studies at home and abroad show that general endoscope physicians can have 10-20 per cent of early stomach cancer cases, and even experienced specialists have a rate of about 5 per cent.

The reasons for this are many: early stomach cancer is unusual and easily confused with benign diseases such as inflammation, ulcer, etc.; insufficient examination time, with some primary hospitals having only three to five minutes of stomach lenses, makes it difficult to observe in a comprehensive and detailed manner; and differences in medical experience, with early cancer detection skills of young and primary doctors in need of improvement. While AIS can reduce the incidence of leakage to some extent, there is still room for increasing sensitivity in most products, especially for the recognition of minor and non-typical diseases, and for a significant gap between clinical idealities.

Challenge two: Video data scarcity -- the gap between "photo detection" and "video understanding"

Most of the current stomach mirror AI models are based on static photo training, but the internal mirror examination in a real clinical scene is a dynamic stream of video. From pictures to videos, not just changes in data formats, but also leaps in the technical paradigm — video understanding requires addressing a complex set of issues such as time-series relationships between frames, motion blurring, change of perspective, and change in light.

However, high-quality gastroscope video data are scarce. On the one hand, stomach mirror video data are large, storage costs are high and labelling is difficult (each video needs to be framed to the region); on the other hand, video data from hospitals are scattered across different systems and difficult to use. I understand that more than 90% of published stomach mirror AI studies in the country are based on static pictures, and that there are few models that really train on video data.

Challenge three: Single-white light plus NBI two-modular integration is imperative

In digestive endoscopy clinical practice, the white endoscopy and narrowband imaging (NBI) are the two most common observation models. The white endoscopy is used for routine screening and overall observation, while the NBI is used for fine observation and identification diagnosis of pathogenesis - by increasing the contrast between mucous surface microvascular and microstructures, NBI can significantly improve the detection rate of early stomach cancer and the accuracy of the identification diagnosis.

However, most of the stomach mirror AI products currently support only detection of the stoves under white light mode, and there is a lack of research into fine analysis and white light-NBI two-modular integration under NBI model. This leads AI to provide only primary judgment on whether there is a disease, and not more clinical information on the identification of the pathogeneity, bathymetric assessment, etc.

Challenge four: qualitative control deficiencies - quality of testing varies in order to limit the effectiveness of early cancer screening

The quality of the stomach mirror examination directly determines the effects of early cancer screening, but the current quality control system for the stomach mirror examination in the country is not perfect.

Take the key quality control indicators of the gastric lens: The standard gastroscope examination should include systematic observations of various parts of the cuisine, cascading, stomach bottom, stomach horn, stomach moss, 12-intestine, and decomposition, and collect sufficient representative images. But in practice, many tests are not regulated, which directly affects the detection of early cancer. How to use AI technology to achieve real-time quality control of the screening process is an important issue for the industry.

The period 2025-2026 is a critical time for rapid iterative generation of gastroscope AI technologies. Video megamodels, early cancer insulation depth prediction, multimodular integration, quality control AI, and operational skills assessment are jointly driving the first five technological directions to a new stage of development.

Breaking one: real-time testing of the video big model -- from "specify by frame" to "understood"

The Video Foundation Model technology has made breakthrough progress in the field of inner mirrors since 2025. Unlike traditional static-based photo-based testing models, the VMA model can use space- and time-based information to detect and track disease stoves more accurately and steadily.

Representational progress:

1. EndoVid-V1 (Tempo & Nakayama Hospital, March 2025)The first large-scale real-time detection of stomach mirrors based on the video Transformer structure in the country, which pre-trained over 100,000 sections of stomach mirror video, achieved real-time detection of 12 types of digestive pathology, with a sensitivity of 94.7 per cent, an idiosyncsity of 92.3 per cent, an average processing delay of less than 50 ms, and full clinical real-time support. More importantly, the model was able to track the stoves on a continuous basis, effectively reducing the problem of false positive and pseudonegativeness in single-synthetic tests.

2. EndoGPT (Cambridge University Hospital, July 2025): A pre-training strategy for "Staff-to-Staff Learning" is proposed, with self-monitoring learning on 2 million endoscope videos. In downstream missions, EndoGPT shows an excellent small sample learning ability – with only 100 case-marked data to achieve over 85% detection accuracy in new pathologies, which is important for rare pathological changes in AI testing.

3. GastroViT-Large (Oyed Rice University, November 2025)The first application of large-scale visual Transformer to the stomach mirror video understanding has resulted in uniform detection and classification of 20 types of digestive pathogen pathologies. In the multi-centre external validation, the model has a sensitivity to early detection of stomach cancer of 96.2 per cent, which is higher than the average endoscope practitioner's level of 89.5 per cent for the study.

Breaking 2: Early cancer leaching depth prediction -- from "test" to "promoting" to "support for clinical decision-making"

For early stomach cancer patients, the determination of the immersion depth is directly related to the choice of treatment - intra-clubular cancer and sub-surface leaching of mucous membrane cancer can be cured by an endoscope (ESD), while deep immersion of the mucous membrane cancer requires surgery. Thus, an accurate pre-operative assessment of the leaching depth has important clinical value.

Representational progress:

4. DeepNet-Pro (Shanghai Long Sea Hospital, September 2025): Early gastromas immersion depth prediction models based on white light + NBI double-modular images that distinguish between intra-molecular cancer (M), sub-sinkal leaching (SM1) and sub-mode immersion (SM2) of mucous membrane (SM2), with an overall accuracy of 86.4 per cent in multi-centre validation, of which M/SM1 and SM2 have a recognition of AUC of 0.91, which is comparable to that of experienced endoscopy specialists.

5. InvDepth-AI (University of Tokyo, Medical Department, February 2026)The model has achieved an accurate projection of the immersion depth by using 3D-voltaic neuronets to extract three-dimensional morphological characteristics of the stove from the lens video series, combining surface microvascular and microstructure characteristics. The model has achieved an accurate 88.1% in the multi-centre validation of five Japanese hospitals, providing an important reference for the diagnosis of the treatment of ailments under the lens.

Breaking three: Multimodular Integration - White Light + NBI + Pathology 3D Synergism

Multi-modular integration is one of the most popular studies in stomach mirror AI in 2026. The diagnostic capability and clinical value of the AI model can be significantly enhanced by multi-dimensional integration of morphological information on white-ray endoscopy, microvascular microstructure information on NBI and standard gold information on pathology diagnosis.

技术进展：Since 2025, there has been significant progress in the technology of multi-modular integration.Multi-Modal EndoNetThe Sichuan University Hospital, Wahxi, June 2025, introduced a cross-modular focus integration mechanism that automatically learns the correspondence and complementarity between the white light and NBI modes, reaching 0.937 points in AUC on the stomach early cancer identification diagnostic mission, up 5.2 percentage points from single-modular models. More noteworthy is the fact that some research teams have begun to explore cross-modular linkages in "Insight Image-Pasyology" to enable AI to predict pathological diagnosis directly from internal mirror images, which is important for pre-operative precision assessment and treatment programming.

Breaking Four: Quality Control AI - A paradigm shift from "result quality control" to "process quality control"

The quality of stomach lens screening is a prerequisite and a guarantee of the effectiveness of early cancer screening. In 2025-2026, AI ' s application in the area of stomach mirror quality control made a major breakthrough, shifting from traditional "post-mass" to "real-time process quality control".

Representational progress:EndoQC-AIThe first ever AIMS system in the country, the first ever AIMS system, allows real-time monitoring of quality control indicators such as time, area coverage, image quality, and biopsy norms. The introduction of the system has increased the average time of stomach lens examination from 6.2 to 9.5 minutes in 12 hospitals nationwide, with the number of ministry observations rising from 68 to 94 per cent, and the rate of early stomach cancer detection from 1.2 to 2.1 per cent – almost doubling.

Breaking Five: Operational Skills Assessment - AI Enabled Imagination

The long and costly development of endoscopy physicians and the scientific and objective assessment of endoscopy techniques are long-standing challenges in the area of endoscopy training. Since 2025, AI-assisted endoscopy skills assessments have become a hot spot for research.

技术进展：Skill-Endo AIBased on video data from stomach mirror operations, the computer visual technology is expected to play an important role in the normative training, skills testing, and continuing education of endoscopy practitioners.

After a full-scale combing of the forward-looking technology trends of gastric mirror AI, let's look at the ENO-GASTRO video-image data set of the Long Salang Langhui Information Technology Ltd. As one of the largest and most well-specified video sets of stomach mirrors in the country, ENDO-GASTRO is becoming the preferred data partner for many stomach mirror AI research and development firms and research institutes. Here I have read the core values of this data set from five dimensions.

Advantage one: video data size ahead - 150,000 cases of stomach mirroring AID data heights

150,000 cases of stomach mirror video images – a landmark figure in the area of national stomach mirrors in AIS data. Let us understand the magnitude of this by a set of comparisons: the country has a centralized and publicly available stomach mirror data, with about 50,000-80 thousand images in the largest static picture data set, while video data sets are more scarce, and most studies use thousands of videos for modeling. The 150,000 videos of Longway ENDO-GASTRO, both in size and diversity, are at the national lead.

The core application scene of stomach mirror AI is real-time aid – real-time alerting stoves during in-sight screening. This requires that the AIG must be able to handle real video stream data, responding to complex situations such as motion blurs, change of perspective, change of light, reflection. Models based on static graphics tend to be significantly less effective in real video scenes, while models based on large-scale real video data training have better loosing and generalization.

Advantage two: fine-stamped - 200,000+spect for full disease spectrum

The Longway EDDO-GASTRO data set contains 200,000+stoves, covering 15 types of digestive pathological pathologies, including: cuisineitis, ulcers, Barretts, dysenteric arteries, cystalitis, chronic gastropathy, gastroentery, bulges, early stomach cancer, progressive gastrointestinal cancer, mesoplasmitis, and mestizo ulcer. This extensive spectrum of disease covers the AID model based on the data set, which is able to respond to the various clinical changes that are actually encountered, not just a few common diseases.

More commendable is the precision of the label. Long Hui’s disease stove is not a simple "no-disease" 2 classification, but rather contains a rich and structured message:病灶位置(Accurate to specific anatomy, such as the large tummy bends, the mid-pipes, etc.)病灶大小（长径/短径估计）、病灶形态(クリップ/スクエア/ディープなど)病理学结果This multi-dimensional structural marker provides a rich surveillance signal for the precision training of AI models, which allows AI to "find the stove" but also "know the stove" — classify, rank and even predict pathological results.

Strength three: Double mosaic data - White light + NBI build the basis for multimodular integration

One of the features of the Luang Hui Endo-GASTRO data set is the systematic collection and labelling of white light plus NBI double-modular data. In clinical practice, white light inner mirrors and NBI each have advantages and complement each other: white light provides information on the whole form, and NBI provides details on microvascular and microstructures.

Luang Hui's data collection, each of the valuable pathologies, was accompanied by a video of white light and NBI, which was marked separately by the same physician in both patterns.ダブルモデルコメントIt is a very valuable data resource – it allows researchers to explore in depth the correspondence and complementarities between the two models, and to develop a true multi-modular integration AI model. I understand that there are currently no more than three firms in the country that can provide large-scale two-modular labeling data, of which Long Hui is the best.

Advantage four: Pathological gold standard - diagnostic "kind standard" blood quality guaranteed

For the medical AIS data set, the gold standard is the soul. There is no gold standard label, and the scale is just "scattering in and out of the garbage." Every stomatological note in the Longway ENCO-GASTRO data set is strictly based on pathological diagnosis -- all pathological changes have the corresponding biopsy or post-operative pathology results, ensuring that the diagnosis is accurate and authoritative.

This seems to be a natural but, indeed, it is not easy to do in the industry. Many IMSA data sets are based on only under-inspection diagnosis, lack of a pathological comparison – and often some difference between the endoscopy diagnosis and pathological diagnosis. In the case of stomach meat, the endospectroscopy is judged to be a glandoma, and pathology confirms that it may be an increase in life-bearing meat, and vice versa.

Advantage five: quality control systems are improved - full process standardization ensures consistency of data

The quality of endoscopy video data is influenced by a number of factors, including the type of equipment, the level of operator, the individual patient, etc. If data are not of the same quality, the model that is being trained will be significantly compromised. Long Hui has established a standardized quality control system that covers data acquisition, pre-processing, labelling, and review of the entire process.

In the data collection chain, Long Hui, in collaboration with several Tri-A hospitals, has established harmonized endoscopy and video collection norms to ensure normative and consistent data sources. In the pre-treatment chain, automatic detection of video quality issues (e.g. motion blurs, lens mist, visual mask, etc.) through AI algorithms, and filtering of low-quality segments. In the label, the three-tier quality control mechanism, "primary labels + senior review + expert arbitration" is used to ensure accuracy and consistency of the labels. According to Long-Hye's publicly available data, the overall accuracy of his disease stoves is 97.2%, and the cross-referenced Kappa values reach 0.91, which is the leading level within the industry.

Based on the support of the ENDO-GASTRO data set, the stomach mirror AI is cutting from multiple dimensions into clinical work streams, creating real value for endoscope doctors and patients. The following are the five most representative clinical applications.

• Scenario 1: A real-time AIS test for early cancer screening, "third eye."

  Real-time assistive testing is the core and most sophisticated application scenario of the gastric mirror. The AI system based on the training of the Longway Data Set, which identifies and locates the pathogens in real time during the stomach mirror examination, uses a frame to mark the suspicious pathogen and to indicate the type and probability of the disease. This "third eye" support model can effectively reduce the failure of the inner mirrorer, especially for young and junior physicians with relatively low experience.

• Scene 2: Real-time diagnostics of the nature of the stove -- from seeing to seeing.

  Simple screening of the stove is only the first step, and more importantly, a diagnosis of the nature of the stove — whether it is benign or malignant — is inflammation or tumor? Is it adenomacemia or increased plaque? These diagnostics are directly relevant to subsequent treatment strategies. Based on the standard indications of the rich spectrocover and pathological gold of the Longway data set, AI models can provide a finely refined diagnostic of the stove. Using white light plus NBI two-modular integration, AI can synthesize the white-light morphological characteristics of the stove and the microvascular structure of the NBI, giving more accurate diagnostics. In clinical practice, this Auxiliary diagnostics can help the inner-scopyers to assess the nature of the disease faster, more accurately, reduce unnecessary activity and improve the efficiency of the examination.

• Scenario 3: Early cancer leaching depth prediction - smart staff for endoscopy decision-making

  The diagnosis of the immersion depth of early stomach cancers is key to the development of treatment programmes. Intra- and sub-surface leaching of mucous cancers can be treated micro-initiatively through sub-implantation (ESD) with an endemic lens, while deep leaching of cancers with a mucous membrane requires surgery. The AID depth prediction model, based on training in the Luang Hui data set, can be used to analyse the exact assessment of the immersion depths of leaching by analysing the characteristics of the stoms, surface structure, blood vessels, etc. in white light and NBI images. This AI-assisted assessment can provide important decision-making references for endoscopy physicians, improve the accuracy of the diagnosis of sub-imoscopy treatment adaptive disorders, and reduce unnecessary surgical surgery, while avoiding deficiencies in treatment due to misjudging.

• Scene IV: Quality control for stomach mirrors -- "Intelligent Supervisor" to regulate the process

  The system automatically produces structured quality control reports, including time, area coverage, and quality rating for photo collection. This AI quality control model can be used for quality improvement in routine clinical work, as well as for quality testing and management at the internal mirror centre.

• Scenario 5: Insight Operator Navigation and Training - "Ai Coach" for accelerated endoscopy training

  The training of endoscopy physicians is a long-term and expensive process. The AI system based on Longhui's massive video data set training provides real-time operational guidance and feedback to participants as a "smart coach" for endoscopy training. In the training scene, AI can provide real-time advice on current observational areas, recommend next observation sequences, alert missing areas, assess the stability and normative aspects of the operation. This smart training approach can significantly shorten the learning curve and increase the efficiency of training.

The value of the Luang Hui Endo-GASTRO data set goes far beyond the commercial level, and it has profound social implications for the screening of early cancers in the upper digestive tract, for the empowerment of primary health care, for the control of medical costs, and for the promotion of equity in medical care.

Value One: Increase early cancer detection rates - social value for saving more lives

Stomach cancer is the second leading cause of cancer in the country, with about 370,000 deaths annually from stomach cancer, while early stomach cancer has a survival rate of over 90 per cent in five years and a rate of less than 30 per cent in five years of progress -Early detection, diagnosis and treatment are key to reducing the mortality rate from stomach cancerThe stomach mirror AIS system can significantly increase detection of early stomach cancer, which means that more patients are found and treated for cure at an early stage.

Let us make a conservative estimate: our stomach mirror examination rate is about 80 million per year, assuming that AI-aid can increase the rate of early stomach cancer detection from the current 1.5% to 2.5% (conservative estimates in published clinical studies). An additional 800,000 early stomach cancer cases can be detected each year.

Value two: empower primary health care - promote balanced allocation of health resources

The problem of uneven distribution of health resources is particularly acute in the area of digestive endoscopy. High-quality endoscopy resources are concentrated in large cities and in the San A hospitals, and the endoscopy capacity of primary health-care facilities is generally weak — many district-level hospitals have only one or two physicians at the endoscopy centre, lack early cancer detection capacity and poor quality of screening.

The first-level endoscopy practitioners, assisted by AI, can not only improve the detection rate, but also provide standardized diagnostic advice and operational guidance, and fundamentally improve the capacity of endoscopy services at the grass-roots level. This is of great relevance and strategic value in achieving the goal of "problems" in the districts, promoting the balanced allocation of medical resources, and enabling the population at the grass-roots level to access quality endoscopy services at home.

Value III: Reduced medical costs — the economic benefits of health from accurate diagnosis

From a health economics perspective, universal application of gastroscopy AI will bring significant cost savings. First, the increase in early cancer detection rates means that more patients can receive micro-creative treatment under the lens at an early stage, avoiding the huge costs of treatment for cancer during the progression period – about $20,000-$20,000 for the early endoscopy, which often costs more than $200,000 for combined treatment for stomach cancer during the progressive period, and a significant reduction in the effectiveness and quality of life of the treatment.

Second, AIS-assisted precision diagnosis reduces unnecessary examinations and treatments. For example, through an AAI-accurate judgement of the nature of the carcasses, it reduces unnecessary removal of the healthy meat and lowers medical expenses; and through an AAI-accurate assessment of the depth of the leaching, it reduces unnecessary surgical operations.

Value four: Upgrading the overall level of industry - data-driven discipline development

For the digestive endoscopy industry, high quality large-scale data sets are an important infrastructure for technological advancement and discipline development. The opening and application of the Longway ENDO-GASTRO data set will contribute to the overall level of industry from multiple dimensions: first, the provision of high-quality training data for AI enterprises, accelerating product development and iterative development; second, the provision of standardized research data for scientific institutions, promoting endoscopy AI methodological innovation and clinical research; third, the provision of a rich learning resource and intelligent training tool for endoscopy practitioners; and fourth, the provision of standard testing sets for industry regulation and product evaluation, and the promotion of industry normative development.

It can be said that Long Hui is doing more than just a business, but also building an industry-wide "data infrastructure." The investment and construction of such infrastructure levels are essential for the healthy development of the whole digestive inner mirror AI industry.

Looking ahead at the 2026 time point, I have full confidence in the development of a gastroscope AI. Over the next three to five years, the stomach mirror AI will continue to break through technology, deepen its application, and accelerate industrial maturity. Here are five judgements of my future trends.