The New Era of Enteric Mirrors: how 200,000 video data sets remodel the early screening system for rectangular colon cancer

By the way, let's start with a sobering set of data: colon cancer is the second most common disease in the country and the fifth most fatality in the country, with about 550,000 new cases and 280,000 deaths each year. Compared to developed countries, our early detection rate of colon cancer is less than 30 per cent, well below 60 per cent in the United States and 70 per cent in Japan.Insufficient supply of intestinal lenses, uneven levels of physicians, and poor quality of examinationsIt is the three core bottlenecks.

In recent years, the rapid development of artificial intelligence has created new hope for early screening for colon cancer. From saloon detection to adenoma classification, from retroscopy control to recognition of the bowel, AI is examining multiple dimensions of enabling colonoscopy. After almost a decade, the colonoscope has moved from laboratory to clinical, from conceptual validation to conventional application.

However, all AI technology competition ultimately points to the same core - data. Without large, high-quality, multimodular labels, advanced algorithms can hardly support a real clinical fall. It is against this industry background that the EDO-COLONO video-image data set of ENCO-COLONO, launched by Long Salang-Hye Information Technology, has attracted my attention. 200,000 cases of enteric lens video, 300,000 + valentoids, 10 intestinal pathology, adenomas/cartological pathology, white light + NBI+ gills, are behind these figures, which are profoundly understood of the prognostic cancer early screening cause and a quest for data quality.

In this paper, I will draw on the professional perspective of a colon surgeon, the clinical pain point of the system in the field of colon cosmology, the frontier technological breakthrough of 2026, the unique value of the Longway data set, and the trend of industry over the next 3-5 years. I hope to provide an in-depth and valuable industry reference for colleagues who are concerned with the development of colon cancer early and enteroscope AI.

Despite the remarkable progress made by intestinal cosmolar AI in recent years, the industry as a whole still faces five core dilemmas in the process of the early screening of real clinical fall and colon cancer. These dilemmas involve both the technical and the data, the quality control and the health economics.

Trouble One: High detection rates -- the double challenge of "see" and "see"

The test of colonoscopy is the gold standard for finding circa and early cancer, but the problem of scavenging remains widespread.The prevalence of staminas from conventional colonoscopy is as high as 20-30%., where flat meat, small meat .<5mm）和位于皱襞后方的息肉漏诊率更高。这些漏诊的息肉中，有一部分可能是进展期腺瘤甚至早期癌，它们的漏诊直接关系到患者的生命安全。

The reasons for this leak are many: inadequate intestinal preparation and faeces to block impact observations; slow retrospection; inexperienced operators with limited capacity to identify flat and micro-modified diseases; complex colon anatomical structures, with corsets and bends easily becoming visible blind areas. While AIS can reduce leakage rates to some extent, there is still room for improvement in the detection performance of most products, especially for flat meat, small and special gravitational meat, and no small gap between clinical idealism.

Adventure II: Undesired adenoma classification -- "uncut" decision-making dilemma

The discovery of sting flesh is only the first step, and more importantly, the determination of its nature — whether it is accumulative or adenomas — is it a tubal adenoma or acne adenoma? Is there a high-level epidemoma? These judgments are directly related to the choice of treatment and the development of follow-up strategies.

The determination of the nature of the holocaust under the inner mirror is currently largely based on the experience of the inner-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthetic-synthesis-styping-synthetic-synthetic-synthetic-styping-synthesis-stystaltic-styping-sy-synthetic-synthetic-synthetic-sty. The improvement of the accuracy of the judgment of the ceresynistic-synthsynthic-synthic

Trouble Three: The quality of the retrospect is a problem -- the quality control problem of "unchecked"

A lot of research has confirmed that the time of retrospection is one of the most important factors influencing the incidence of adenomas - the longer the resort, the higher the adenomas. The standard time for resortation recommended in the international guidelines is at least six minutes, but in practice, many endoscope centres have averaged less than four minutes. Oversighting directly results in higher incidences of salivation and adenoma, which seriously affects the effectiveness of colon cancer screening.

Traditional quality control relies mainly on after-action spot checks and manual assessments, which are inefficient, limited coverage and less objective. Real-time, automatic and objective assessment of how to achieve endoscope quality is the key to improving colonoscopy quality and adenoma detection.

Scenario 4: scarcity of data marked - data bottlenecks in salivating AI

The data challenges for colon mirrors are even more severe than for gastroscopes. On the one hand, colonoscopy videos are longer, data are larger and labeled at higher cost; on the other hand, saloon meat is in a variety of shapes, sizes and locations, and is much more difficult and complex than static images.

The current internationally open intestinal lens data sets are generally smaller: CVC-ClinicDB has about 612 pictures, Kvasir-SEG has about 1,000 pictures, and HyperKvasir has about 110,000 images (but only with limited label levels). The situation is even more acute in the country — high-quality, large-scale intestinal lens video data are scarce, which directly constrains the performance of the intestinal coscopy AI model and product overlay. Many AI companies rely on self-mining data for model training, with limited data and insufficient diversity, making model-wideization less able to stabilize their real clinical scenes.

Scenario 5: Poor screening path -- "Sift for what, who, how" systemic problem

The screening of colon cancer is a systematic project that involves a number of aspects, including screening of the population, initial screening methods, precision pathways, follow-up management. The current system of screening for colon cancer in the country is not perfect, and the screening path is yet to be optimized – which groups should be prioritized – what methods should be used for screening first (scrub blood? faeces DNA? blood markers?) how do first-scrudents conduct a colon-scopy examination? How do the screening of the saloons be managed and followed up?

AI technology is expected to play an important role in optimizing the path to screening for colon cancer – from high-risk group identification, primary screening risk layers, to colon-scopy quality control, and meat treatment decisions, to follow-on programming, where AI can provide intelligent support at all stages of screening the entire process. But all this is achieved on the premise that there is real world data on a large, high-quality, process-wide scale.

The period 2025-2026 is a critical period for rapid chronography of intestinal cortex technology. The macro-speculation model, real-time classification of adenomas, brain assessment of retrospect, recognition of the appendix, and optimization of the path to screening for colon cancer - these five technological directions are jointly driving the intestinal cortex AI into a new stage of development.

Breaking one: Big model of sabbatical detection -- leaps from "synthetic detection" to "video understanding"

Since 2025, as video-mode technology matures, the colon-scopy meat detection is leaping from traditional "synthetic" to "video-understanding." The video-mode allows for more accurate and stable detection and tracking of saloon meat using both spatial and time information, effectively reducing omissions and errors.

Representational progress:

1. Polyp-ViT-Large (MSA Asia Institute & Middle Mountain Six College, February 2025)The first large-scale in-country cortex detection model based on a large video Transformer structure, which pre-trained over 80,000 in-intestinal mirrors, achieved real-time detection of various types of salivating meat, reaching 95.8 per cent sensitivity and 93.1 per cent speciality, with processing delays below 40 ms, fully meeting the clinical real-time support needs. In the multi-centre external validation, the model increased the adenoma detection rate (ADR) by 34 per cent in relative terms and with significant results.

2. Endo-Colon Foundation Model (Méo Clinic, August 2025)The basic model of self-supervised colonoscopy, which is based on half a million sections of colonoscopy video, allows for a fine-tuning of various downstream tasks, such as ration meat, holocaust classification, quality control, etc. The model is 97.2% sensitive to meat stinging, and is particularly good at flat meat testing, with 93.5% sensitivity, which is nearly 15 percentage points higher than the traditional model.

3. ColonGPT (Queen Mary College, University of London, December 2025): The introduction of the "Spatial and Space Focus Integration" mechanism, which focuses on both spatial and temporal dynamics of the saloon meat, has effectively enhanced the robustness of complex situations such as motion blurring and change of perspective. In a comprehensive assessment of five open data sets, the mDice of ColonGPT reached 0.867, updating the world record at the time.

Breaking through II: Real-time classification of adenomas -- critical spans from "test" to "diagnosis"

Spectrum testing is only the first step, and the ability to judge the nature of the stingling (incremental vs adenomas) is more clinical. In 2025-2026, with advances in multimodular imaging techniques and AI algorithms, there was a major breakthrough in real-time classification of adenomas.

Representational progress:

4. AdenoClass-AI (Chongshan Hospital, University of Jordan, May 2025): A real-time classification model of adenoma based on white light + NBI dimodified images, which distinguishes between antenna, adenoma (pipecarb, velvet, velvet tumour) and early cancer. In the multi-centre external validation, the model identifies AUC as 0.926 for adenomasary meat, and AUC as 0.943 for early cancer, which is comparable to the level of experienced inner-scopy specialists.

5. i-Scan AI Classification (University of Medicine, Vienna, March 2026)The model achieved an overall classification accuracy of 89.7% in the European multi-centre study, with a negative projection of 94.1% for adenomas, meaning that the AAI judgement that over 94% of the acetoids are not really adenomas and avoid unnecessary removal.

Breaking Three: Retro-Quality Smart Assessment - AA Enabling Power Control Core Catcher

Retrospective quality is a key factor in the effects of colonoscopy and is the core application scene of AI quality control. Since 2025, the rescoping-quality intelligence assessment technology has made significant progress, moving from single-time monitoring to multidimensional integrated assessment.

Technology Progress：ColonQC ProMore importantly, the system can identify the arrival of the appendix in real time, and automatically record the arrival of the colon, which is one of the most important qualitative indicators for colon examination. The system was introduced in multi-centre clinical testing nationwide, with the average exit time for the colon lens increased from 5.2 to 8.7 minutes, the adenoma detection rate increased from 23.1 to 31.8%, and the arrival rate of the cavity increased from 91% to 98%.

Breaking four: Auto-Recognition of the appendix - Checking the integrity of the smart proof

The arrival rate of the appendix is a core indicator of the integrity of colonoscopy - an examination that does not reach the colon can be considered incomplete. However, confirmation of the arrival of the colon is long-term dependent on manual records, subjectivity, data inaccuracy, etc. The emergence of AI technology provides new solutions for auto-literacy.

Technology Progress：Since 2025, blinding techniques based on in-depth learning have become more sophisticated.CecumAIThe technology can be used not only for quality management, but also for the examination of reports for automatic generation - key parameters such as system auto-recording of goggles, time of goggles, time of goggles, time of exits, etc., and reducing the workload and error of manual records.

Breaking five: Optimizing the path to screening for colon cancer - Accurate screening system driven by AI

The rectum cancer screening is a multi-stage system project, and AI technology is moving from a single-link aid tool to a process-wide optimization engine.

Representational progress:CRC-Screen AIThe system, based on large-scale mass screening data training, enables individualized screening strategies for populations at different risk levels to optimize resource allocation while ensuring the effectiveness of screening. In modelling, the introduction of an AIE-optimal screening path can increase the early detection of colon cancer by more than 25% with the same resources.

After the system combes the forward-looking technological trends of enteroscope AI, let us look at the ENDO-COLONO video-image data set of Long Salangye Information Technology Ltd. As one of the largest and most well-specified in-country intestinal mirror video data sets, ENDO-COLONO is becoming a core data partner for many enteroscopeAI research and development firms and research institutes. Here I have read the core values of this data set from five dimensions.

Advantage one: the industry at scale is ahead - 200,000 cases of video-building barriers to the intestines

200,000 cases of intestinal mirror video images – the largest in the country’s public information available today – are the largest in the country’s intestine mirror video data set. Let’s use a set of comparisons to understand the gold content of this scale: the largest internationally in the open intestinal image data set, Hyper Kvasir, is about 110,000 images (from about 4,000 tests), while the data used in published research in the country are generally in the range of tens to tens of thousands.

The significance of scale is not only "big" but also "full" and "precision." The magnitude of 200,000 cases means a sufficiently rich diversity — different equipment, different hospitals, different physicians, different groups of people, different types of disease, different degrees of severity — which is fundamental to modelability. For an in-depth learning model, there is a near-readarithmic relationship between data volume and performance, with 200,000 cases of size sufficient to support a truly big video-film training in the form of a large-scale intestines model, which is the most scarce core resource in the industry today.

Advantage two: Spectrum is fine - 300,000 + full disease spectrum covered

The Longway EDDO-COLONO data set contains 300,000 + crotoids covering 10 types of intestinal pathological changes, including: increased gravitational meat, inflammatory meat, mestizo meat, tube adenoma, adenoma, tubular adenoma, scanic adenoma, high-level epidermal tumors, early colon cancer, procinary carcinoma, and procinary cancer. This extensive spectrum of pathological changes allows AI models based on the data set to respond to various types of changes encountered in clinical practice, not just to simple tests of "no carcasses".

More commendable is the depth and structure of the label. Each of the placards of Long Hui contains a wealth of information:解剖位置(Accurate to colon sections and specific locations)大小估计(Long/short)形态分型(in Paris)Pathology ResultsThis multi-dimensional structural marker provides a rich monitoring signal for the precision training of the AI model. In the case of adenomas, the data concentration not only indicates adenomas but also sub-species such as tube-state adenomas, tactile adenomas, tubular acoustic adenomas, and the upper skin tumours - so that the AI model can learn to learn the fine differences between sub-types and achieve a true "optical active examination".

Advantage three: TIMAGE DIVERSITY - White Light + NBI + Lactose Build Multimodular Data Base

A key feature of the Luang Hui Endo-COLNO data set is the systematic collection and labelling of white light + NBI+ gill trimester data. Each of these three molluscs has advantages and complements each other: white light provides overall morphological observations, NBI enhances the contrast between microvascular and microstructures, and fat dyes can show more clearly the microforms and indentular pathologies of the mucular surface.

Luang Hui's data collection, which is used to synchronize videos of white light, NBI and tipo-chromium dye patterns for each clinical saloon, is then used to match this.Three-modular PairingData sets are scarce both at home and abroad. It provides a valuable data base for research and development of multi-modular integration AI – researchers can explore in depth the complementarities and integration strategies between the different models, and develop more accurate AID diagnostic models. From a clinical point of view, a three-modular integration AI implies higher diagnostic accuracy and more reliable clinical decision support, which is important for reducing unnecessary carving and improving the effectiveness of clinical interventions.

Advantage four: Pathological gold standard - "Gold standard" certification for adenoma/cancer spectrometry

For the intestinal lens AIS data set, the pathological gold standard is the core value. Without a pathological endoscope, the diagnosis is questionable - there is often a difference between the eye judgement under the endoscope and the pathological diagnosis. Each of the signs of the ENCO-COLONO data set has the pathological results, all of which are based on pathological diagnosis, which essentially guarantees the accuracy and reliability of the data.

Particularly noteworthy is the severity of ryanhui’s adenoma styma styma. Each adenomas are detailed pathological styma (including tissue type (pipeline/fluar/breath-like fur), heteroacity (low/high-level), cancer mutation, etc. This finely refined pathological styma, which enables AI models based on the training of the data set to learn more essential pathological characteristics, thus achieving a precise judgement of the nature of the valentine. For early screening of adenomas, the precise distinction between adenomas and non-adenoma styma styma, recognition of upper-synthetic changes and early cancers is key to determining the effectiveness of the screening.

Advantage five: Full process quality control - complete data links from lenses to mirrors

Another important feature of the ENDO-COLNO data set is its whole-process data collection and quality control system. Unlike many data sets that collect only video footage containing the disease stoves, the ENDO-COLNO data set keeps a full intestinal examination video, from the beginning of the goggles to the end of the exit mirror, including a complete record of the stages of the goggles, the blind, the backsight.

This whole process data is essential for quality control AI development. These data will provide training in multiple quality control AI models, such as retrospect time monitoring, respect stability assessment, goblin arrival recognition, and area coverage assessment. In addition, Longway has established a rigorous data quality control process: from standard data acquisition development to automatic video quality screening, to manual labelling and multi-level review, each has clear SOP and quality control indicators. According to publicly available data, the disease stoves of the Longecomposer data set have 96.8% accuracy, cross-referenced Kappa values of 0.90, which are industry-leading.

Based on the support of the ENDO-COLONO data set, the enteroscope AI is cutting from multiple dimensions into clinical workflows and protocinary cancer screening systems, creating multiple values for endoscopy physicians, patients, and the public health system.

• Scene one: Real-time test for salivating meat -- Smart partner for dropping the drop rate.

  Real-time assistive testing is the core and most sophisticated application of intestinal mirror AI. The AI system, based on the training of the Longway Data Set, can identify and mark carnival meat in real time during colonoscopy retrospect, displaying the saloon area in the inner mirrors with a frame or masked high-lighted image, and indicating the size and type of probability of the saliva. This "second-eye" support model can effectively reduce salivation rates – especially for flat meat, small-silt meat, and crinkled meat – and the AIS’s support value is particularly significant.

• Scene 2: Real-time classification of adenomas - Smart Assistant for Optical Work

  The AA adenoma classification system, based on the training of the Longway Data Set, can provide real-time character judgement of the saloon char, by analysing the characteristics of multiple-modular images such as white light, NBI and curry, distinguishing between increased valentine, adenoma and early cancer, and giving corresponding confidence. This "optical activity" capability has important clinical application: it can be considered that no need to remove the disease, and that unnecessary medical expenditure and risks are reduced; and that more active treatment and follow-up are needed to determine the disease's transformation to high-level adenoma or early cancer. Studies have shown that high-quality AIS is accurate enough to reach an experienced inner-sight expert level, providing an important reference for clinical decision-making.

• Scenario three: Retro-quality intelligence control -- "Intelligent Supervisor" for quality check.

  The ACMS, based on the development of the Longway process video data set, allows real-time, comprehensive quality monitoring and assessment of the retrospection process: autocalculating the time of resuppression, assessing the stability of the resupplied mirror, identifying the observation coverage of the various parts of the colon, testing the cleaning of the enteric tract, judging the quality of the vision, etc. The system can automatically alert when the resuppression is too fast, recommending re-observation in areas that are inadequately observed, and automatically generating structured quality control reports at the end of the examination. This ACMS model has been confirmed by several clinical studies that significantly improves the rate of adenoma detection. For the management of the ICEM, AC can also be used as a key tool for improving the overall service capacity of the ICEM in a variety of areas, including the performance appraisal of physicians, continuous quality improvement, and training effectiveness assessment.

• Scene 4: The appendix to auto-confirmation -- "A smart witness" to check integrity.

  The cadets are the core indicator for measuring the integrity of colon examination. The AID, based on training in the Longway Data Set, automatically confirms whether or not to reach the cadets by identifying the signo-section structure, such as the retrievable valve, the appendix opening, and automatically records the arrival of the cadets. This function appears simple, but in practice has important clinical and management values: at the clinical level, the cadets confirm that close-end colonosis due to failure to reach the cadets can be avoided; at the managerial level, auto-recording is more objective and accurate than manual statistics, and can be one of the core indicators of central endoscopy control; at the patient level, the clear arrival of the botage in the examination report enhances patient ' s confidence in the quality of the examination.

• Scenario 5: Optimizing the path to screening for colon cancer -- the accurate screening of the "AI brain"

  Based on Longway data sets and accompanying population screening data, it is possible to develop an AIMS-based optimized screening path for colon cancer, from high-risk group identification (risk stratification based on demographic information, lifestyle, family history, etc.), to initial screening programme recommendations (different primary screening strategies for different risk groups), to colon-scopy quality control (AI-aided improvement of the quality of screening colons), and to follow-up screening management (individualized follow-up programmes based on the types of disease detected). This AIS-driven precision screening system can maximize the early detection and cost-effectiveness of rectal cancer in a country with limited medical resources, which is particularly relevant for a large population with relatively limited medical resources.

The value of the Luang Hui Endo-COLONO data sets goes far beyond the business sphere of individual enterprises, and has far-reaching implications for the promotion of protecs, the upgrading of primary health care, the control of medical costs and the improvement of the public health system.

Value I: Prevention of colon cancer - "Physical value of undiseased"

The development of colon cancer is a multistep process of increasing glandoma cancer from normal mucous membranes, which usually takes 5-10 years or more. This means that we have enough time to intervene -- if we can detect and remove adenoma before it changes, we can effectively prevent it.colonoscopy screening is the most effective current stage-I prevention of colon cancer。

The AI-assisted detection system, based on training in the Longway Data Set, can significantly increase the incidence of adenomas. Let's make a conservative estimate: we have about 30 million colonoscopy tests per year. If AI-aid increases the adenomas detection rate from 25 to 32.5% (a relative increase of 30%, conservative estimates in published studies), an additional 2.25 million adenomas can be detected each year. If 80% of the adenomas are removed in a timely manner, based on adenoma rates of 5-10%, we can prevent the occurrence of 9-180,000 straight-intestinal cancers each year.

Value two: empowers primary care - the technological underpinning of the intestine screening process

The difficulty of rectal cancer screening lies in coverage – more people of appropriate age must be given high-quality colonoscopy screening, which must be sunk to the grass-roots level. However, the general weakness of the endoscopy capacity in our primary health-care facilities and the inability of many district-level hospital centers to conduct high-quality colonoscopy independently constitute important bottlenecks to the increase in screening coverage.

The system of colonoscopy, based on training in the Longway Data Set, allows for the endoscopy diagnostic capability and quality control standard "codes" of top-level hospitals to be incorporated into algorithms and digitized into the base level. The primary endoscope physician, assisted by AI, not only improves the rate of salivation and diagnosis accuracy, but also provides standardized operational guidance and quality control, and fundamentally enhances the capability of endoscopy services at the grassroots level. This is of strategic importance in promoting the sinking of colon cancer screening, achieving the goal of "early cancer screening in counties" and providing access to quality screening services to a broad range of people at home.

Value III: Reduced medical costs — benefits of accurate screening in the economics of health

From a health-economic perspective, AI-enabled colon cancer screening is a significant cost-effectiveness. First, early treatment can significantly reduce treatment costs – the cost of treatment for early colon cancer is around $30,000, while comprehensive treatment for late colon cancer tends to exceed $300,000, and treatments and quality of life are far less effective than earlier. Second, accurate diagnosis of AI-assisted care can reduce unnecessary medical expenses – through an accurate AI judgement of the nature of meat, unnecessary removal of healthy meat can reduce medical costs and risk of complications.

More importantly, at the population level, the cost of advanced treatment can be saved by several yuan per $1 investment in colon cancer screening. According to the national health economics study, the cost-effectiveness of colon cancer screening is about $20-$30,000 per quality-adjusted life year (QALY), well below the WHO recommended threshold of three times per capita GDP, and is a cost-effective public health intervention.

Value four: Upgrading industry - data-driven discipline building and industrial development

For the digestive endoscopy industry and AI industry, high-quality large-scale data sets are the core infrastructure that drives technological progress and industrial development. The opening and application of the Longway ENCO-COLONO 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 methodological innovation and clinical research in enteroscopy AI; third, the provision of a rich learning resource and smart training tool for the training of endoscopy physicians; and fourth, the provision of standard testing sets for industry regulation and product evaluation, and the promotion of industry normative development.

In addition, the accumulation and application of intestinal cosmopolitan data will facilitate the cross-fertilization of relevant disciplines - the multidisciplinary intersection of digestive endoscopy, artificial intelligence, epidemiology, health economics, etc., which will lead to more innovative research orientations and application models that will drive the development of the whole prostointestinal cancer control industry.

Looking ahead at the 2026 time, I have confidence in the development of colon cosmoscope AI and colon cancer screening. Over the next three to five years, colonoscopy AI will continue to break through technology, deepen its application, and accelerate industrial maturity. Here are five judgements of my future trends.