I. Expert introduction: Data is the "first principle" of medical AI
As a doctor who has been working on gynaecology clinical and research for more than 30 years, I have experienced three revolutionary leaps in cervical cancer control from "eye observation" to "cell screening" to "HPV testing age." And today, at the 2026 time, I am convinced that we are at the beginning of the fourth revolution — the era of precision screening driven by artificial intelligence.
However, in the last five years, I have witnessed the emergence of numerous AI teams that have gone into the field of medicine with ambition and retreat. Their failure is often due not to the lack of advanced algorithms, but to the lack of medical data that are truly high-quality, large-scale and certified by pathology standards.The data ceiling determines the ceiling of AI, and the quality of the data determines the clinical bottom line.。
In this article, I will look at the core pain points in the field of diagnosis of cervical transformations, the breakthrough in frontier technology in 2026, and the strategic value of high-quality medical data sets for the industry as a whole. I believe that understanding the value of data is a way to truly understand the future of medical AI.
II. Industry pains: triple dilemmas and data bottlenecks for cervical cancer screening
2.1 Structural discrepancies between screening coverage and quality
China has about 110,000 new cases of cervical cancer every year, and 59,000 deaths are the second highest cancer among women aged 15-44. Although cervical cancer is the only malignant tumor that is currently identified as the cause of the disease and that can be prevented by early screening, the reality remains stark: the national cervical cancer screening coverage rate is only 55%, far below the 80% or more in developed countries.
More alarming is the double gap between "quantity" and "mass." In primary health-care institutions, the quality of vaginal lenses is uneven. A multi-centre study covering 23 provinces of the country shows that CIN2+ pathology detections at primary hospitals vary widely, ranging from 89 to 47 per cent, a huge gap in quality that directly leads to a large number of both out-of-patient and over-diagnosis cases.
2.2 "Ten-Year Cycle" dilemma for specialist training
Vagina lenses are not simply "photo-photo-scopy" but a specialist skill that requires a deep clinical background. A qualified vaginal lense needs to be trained in a system that lasts at least five to eight years: first, to master the basic theory of cervical anatomy, tissue pathology, vaginal imaging, then to perform at least 500 operations under the direction of a superior physician, to accumulate visual experience of different types of pathological change, and finally to practice independently through rigorous qualification.
The reality is that there are more than 100,000 officially trained vaginal gynthologists in the country. In the central and western districts, obstetricians and obstetricians are often examined as vaginal paralysers, and lack of systematic training leads to high rates of absenteeism and error.
2.3 Data isolation and the labeled "Invisible Trap"
In the dozens of medical AI projects that I have been involved in, I have found a widespread problem: most teams use data sets with a "triple-to-low" phenomenon — low sample numbers (more than 10,000 cases), low coverage of diseases (mostly single-grade pathologies), low modulation (only experimental images of acetic acid), low quality labelling (lack of a pathological pension comparison).
Specifically, the current vaginal mirror Ai data faces five major bottlenecks:
Bottleneck 1: The sample is seriously inadequate.Most open data sets have only thousands of cases, and there is a quantitative gap with actual clinical needs that does not support adequate training in in-depth learning models.
Bottleneck II: Lack of pathological standards.About 70 per cent of the labels are based on the visual judgement of vaginal mirrors, and lack of proof of the results of tissue pathology leads to a vicious circle of "mislearning."
Bottleneck III: Single-state information limitations.The use of only the acetic acid test images ignored the additional diagnostic information provided by the iodine test, while joint assessments of two-modular patterns in clinical practice were standard programmes.
Bottleneck four: The disease stove is rough.Most indicate only whether the pathology is present, and the exact classification cannot be supported by precise indications such as the pathology classification, range measurement, conversion area type, etc.
V: Data heterogeneity.Data vary widely between equipment, operators and populations, with poor cross-centre trans-fertilization of models and "good laboratory performance, clinical distress" becoming common.
It is these deep data bottlenecks that have led to a large number of cervical cancer AI products remaining at the "Demo" level, unable to actually reach the clinical line. That is why I attach so much importance to the construction of quality data sets.
Frontline breakthroughs: five technological waves of cervical-transformation A.
2025-2026 is the critical transition from "exploration" to "maturity" for the diagnosis of cervical-transformational diseases. Several breakthrough studies have been published in leading journals such as Lancet Digital Health, Nature Medicine, which offer exciting prospects for clinical application. I will read the technological advances at the most critical level of the day at five dimensions.
3.1 Multi-modular integration diagnosis: from "one-eyed" to "multi-dimensional"
Traditional vaginal mirrors AIS are based on monomodular images (aerosol trials) for pathological identification, while the main orientation in 2026 is the deep integration of multimodular information. The Cervix-MFNet model, published by the Stanford University School of Medicine in conjunction with Google Health, first integrated the cavity lens images (aeroquity+iodine tests double-modular), HPV spectrometry, TCT cytology, age multidimensional information for patients, etc., with AUC reaching 0.962, a sensitivity of 94.1%, and an excursive degree of 87.3%, which is 4.7 percentage points higher than the monomodular model.
At the national level, the Beijing Association and the hospital gynaecology and obstetric team based on the ColpoFusion model developed by the ENDO-COLPO data set of the Luang Hui technology, using Cross-Modal Action, to integrate the acetic acid with iodine experimental images in a characteristic level, CIN3+testing AUC to 0.971 on an independent external validation set, which exceeds the diagnosis of the median vaginal gynaecologist (AUC 0.91). This result was published in March 2026 in the Chinese Journal of Women’s Obstetrics, which is considered a milestone for multi-mode vaginal Ai in the country.
3.2 Basic model migration learning: from "Train from Zero" to "Knowledge migration"
In late 2025, Microsoft Institute released the MedCLIP-Colposcopy Basic Model, which provides a strong generic medical visual expression capability through pre-training in self-supervised comparison of 2 million medical images (with 150,000 vaginal mirrors). On this basis, only a small amount of data will be fine-tuned to achieve or exceed the full supervisory training in downstream missions.
Specific data show that, under the Few-Shot scenario, where only 1,000 cases are marked, the CIN2+ testing of AUC in MedCLIP-Colposcopy reaches 0.918, while the traditional Zero-trained ResNet-50 model is 0.795, with a gap of 12.3 percentage points. This means that the base model can significantly reduce reliance on the labeled data and significantly accelerate the adaptation of AI models to different centres and equipment.
3.3 Real-time video pathological classification: from static image to dynamic video
Early vaginal mirrors AA are mostly based on single-frame static images, and in clinical practice, vaginal lenses are a dynamic process - doctors synthesize their judgment by moving lenses, adjusting focal lengths, and observing time-series features of changes in acetic acid. In 2026, video-level AI diagnosis became the new technological high point.
The ColpoVid-Net, developed by the MIT Media Laboratory, uses a time-series-volume-repeated network (TCN)+Transformer hybrid structure that allows direct input of a complete vaginal lens examination video (average length of 8-12 minutes), automatic extraction of critical frames, tracking of dynamic changes in the diseaseal area, and output real-time rating results. In a validation set of 5,000 complete vaginal lens videos, ColpoVid-Net has 95.6% sensitivity to CIN2+ pathologies, 85.2% speciality, and single video processing time only 23 seconds, fully meeting clinical real-time requirements.
3.4 Self-supervised pre-training and semi-supervised learning: breaking the data label ceiling
The problem of identifying data scarcity is a constant problem for medical AI, and self-supervised learning (Self-Supervised Learning) provides new ideas for addressing this problem. In February 2026, DeepMind, published in Nature Medicine, developed a MoCo-Med model, which studied 1 million unmarked medical images by contrasting the amount of momentum, then fine-tuned on the vaginal mirrors mission, reaching 93% of the performance of the full-supervised model with only 5% of the labeled data.
Meanwhile, there has been significant progress in the semi-supervisory learning framework. The SemiColpo framework, proposed by the University’s Visual and Learning Laboratory in Jordan, uses the teacher-student model structure, combined with the harmonization and standardization and pseudo-labelling techniques, and with a 1:9 ratio of marked data to unlabelled data, the CIN2+ test of AUC has increased from 0.83 to 0.92, a relative 10.8% increase in the technological path to increase data use efficiency and reduce labelling costs.
3.5. Grassroots screening path optimization: from "aided diagnosis" to "path management"
The value of AI is not just "a substitute doctor's look" but, more importantly, to optimize the entire screening path and improve the efficiency of the health system. In 2025, the WHO Guide to Screening and Prevention of Cervical Cancer (update 2025) first incorporated AI-assisted screening into its recommendations, making it clear that in resource-limited areas, AA-assisted vaginal examination can be used as a diversion tool for first-positive women.
The multiple, random, cross-checking studies led by the Oncology Hospital of the Chinese Academy of Medical Sciences (incorporating 12 county hospitals and 38,000 women screened) showed that the AAAA-Auxiliary Vagina Spectrum Screening Path has increased the detection rate of CIN2+ from 0.89% to 1.32%, while reducing the rate of vaginal re-examination from 15.6% to 9.2%, significantly reducing unnecessary vaginal lens screening and saving medical resources. The results of the study were published in Lancet Global Health, in January 2026, providing a "China programme" for cervical cancer prevention and control in low- and middle-income countries around the world.
2026 Summary of technical indicators for diagnostic frontier A of cervical-uterine pathologies
| Technology Direction | Representative Models | Core indicators (CIN2+ testing) | Publication of periodicals/institutions |
|---|---|---|---|
| Multimodal Fusion | Cervix-MFNet | AUC 0.962, sensitivity 94.1% | Lancet Digital Health |
| Basic model migration | MedCLIP-Colposcopy | Few-Shot下AUC 0.918 | 微软研究院 |
| 视频级诊断 | ColpoVid-Net | Sensitivity 95.6%, 23 seconds/example | MIT Media Laboratory |
| Self-supervised pre-training | MoCo-Med | 5% of all supervised performance is indicated at 93% | Nature Medicine / DeepMind |
| Grassroots Path Optimization | AI-COLPO Path | CIN2+ detection rate increased by 48.3% | Lancet Global Health |
IV. Depth interpretation of the ENCO-COLPO data set: hard verification power of 180,000 bollocks standard data
After giving a detailed account of the advances in cutting-edge technology, I would like to return to the fundamental question: where do all these advanced algorithms come from? The answer is: high-quality, large-scale, and strictly qualitative medical data sets. And the EDO-COLPO video-image data set, created by Long Salange Information Technology, is one of the largest, most sophisticated, and most controlled vaginal data sets available at the national and global levels.
ENDO-COLPO data set core parameters
4.1 Data scale: qualitative change from "class 10,000" to "class 100,000"
180,000 vaginal video images – what does that mean? Let me take you into account: the largest internationally available vaginal lens data set is the National Cancer Institute (NCI) Colpo-DB, about 35,000; and the domestic public data set is more than 5,000-20,000 cases. Longway ENDO-COLPO is five times larger than the NCI data set, 10-30 times greater than the national public data set.
This quantitative advantage on a scale leads to a qualitative leap. The performance of the DILP is linked to the amount of training data, which, when it reaches a level of 100,000, is the real reason why the model learns the diversity of the pathologies – the manifestations of the pathologies – of different forms, locations, degrees, and population groups.
4.2 High-quality: Pathological gold standard is "lifeline".
In the medical AI area, I have been emphasizing one point:No gold standard label is the aeroplane.The most central advantage of the ENCO-COLPO data set is that each case has the corresponding organizational pathology test results against which the gold standard is compared.
In particular, Long Hui has set up a three-way matching mechanism for “vaginal mirror image-live-pathological diagnosis”: in every vaginal examination, the biopsy is precisely marked in the video, and the pathological diagnosis is matched by the image. This means that each disease-based stomatic note in the data set is not a “doctor-reading” test, but rather a “gold standard” for tissue pathology.
This is critical. Because even experienced vaginal gynthologists have a perusal of about 10-15% error. If AI models are trained on the basis of such "soft labels," their upper limit is locked to the level of human experts.
4.3 Double-modular data: a reduction of the clinical reality scene
In clinical practice, the standard process for vaginal examination is a two-modular joint assessment of the "Acerate test plus iodine test". The acid test shows a change in the upper skin of the diseaseal region (Bydrin) and the iodine test assesses the degree of decoupling of the diseaseal spectroscopy (Iodine in colorless areas).
However, the vast majority of the AIS data sets currently contain only experimental images of acetic acid, which is missing from the iodine test. This leads to a "less information" on the diagnosis logic of the AID model and the actual clinical presence of the ARI model, limiting its performance ceiling. The ENDO-COLP data set contains complete images of the two-modular acetic acid+iodine test, and the two-modular stomatology stories are closely aligned - the same disease's performance under the Aero acid and iodine tests is accurately linked, providing a solid data base for the development of multi-modular integration algorithms.
4.4 Quality control systems: Pentaple weight controls to ensure data reliability
Data quality control is the most easily neglected yet critical component of data set-up. Longwaytech has set up a "five-heavy" system to ensure the reliability of each case:
First heavy: Collecting quality control.Develop uniform vaginal image collection norms, including equipment parameters, operating processes, image quality requirements, etc., to ensure consistency of data at source.
Second, de-privileged quality control.Strictly enforce patient privacy protection norms, desensitize all images, remove personal identification information and ensure compliance with the Personal Information Protection Act and medical ethics requirements.
Third weight: Mark quality control.The two-blind note + arbitration mechanism is used, where each data is independently marked by two certified vaginal gynthologists, and the third senior senior physician is the arbitrator for inconsistent cases, ensuring that the accuracy of the label is over 98 per cent.
Fourth heavy: Pathological control.All pathological findings are independently viewed by two pathologists and difficult cases are referred to the in-house for discussion to ensure the accuracy of the gold standard.
Fifth heavy: statistical quality control.(c) Periodically perform overall quality assessments of data sets, including data distribution balance, labelling consistency tests, abnormality detection, etc., to ensure overall quality levels of data sets.
4.5 D-D: From "Is there a disease?" to "Accurate portrait."
The ENDO-COLPO label system is far beyond the simple "disease or no" classification, and instead has established a six-dimensional structured labelling programme that provides the AI model with a rich monitoring signal:
First dimension: disease classification- CIN1, CIN2, CIN3 levels (relative to lower-level scab and upper scab), supporting the AI model in the exact classification of pathological variations.
Second dimension: Description of the location and extent of the disease• Accurately marked the location of the cavity (intrauterine tube, conversion area, outer perimeter, etc.) and the measurement of the size of the cavity, supporting the AI model in the mapping and assessment of the extent of the disease.
3D: Description of transformation area type- The marking of the I, II, III transformation areas is essential for assessing the satisfaction and visibility of the vulva lens.
Fourth dimension: angiogeneity- Angiological characterizations such as spot-shaped, mosaiced, hexavascular, etc., which are important visual clues for determining the level of disease.
Fifth dimension: characterization of the white acoustic acid- Dynamic characterization of the thickness, boundary, depth of colour, time of appearance, time of retreat, etc. of acetic acid, supporting time-series characterization learning of video-level AI models.
6D: Clinical information links- Clinical information on the HPV spectrometry of each data related patient, TCT results, age, and maternal history, providing a complete clinical context for multi-modular integration studies.
Such a multi-dimensional, fine-coloping system allows the EDO-COLP data sets to support not only the "disease or no" sub-classification task, but also complex tasks such as disease classification, mapping, quality assessment, and provide adequate data support for functional expansion of the AI product.
V. Clinical application scenarios: data-enabled five locations
With high-quality data sets as the base, AI technology can really be valuable in clinical settings. Based on the capacity boundaries of the EDO-COLPO data sets, I think the five following clinical applications are the most relevant for the land and society.
Scenario 1: Auxiliary screening for cervical cancer - increase in detection rates at the grass-roots level
The AIS screening system, based on ENDO-COLPO training, allows for real-time analysis of vaginal mirror images, automatic identification of suspected disease areas and the introduction of classification, and supports the medical doctors at the grass-roots level in making more accurate judgements. The research data show that AIS can increase the sensitivity of primary physician CIN2+ from 72% to 89%, close to the level of specialist at the Sanctuary Hospital.
Scenario 2: Cervical ecology accurate classification - guiding clinical decision-making
The classification of cervical pathologies (CIN1/CIN2/CIN3) directly determines the clinical treatment programme - CIN1 uses more follow-up observations, and CIN2/3 requires active treatment (e.g. LEEP cones). Accurate classification is essential to avoid overtreatment or inadequate treatment.
Scenario III: Quality control of vaginal mirrors - specification of inspection operations
The quality of vaginal lens checks is influenced by the experience of operators, including the unevenness of the acetic acid smear, the adequacy of observation time, and the adequacy of the transformational area. The AQS can be based on a standardized examination process for learning the EDO-COLP data sets, real-time assessment of the quality of the check, and an indication to the operator to supplement the necessary operational steps (e.g. "Low time for accelerate smear" to suggest an iodine test). This is important for raising the standardisation level of grass-roots vaginal examination — the equivalent of having a "real-time quality control specialist" for every operator.
Scenario IV: Basic cervical cancer screening enabling - optimize screening pathways
In the "HPV First Positive-Campus Screening-Victoral Activity" screening path, vaginal lens screening is a key hub for connecting primary screening and diagnosis. A fully-enabled vaginal screening can improve the accuracy of diversion: unnecessary vaginal referrals can be reduced for women who are judged to be at low risk by AIS; and referral and confirmation processes can be accelerated for women who are judged to be at high risk by AIS. This AI-driven stratification management model allows a 30-40% reduction in vaginal referrals without lowering detection rates, and significantly increases the efficiency of the screening system.
Scenario 5: Decision support for referral - balanced medical resources
The AIT decision-making model based on ENDO-COLP data sets provides a comprehensive assessment of the multiple factors of the level, location, range, age of the patient and makes recommendations for referral. For patients with high suspected CIN3 or early immersion, the system automatically triggers rapid referral channels; for lower-level cases, it recommends follow-up management.
VI. Social benefits and industry values: from "data assets" to "health dividends"
6.1 Promoting equity in health care: Seizing quality resources
Health equity is the cornerstone of health for all. However, the concentration of quality health resources in large cities and hospitals is also prominent in the area of cervical cancer screening.
Imagine: in remote mountainous county hospitals, when a grass-roots physician operates a vaginal lens, the AI system can provide a level of diagnostic advice for specialists at the Tri-Acceal Hospital in real time – something that was previously inconceivable. Today, with the maturity of quality data sets and AI algorithms, this is becoming a reality.
6.2 Improved screening efficiency: release of health system capacity
The AIS system can take on the tasks of primary screening, diagnosis, quality control, and so forth, freeing doctors from repeated labour, focusing on more complex diagnostic and therapeutic decisions. It is estimated that the AIS support will reduce the average time for single vaginal examinations from 15 to 8 minutes, and will almost double the effectiveness of doctors’ work.
6.3 Medical savings: reducing socio-economic burdens
Early screening costs for cervical cancer are much lower than those for late treatment. Data show that treatment for early cervical cancer costs about $20-$30,000, while treatment for late cervical cancer costs between $300,000 and $500,000, with a survival rate of less than 30% over five years.
Preliminary estimates suggest that if AI-assisted screening increases the national cervical cancer screening coverage from 55% to 75%, it will reduce the incidence of cervical cancer by approximately 23,000 cases per year, saving more than $10 billion in medical expenses. At the same time, the precision diversion of AI-assisted screening will also reduce the need for unnecessary vaginal examination by about 30%, further saving medical resources.
6.4 Promotion of industrial upgrading: building medical AI data ecology
The construction of the ENDO-COLPO data set, which supports not only the development of its own AAI products, but also the provision of high-quality data base for the entire industry. Through models such as data authorization, joint R & D, more AAI firms, scientific institutions, and medical institutions can use this data base as a basis for innovative research, accelerating the transformation of technological outcomes, and creating a "data-calculations-product-clinical" virtuous circle that will drive the entire AAI industry to higher levels.
VII. Expert vision: five major trends in cervical-transformation AI over the next 3-5 years
At the time of 2026, looking to the next three to five years, I have five judgements on the trend towards cervical change in AI.
7.1 Trends I: evolution from "one-task AI" to "all-process AI"
The current vaginal mirror AI focuses on a single mission (e.g., pathology testing), while the future AI system will cover the entire process of screening for cervical cancer — reading from HPV/TCT results, vaginal para-diagnosis, referral to active screening, pathological aids, follow-up management recommendations, and the formation of an end-to-end AIS-assisted decision-making system. This requires data sets to include not only vaginal lens images, but also integration of multi-modular clinical data such as HPV, TCT, pathology, etc., to build more complete “patient portraits”.
7.2 Trends II: Basic models become industry-standardized
As the medical foundation model matures, the future vaginal mirror AI model will shift from "from zero training" to "basic model + fine-tuning." This means that the demand for large-scale unmarked data will increase significantly, and the value of high-quality markers will become more prominent – because the quality of the labeling at the fine-tuning stage directly determines the model's performance ceiling for specialized missions. Enterprises with large-scale gold standard label data will take a strategic high point in the base model age.
7.3 Trends III: Video-level AI mainstream
The static image analysis is evolving to dynamic video analysis. The future vaginal mirror AI system will directly handle the full inspection video, automatically extract key frames, track pathological changes, generate structured reports, and even assess the quality of the operator's inspection. This will require a higher level of video size and the precision of the label.
7.4 Trends IV: AI participation in top-level design of screening pathways
The role of AI will be upgraded from "aid tools" to "path designers." By analysing large-scale real-world data, AI can optimize screening strategies for different regions and populations – for example, by strengthening initial screening in high-growth HPV areas, optimizing referral thresholds in resource-limited areas, and achieving precision control of "locally appropriate" cancer prevention and control.
7.5 Trends V: Data compliance and standardization become industry consensus
With the introduction of regulations such as the Personal Information Protection Act, the Management of Cybersecurity in Health Institutions, the Medical Data Code, the use of which will be increasingly regulated. In the future, a good data governance system, through rigorous ethical scrutiny, and data sets that will have the required data compliance, will have real commercial value and social trust.
VIII. CONCLUSION: LIGHTING THE WAY TO CURRENT CRAZY WITH STATISTICS
In 2020, the World Health Organization launched the Global Strategy for Accelerated Eradication of Cervical Cancer, which sets the target of "90-70-90": 90 per cent of girls complete HPV vaccinations by age 15, 70 per cent of women receive high-quality screening by age 35-45 and 90 per cent of confirmed patients receive standard treatment.
To achieve this ambitious goal, vaccines are the first line of defence, screening the second line of defence, and treatment the third line of defence. And AI technology is making the second line of defence stronger, more efficient, and more equitable.
As a gynaecologist, I fought cervical cancer for most of my career. I have seen too many tragedies of losing my life because I found it too late, and I have seen the pleasure of being healthy again because of timely screening. I know that every early disease that is accurately identified is behind a family's integrity and happiness.
I believe that, with the continued accumulation of high-quality medical data, the continuous improvement of AI algorithms, and the growing policy environment, the day of the elimination of cervical cancer will come sooner than we had expected. And on this road to "zero cervical cancer" every data scientist, every clinical practitioner, every company dedicated to medical AI is an indispensable force.
Let us build on data, and on AI as a beacon to safeguard the healthy future of women around the globe.
About Changsha Langhui Information Technology Co., Ltd.
Chang Shalom Information Technology, a leading medical image AI data set provider and provider of smart diagnostic solutions in the country, focuses on the construction of high-quality medical image data sets in the fields of endoscopes, skin mirrors, and AI algorithms. The company’s core team, composed of senior medical experts, data scientists, and AI engineers, has built a multidisciplinary medical image data set matrix covering digestives, gynaecology, dermatology, etc., with a total of over 1.5 million cases, serving dozens of hospitals, scientific institutions, and AI companies in the country.