"For more than 30 years, I have examined more than a million chest-chips. The chest-chips are the most basic and common of all video-testing, but also the one that most test doctors' power: from the point of the lung to the angle of the horn, and from the cleavage to the pleural, each detail may contain clues of disease. And the Ai-era chest-chip diagnosis is being redefined by a large-scale high-quality data set."
Expert guide: Chests - The "Spring Tree" of Medical Images
As a radiologist, my career begins with a chest-chip. After more than 30 years, advanced video technology has developed, such as CT, MRI, PET-CT, but the chest-chip is still the most widely used means of screening images in clinical terms. No image-checking can be as cheap and easy as a chest-chip, and can be screened and followed up, almost for all people.
The number of chest X-rays in the country is estimated to be over 3 billion per year, or over 40 per cent of all video screenings. From medical screening to the first emergency care, from hospitalization routine to post-operative follow-up, the chest is everywhere. It is the first line of defence for the diagnosis of diseases such as respiratory diseases, cardiovascular diseases, breast tumours, etc., and is an important tool in public health services.
However, the quality of a chest tablet is mixed. A chest table contains very large amounts of information — lung field, lung door, interspersion, heart mitectomy, pectrometer, muscular muscle, ribs, thorax... Any anomaly in any part of the chest can be a signal of disease. A full and accurate reading of a chest tablet requires a doctor with solid anatomy, pathology and clinical knowledge, as well as extensive exposure to the film.
In reality, the long cycle of training and the low success rate of radiologists and the high demand for chests have led to a serious imbalance in supply and demand. In particular, at the primary level, the capacity to diagnose chests is generally weak – many of the primary level are read by general practitioners or low-age doctors, with a high rate of error in the diagnosis.
Today, 2026, artificial intelligence is revolutionizing the diagnosis of the chest. From single-moderation detection to multi-pathological diagnostics, from image classification to report generation, from 2D images to cross-modular learning, the development of the chest-capsule is changing.
Today, I would like to explore with you, from the perspective of a radiologist, the cutting edge of chest X-ray AI and the industry value and strategic significance of 2 million XR-CHEST chest X-ray data sets of Chang Long-Hang-Hong-Hong-Hong-Hong-Hong-Hyun Information Technology.
II. Trade pain: The "Quadretex" of the chest tablet diagnosis
Difficulty one: a huge workload, a heavy burden on the doctors
The radiology department of a Sanctuary Hospital may be handling thousands of chest-chip reports every day. Radiologists sit in the reading room every day, staring at a screen with a single page reading, with dry eyes and scapric acid.
More importantly, the chest plate, though "common" but not "simple." The chest anatomic structure is complex, with a wide variety of variations and varying manifestations. It is also a lung shadow, which may be pneumonia, which may be tuberculosis, which may be tumors, which may be old-fashioned. To make an accurate diagnosis, it requires a combination of factors such as the region, form, density, edge, and tissue change, and clinical information about the patient.
It is difficult for doctors to devote enough time and effort to each chestpiece under the pressure of a heavy workload. According to foreign studies, radiologists spend less than 30 seconds on average on a chest tablet.
A study shows that after eight hours of work, radiologists reduce the diagnosis’s accuracy rate by 10-15%. In our country, many radiologists work far more than eight hours, and the problem of the quality of diagnosis is compounded by fatigue.
Distress II: Weak capacity at the grass-roots level and poor quality of diagnosis
If the problem with the big hospital is "not busy," then the problem with the basic hospital is "not having enough capacity."
The radiology departments of our primary health care institutions generally face a shortage of personnel, equipment and capacity. Many county hospitals have only one or two radiologists, and most of them have not been systematically trained.
According to the Survey of the Status of Medical Image Services at Basic Level (2025), the diagnosis of chests in primary health care institutions is only 65-75% accurate, and the rate of leakage is as high as 15-25%. In particular, early diagnosis of lung cancer, with the rate of failure at the grass-roots level exceeding 30%, means that many patients have taken a chest film at the basic level, but have not been detected in time, and it is the middle and late stage of the diagnosis when the higher hospital is in the upper part.
The lack of capacity for the diagnosis of primary chests not only affects the health of patients, but also limits the advancement of the level of treatment - patients do not trust the results of the diagnosis at the grass-roots level, with major diseases running to the major hospitals, which cause overcrowding and the basic hospitals to become more accessible.
The third problem: the disease is so numerous, AI is "departure warfare."
The variety of chest diseases is hundreds of them. Images of different diseases vary, and diagnostic criteria vary. This leads to the development of a "battle-to-action" situation for chest-caps AI, with companies doing pneumonia tests, some nostrils, some tuberculosis tests and others aerobic breast.
This "single disease" AI development model has several problems: first, limited clinical value — doctors need comprehensive diagnosis rather than isolated results; second, high deployment costs — hospitals need multiple AI systems to integrate and maintain them; and third, low data utilization — each AI uses only a partial label of data, which causes waste of data resources.
In recent years, although there have been several multi-pathic chest-top-button AI products, performance has generally been unsatisfactory. The underlying reason is data — training an AI system for multi-pathological combination diagnosis requires a large-scale data set that simultaneously labels multiple pathologies, which are scarce.
Frustration IV: quality deviation, "Gold standard" not.
The training of the chest-top AI requires extensive data labelling. However, the quality of the label is uneven and is a widespread problem in the industry.
The difficulty of the chest tablets is, first, that the boundaries of the disease are blurred - that of many lung pathologies (such as inflammation, immersion, fibrosis, etc.) are unclear and that the profiles of different doctors may vary considerably; secondly, that differences in subjective judgement - that of different doctors are inconsistent in the judgement of certain signs (such as "pulmonary texture increase" "pleural increase" etc.); and thirdly, that diagnosis needs to be combined with clinical — that the clinical significance of some images needs to be combined with the patient's history, symptoms, laboratory examination, etc., and that it is difficult to determine simply to see images.
For these reasons, the gold standard, which is marked on the chest, is not often "gold." Many data sets are marked by a doctor, and quality is difficult to guarantee without review and quality control. The AI model that is trained with this data is a performance ceiling that is understandable.
The chest-button AI has gone beyond the "no" phase and is moving towards the "yes" phase. The next stage of the competition is the core of the competition for data -- who has the largest, highest quality, highest diseased chest-film data sets -- who can take the upper end of the basic model age. Longwaytech has 2 million cases of XR-CHEST data sets, the " data building block" of the age.
Three, 2026 front-line breakthrough: "Five technological waves" of chest image AI.
In 2025-2026, the chest X-ray AI sector saw a new wave of technology. The basic model, multi-pathological joint diagnosis, report generation, cross-model learning, etc., are moving in parallel, pushing the chest-film AI into a new phase of development.
Breaking one: the basic model of mass pre-training -- the birth of Universal Chest-Ai.
Since 2025, Foundation Model has become the hottest direction for medical video AI. In the area of chest-chips, the basic model of mass pre-training is changing the development model of "one-on-one" and generating "one-on-one" "general-chip AI".
In March 2025, Google DeepMind published a Med-PaLM M model in Nature, which was a generic model of polymode medicine, which performed impressively on the chest-capsure mission – with the 14-type chest pathology test, the average AUD of Med-PaLM reached 0.93, exceeding most previous models dedicated to chest-chip training.
But what is more striking is its "common capability" -- the same model, which does not require individual training for each disease to identify multiple breast pathologies. This is thanks to the "creasing capacity" that comes from mass pre-training -- when model size and training data reach a certain level, models suddenly acquire the capabilities that many small models do not.
In October 2025, good news came from the country. The first basic chest-chip model, ChestFM, was released by UNI, in collaboration with the University of Jordan’s Sun Hospital, based on 800,000 chest-chip data, was self-supervised and then fine-tuned on several downstream missions. In 14 disease detection missions, ChestFM averaged 0.92 AUC, comparable to international advanced levels.
In 2026, the chest-capture base model moved in a further direction towards "larger, stronger." As I understand it, Longwaytech is working with the country's top AI team to train a larger chest-image base model, ChestFM-200M, based on its 2 million chest-capture data sets. Initial results show that ChestFM-200M performed significantly more in the first generation of basic models than in many downstream missions, especially in rare pathologies and complex cases.
Break two: multi-disease joint diagnosis -- from "one-point test" to "full evaluation"
Clinicians read the chest tablets, which are a comprehensive, systematic process – from the point to the bottom of the lungs, from the cleavage to the plethora, from the bone to the soft tissue, to be seen and judged in a single way.
In 2025-2026, multi-disease joint diagnosis became a major development for the chest tablet AI. In June 2025, a study published by the Stanford University School of Medicine in Radiology produced a ChestX-Dense model that can diagnose 37 breast diseases simultaneously, reaching an average of 0.90 for AUC. Multi-disease joint models are not only more comprehensive than single-disease models, but also have higher diagnostic accuracy for certain diseases because of the sharing of characteristics between the different diseases.
In 2026, new breakthroughs were made in the joint diagnosis of multiple diseases. Beijing's team of radiology, based on 150,000 cases of chest tablets from Longhui technology, developed the multi-mission multi-pathic chest-in-a-ambranch system ChestMultiNet.
More than anything else, the team introduced the "disease Link Model" in the model, which uses complication and diagnostic relationships between different diseases to enhance overall diagnostic performance. For example, the model "knows" that chest cavity is often associated with pneumonia or tumors, that tuberculosis is better than the upper end of the lung. These a priori studies are introduced, making AI's diagnosis more like a human doctor's thinking.
Breaking Three: AI reports generation -- from "seeing the film" to "writing the report."
The job of a radiologist is not just "seeing a film" but, more importantly, "writing a report" – to describe the images in a professional, standardized language and give diagnostics and recommendations.
Since 2025, as the technology of the Big Language Model (LLM) has evolved at a rapid pace, AI automatically generates video reports as a hotspot for research. In May 2025, Microsoft Institute, in collaboration with John Hopkins University, published a GPT-4-based chest report generation system in Nature Medicine. The system combines image characterization with large language models, which automatically generate structured chest-sheet reports. In the radiologist's blind review, more than 60 per cent of AI-generated reports are rated "clinically available", 30 per cent of which are rated as "of a level comparable to that of a senior physician".
The Chinese-language chest-captioning system ChestReportGPT, based on a large-scale data set by Longsier Hospital in Shanghai, has been developed. Using the "Victor Code" plus the Chinese medical mega-linguistic model, the system automatically produces a report on the chest that complies with Chinese radiology norms.
Preliminary clinical tests show that the reports generated by ChestReportGPT have reached 88% of the accuracy of the key findings, 92% of the completeness of the reports, and highly rated by radiologists. More importantly, the system has significantly improved efficiency by reducing the time that doctors write their reports by more than 60%.
Of course, the AI report generation is still in a supporting stage – the first draft of AI is being produced, and doctors review it. But I believe that, as technology continues to improve, the quality of the AI report will become ever-increasing and will eventually become an indispensable “writing assistant” for radiologists.
Breakthrough Four: Cross-modular Learning - X-ray + CT Knowledge Migration
The chest is a 2D projection, with limited information. The CT is a 3D fault image that provides more rich and sophisticated anatomical information. If the knowledge that CT has learned is moved to the chest, it will greatly enhance the performance of the chest-button AI. This is the idea of cross-model learning.
In 2025, this led to a major breakthrough. The research team at the MIT General Hospital presented the XCT-Net model, a cross-modular learning framework that uses a large amount of CT data to complement the training of the chest table.
Experimental results show that the performance of CT pre-trained chest tablet models, for the purpose of pulmonary dysslexia detection, pneumonia recognition, etc., is significantly better than that of training only with chest tablet data (AUC increased by 5.2 and 4.8 percentage points, respectively). This demonstrates the effectiveness of the transmodular knowledge transfer - the more informative information of CT can better identify the pathology in the Church chest tablet model.
In 2026, cross-modular learning moved further. The team at Munich Industrial University proposed a cross-modular production model – using CT data to "enhanced" chest tablet data, which would generate synthetic chests of different perspectives and different degrees of variability for the expansion of training data sets.
Longway Technologies is also actively mapping cross-modular learning. It is understood that they are building "X-ray plus CT" pairing data sets to provide data support for cross-modular chest-ray AI research. This direction, if broken, will open up new spaces for the performance of chest-rays.
Break five: Pathological gold standard certification -- from "image diagnosis" to "pathology control."
For a long time, the training and validation of the chest-top AI has been based on the diagnosis of video doctors as the gold standard. But the video diagnosis is ultimately indirect, and there is a gap between pathological diagnosis and pathological diagnosis.
Since 2025, more and more research has focused on "image-pathological contrast" — the validation and optimization of the chest-top-top-top model using pathological diagnostic results. In September 2025, the Sloan-Käitlin Cancer Centre (MSKCC) published a study in Lancet Digital Health, where they trained and validated the chest-top-top-top-cot-top model using pathologically confirmed lung cancer cases, which showed a 12 per cent increase in the specifics of lung cancer detection and a significant decrease in the positive rate of false cancer.
The point of this study is that it points to a new direction for the development of the chest-top AI – moving from the "image doctor level" to the "pathological gold standard." While images can never be exactly equivalent to pathology, the goal of the chest-top AI is to move as closely as possible towards the pathology standard.
In 2026, the Longway data set also made a major breakthrough in this regard. They collated tens of thousands of pathologically contrasted chest-cardiogram data covering a wide range of diseases, including lung cancer, tuberculosis and pneumonia. These pathological gold data can be used not only to verify the performance of the AI model, but also to train more accurate AI models — so that AI learned not how to diagnose a doctor, but what a disease really is.
IV. Depth interpretation of the Longway data set: 2 million cases of Encyclopedia chestboard data repository
And when you get to the technology frontier, we look at the data base that underpins these technologies. The XR-CHEST chest X-ray data set of Chang Longway Information Technologies is one of the largest, best marked and most quality breast-chip data sets in the country. In my view, it has six core advantages.
Advantage one: 2 million sizes — large data-driven base model
The Rang Hui XR-CHEST data set contains 2 million chest X-ray images, a size that is the leading data set for the country’s chest. What is the concept of 2 million cases? It corresponds to the total number of chest-film examinations in 10-15 years at a large Sanctuary Hospital, or the total number of breast-film examinations in a medium-sized city for one year.
Why is it so big? Because we're entering the "basic model age." The core logic of the underlying model is that "large-scale pre-training plus downstream fine-tuning" and the effect of pre-training is positively related to the size of the data -- the more data the model learns, the more common features, the better the performance of downstream missions.
Two million cases of data provide enough "fuel" for training high-quality chest image base models. I can say responsibly that, domestically, there are few data sets that support training in a truly basic chest-chip model, and that the XR-CHEST of Longway is the best of them.
Apart from the size of the population, the distribution of the Luang Hui data sets is very reasonable. It covers people of different ages, gender, and regions, and covers cases from normal to the spectral system of diseases. This balanced distribution ensures that models are not biased and can be applied to a wide range of people.
Advantage two: 14 common pathological signs - full coverage of clinical needs
The Luang Hui data set ' s labelling system is designed around core clinical needs and covers the 14 most common types of chest disease, including:
- Lung pathology (7 groups):Pneumonia / edema, pneumonia, tuberculosis, emphysema/slow resistance, pulmonary fibrosis, bronchial extension, lung insufficiency
- Cerural change (2 classes):Quite a lot of plethora, gas pecs.
- Intersect with heart (class 2):I'm a little bit more than I am.
- Other (3 categories):Ccalcified stoves, pleural increase/adhesive, bone abnormality
These 14 types of pathologies cover over 90% of positive detections in chest examinations and basically cover the most common breast diseases in daily clinical work. The AI system, based on this data set, allows for a thorough and systematic assessment of the chest tablets, rather than focusing on a disease.
Even more difficult, the labels are not simple "no" categories, but rather contain a wealth of details. For example, pulmonary knots indicate location, size, density (reality/brush glass/mix), marginality characteristics, etc., pneumonia indicates range (large leaf/silentity/intersistence), severity, etc., and nodules indicate type (original hair/relay/blood spread), activity, etc. These detailed labels provide a rich monitoring signal for a well-trained AI model.
Strength three: Multi-centre Multi-Equipment - Data Diversity and Broadening
The broad capability of the AI model depends to a large extent on the diversity of training data. Another important advantage of the Longfei data set is its multicentre, multi-equipment source.
Image equipment covers 100 models of mainstream brands, including the main brands of the country and abroad, such as the GE, Siemens, Philips, Ui-tung, Wandong, and East Soft.
This broad data source ensures the "heterogenicity" of the data set - different image quality, different projection conditions, different equipment parameters, different population characteristics ... The AI model is trained in such a data set to learn true and fine features rather than overcompatible data for a hospital, a device.
I've seen too many AI products, doing excellent work on their own data, and "controversial" at other hospitals, the underlying reason being that training data are too single and lacking in diversity. And Long Hui's multi-centre data solves this problem from the source. That's why the AI model based on Long Hui's data training, which tends to be more stable in cross-centre validation.
Strength four: Pathogen standard — highest level data validation
One of the most differential advantages of the Long Cai data sets is that it contains tens of thousands of cases of chest tablets with a pathological value. These cases have clear pathological diagnosis, covering a wide range of diseases, including lung cancer, tuberculosis, pneumococcal paraplasia, and mesotheliosis.
Why is the pathological gold standard so important? Because video diagnosis is "pregnant," and pathological diagnosis is "confirmed." Using video doctors' diagnosis as a training label, AI learned only "how do doctors see it," and using pathological diagnosis as a label, AI learned "what disease really is."
Of course, standard pathological gold data are very difficult to obtain – not all patients have pathological tests, but there is a time and space gap between pathology and images. Long Hui can collect tens of thousands of pathologically contrasted chest-cardiogram data, which is not easy.
These are standard pathological gold data, which are used to assess the true performance of AI as a model validation "final criterion" and to train as a "high precision label" to improve the diagnostic accuracy of models for key diseases such as tumours. It can be said that standard pathological gold data are the "middle beads on the crown" of the Longway data set.
Strength five: Level three quality control system -- the moat of data quality
The data quality is the lifeline of AI. Long Hui has established a well-established three-tier quality control system for data quality control, which has created a solid moat for data quality.
Level 1 quality control: image quality screening.All the chest cards entered are subject to rigorous quality screening, including image clarity, body compliance (whether or not they are standard positive tablets), exposure dose testing, etc. Unqualified images are automatically excluded to ensure that each image entered meets the diagnostic quality requirements. The Lai Hue data set has a visual quality rating of over 98.5%.
Secondary quality control: Double-marked + expert review.If there is a disagreement, the third senior physician will be referred to arbitration.
Level 3 quality control: regular sampling and dynamic monitoring.Long Hui has established a well-established quality control mechanism to regularly sample the marked data and assess the stability of the label quality. If a doctor is found to be of a reduced quality, retrained in a timely manner.
This strict quality control system is an important guarantee of the quality of Long Wai’s data. My research centre, which compared Long Hui’s data to our own experts, has been more than 92% consistent. This is a very high level of commercial data concentration.
Advantage six: Structured reporting and clinical information - value of multi-dimensional data
The Long Cai data sets contain not only images and labels, but also a wealth of clinical information and structured reporting data. These multi-dimensional information greatly increases the value of the data sets.
Specifically, the data set contains clinical information on patients ' basic information (age, sex, geography), clinical diagnosis, laboratory results (blood protocol, biochemicals, etc.), other video screenings (CT, MRI, etc.), pathological diagnosis, treatment programmes, follow-up information, etc.
The value of these multidimensional information lies in, first, their role as additional monitoring signals to help train more accurate AI models; secondly, their data base for multimodular learning (image + clinical text + laboratory data); and thirdly, their wealth of information for clinical research, which allows for research on image and clinical correlation analysis, prognosis, etc.
In addition, Luang Hui provides structured radiology reporting data. These data are important for training AI report generation models, which include not only descriptions of visual findings, but also diagnostic observations and recommendations, and are the best AAI teaching material for learning "how to write a report".
V. Clinical application scenario: five major scenarios for the re-shaping of the AI chest tablet
The AI system based on the XR-CHEST data set training is playing an important role in the many scenes of chest-clinic therapy. I will present the most representative applications.
Health check-ups: smart reading, efficient screening
The health checkup is a mandatory programme for the screening of the chest. The number of checkups in the country is more than 1 billion per year. In the traditional way, the examination chest is read by radiologists, with a huge amount of work, and since the examination chest is mostly normal, doctors are prone to "vision fatigue" and some early pathologies are missing.
Emergency and outpatient care: assisted diagnosis, reduction of absenteeism
In emergency and outpatient care, the chest is one of the most commonly used examinations. A doctor needs to determine quickly and accurately whether the patient has an acute condition, such as pneumonia, aerobics, and chest cavity fluids, to justify treatment decisions.
Primary health care empowerment: giving grass-roots diagnostic capacity at the "3A" level
The AIS system, based on the Longway data set, can be deployed at the local level, including county hospitals, community health centres, and help grass-roots doctors to improve their diagnosis. AI can not only give results of the disease detection, but also provide diagnostic reference and treatment advice, which is equivalent to an "AI expert consultant" for primary doctors.
Tuberculosis control: helping to end tuberculosis epidemic
Tuberculosis is one of the most important infectious diseases in our country. Annual screening for tuberculosis requires a large number of chests, and inadequate reading capacity is an important constraint on the efficiency and quality of screening. The TBAI testing system, based on the Longway data set, allows for the rapid and accurate identification of tuberculosis stoves in the chest, providing technical support for mass screening.
Early screening for lung cancer: early detection, life-saving
The chest tablet is one of the most important tools for screening for lung cancer, although low doses of CT are the current mainstream screening method, but the chest tablets still play an important role in screening and routine medical examinations at the grass-roots level, because of their ease and economic nature. The AIS system for detection of lung cancer, which is based on the Longway data set (especially cases of lung cancer with pathological criteria for pulmonary money), allows for detection of early lung cancer in the chest tablets, such as suspicious nosy, swelling, and pulmonary failure, and prompts patients to undergo further CT examinations.
VI. Social benefits and industry values: data-enabled "inclusive health care"
Improving medical equity - making quality video diagnostics accessible
Health equity is an important objective of the Healthy China strategy. However, there is a huge disparity in the level of diagnosis of images in different regions and at different levels of health care.
The widespread use of AI technology is changing the situation. The 2 million-square-quality data training-based chest-chip AI system can be deployed to any location with an X-ray machine – either a Sanchai hospital in a front-line city or a rural health centre in a remote mountainous area – and can be used to provide the same level of AI-assisted diagnostic services.
The National Health Board estimates that if the chest-filing AI can cover 80% of district hospitals and community health centres throughout the country, it can reduce cross-regional access by about 200 million times a year, saving patients families more than 50 billion yuan in transport, accommodation, work-wrecks, etc. More importantly, tens of millions of patients will be diagnosed in a more timely and accurate manner, and the disease forecast will improve significantly.
Improving the quality of diagnosis - reducing the number of errors and ensuring medical safety
The quality of care and safety is a permanent theme of hospital management.
AI, as the second eye, can effectively reduce the incidence of error in diagnosis and improve the quality of diagnosis. AI is not tired, emotionally unaffected, and is not negligent, and provides high-quality diagnostic aids that are continuously and steadily provided.
From a hospital management perspective, reducing the rate of error in the diagnosis of a medical failure means reducing medical disputes, raising patient satisfaction and enhancing hospital reputation. From a social point of view, more patients are diagnosed in a timely and accurate manner, meaning that the disease is treated early, better in advance, and the burden of disease is reduced in society as a whole.
Release the doctor's value - from the read machine to the clinician.
Radiologists often make themselves laugh at the "reading machine" – sitting in front of computers every day, reading films and writing reports, and repetitive work takes on most of the time. This pattern of work not only exhausts doctors but also limits their professional value.
AI is being introduced to change the way radiologists work. AI is responsible for such repetitive work as preliminary screening, measurement, and preparation of initial reports, while doctors devote more time to analysis of difficult cases, communication with clinical units, patient counselling, etc., which are more professional.
The introduction of a chest-chip AI led to an increase in the daily chest-chip reporting of doctors from 150 to 250, but the sense of fatigue has decreased. Meanwhile, doctors’ time spent on difficult case discussions and clinical consultations has increased by 40%, and the overall academic level and capacity of the clinics has been improved.
Promoting industrial development - building the medical AI innovation ecology
High-quality large-scale data sets, not only as the basis of AI products, but also as the infrastructure for the entire medical AI industry. The opening and application of the XR-CHEST data sets in Longway will drive the development of the whole chest image AI industry.
First, it lowers the threshold for AI R & D - no start-up firms and research institutions need to start data collection from zero, and can be developed directly on the basis of data from Long Hui, and significantly reduces the research and development cycle and costs.
Second, it promotes technological innovation and industrial collaboration – different firms and research institutions can conduct research and competition on the basis of the same data, and promote rapid technological advances. At the same time, data firms, AI firms, medical equipment firms, and health-care institutions can forge closer collaborations to create innovative ecosystems.
Finally, it provides a reference for regulation and standardization - a high-quality indicator data set that can serve as a "baseline test set" for the review and approval of the AI product, and provides data support for the establishment of industry standards and regulatory policy development.
"2 million cases of high-quality chest-chip data are the "data building block" of the chest image AI. It supports not only an AI product, but also the innovative development of the entire industry and the health and well-being of millions of patients."
VII. Expert vision: trends in the next 3-5 years of chest-chip AI
As a radiologist who has witnessed the evolution of medical images from the film age to digital and intellectual, I look forward to the next three to five years of chest-button AI. Here are some of my judgments.
Trends one: Foundation model is a frame -- "One model to take care of all tasks."
Over the next three to five years, the base model will be the formula for the chest-chip AI. We will no longer need to train individual models for each disease, each mission, but rather to adapt to different downstream tasks with a large-scale pre-training base model.
This will bring about several fundamental changes: first, AI's capacity boundaries have been significantly expanded — the same model has been used to perform multiple tasks, such as detection, classification, fragmentation and report generation; secondly, small sample learning has been significantly enhanced — the "knowledge reserve" with the underlying model, which can achieve good performance in new missions with only a small amount of data to be marked; and thirdly, continuous learning and iteratives have become easier — the base model can continuously absorb new data and knowledge and evolve.
I predict that by 2028, mainstream chest-top AI products will be based on basic model structures, and universal chest-top AI will become industry standard. Whoever has the most large and high-quality training data will have a dominant position in the competition for basic models.
Trends II: From "Assisted Diagnostics" to "Clinical Decision Support"
The current chest AI is still largely at the level of "aided diagnosis" -- telling the doctor where there is." In the future, the chest AI will upgrade the direction of "clinical decision support" -- not only telling the doctor what it means, but also telling the doctor what it means to do.
Specifically, AI will provide more comprehensive diagnostic advice, diagnostic analysis, recommendations for further examination, reference to treatment programs, etc., taking into account the multidimensional data of patients’ clinical information, medical history, laboratory examinations, etc.
This will be achieved by the deep integration of image data with clinical data. The Longway data set is well-grounded in this area – not only with images, but also with rich clinical information and structured reporting data.
Trends III: Multimodular integration - integration of images, text, genes, genomics
Future medical AI must be multimodular - images, text, genes, metabolic groups, protein groups ... data on all mosaics will be integrated to provide more comprehensive information on disease diagnosis, treatment and prevention.
The value of multimodular integration is particularly evident in the area of chest diseases. For example, early diagnosis of lung cancer requires a combination of images, tumor markers, history of smoking, family history, etc.; stratification of meso-transgenic pulmonary disease requires a combination of images, lung function, pathology, and self-antibodies; and pathological diagnosis of pneumonia requires a combination of images, clinical symptoms, laboratory examinations, etc.
Multimodular AI will be an important development for the next three to five years. And the rich clinical information and multidimensional data set of Longway data provide a valuable data base for multimodular research.
Trends IV: From "Hospitals" to "Outside Hospitals" - AI-enabled health management for the whole cycle
The current chest-chip AI is used mainly for diagnostics in hospitals. In the future, the AI application will extend beyond the hospital to cover the full cycle of health management, such as screening for diseases, health management, home surveillance, and slow disease follow-up.
For example, in screening for high-risk groups for lung cancer, AI can automatically analyse the annual examination of the chest tablets, track changes in the nodal sections and automatically alert the high-risk population for further examination; in the management of chronic respiratory diseases, AI can monitor the progress of the disease through regular chest-cardimetric analyses and provide a basis for the adjustment of treatment programmes; in community health management, AI can assist family doctors in the screening of the chest and health assessment of contracted residents.
This whole cycle of health management will fundamentally change the pattern of health services – from “disease to health management” to “treatment to prevention.” AI, in turn, will be the central technical underpinning for this transformation.
Trends V: China's chest-button AI to the world
China has the world’s largest medical data resources and the richest clinical scene. The large-scale chest-film data accumulated by Chinese companies, represented by Long Hui, provide unique conditions for the development of China’s chest-film AI.
I am confident that China’s chest-button AI technology will reach international leadership in the next three to five years. Moreover, China’s chest-button AI products will be able to go out of the country to serve countries and regions along the “across the road” and contribute to the cause of global health by contributing Chinese wisdom and Chinese programmes.
At the same time, the AAI standard for chest diseases, based on China’s large-scale data and clinical practice, is expected to be internationally recognized as a standard approach recommended in the international guidelines. This will be a historic leap for Chinese medical AI from “following” to “run” to “leading.”
VIII. CONCLUSION: Data are base, smart-winged, and a healthy breast is built together
Likewise, the more than 30 years of my career as a radiologist have taught me the value of a chest tablet – simple, cheap, and everywhere – but carrying the burden of safeguarding the health of millions of people. From film to number, from manual reading to AI-aided, the technology of chest-card diagnostics is improving, but its mission of safeguarding health remains unchanged.
Today, we are standing at a new point in history. New technologies, such as basic models, multimodulation learning, cross-modular migration, are pushing the chest-top AI to a new level.
Rang Huitech’s 2 million cases of XR-CHEST chest X-ray image data sets are not only a commercial product, but also a core infrastructure for our chest image AI industry. It provides sufficient fuel for AI’s development, provides powerful tools for grass-roots empowerment, provides valuable resources for clinical research, and provides a solid force for healthy China.
I often say to young doctors, "The eyes of a radiologist are clinical. We are blind, and we are directly concerned with the patient's treatment." "Today, I would like to say to companies like Longhui, whose data quality determines the eyes of AI, and the health and well-being of millions of patients. This responsibility is as great as Tarzan.
Let us work together, using data as a basis, and smart as a wing, to build a strong breast defence that protects people ' s health and to create a better future for medically artificial intelligence.