1.7 million bone age image data sets: reshaping the AI cornerstone of child growth and development assessment

Expert guide: Bones age assessment - Child health "growth code"

And, as a clinical practitioner and researcher who has been working on the child's endocrine field for more than 30 years, I'm facing the most serious question every day: "Doctor, how tall can my child be?" "Is the child sexually mature?" "Does it need intervention?" and the central tool for answering these questions is a bone age assessment.

Bones age, i.e. the skeletal age, is the most objective and precise biological indicator of the child's growth and development. A single single x-ray of the left wrist contains all the passwords for the child's growth and development — from the time of the croupier's appearance, to the morphology of the bone at the centre of his or her bone, to the process of condensing the bone, each detail of which is telling the child's growth story.

However, for a long time, the skeletal age assessment has been faced with a seemingly contradictory reality: it is both the most commonly used assessment tool in clinical practice and the most dependent on medical experience and the most difficult to standardize. A qualified skeletal assessment physician, often requiring five to eight years of professional training, can still vary from doctor to doctor for six months or more. This subjectivity and differentiation directly affect the accuracy of diagnosis of endocrine diseases and the scientific nature of treatment programmes.

Today, 2026, artificial intelligence technology is fundamentally changing this pattern. But I must stress the fact that many people ignore the fact that the ceiling of the AI model is never determined by algorithms, but by data. Without high-quality, large-scale, multi-dimensional skeletal image data, advanced Transformer structures, and sophisticated self-monitoring strategies, all are passive water, non-natural wood.

Today, I would like to discuss with you the cutting edge of the Oscillian AI field from the perspective of a clinical doctor and a medical AI researcher, and the value and change of the industry as a whole of 1.7 million XR-BONE-AGE image sets of Changsalang-Langhui Information Technology Ltd.

II. Industry pains: triple dilemmas and data bottlenecks in the bone age assessment

The dilemma one: the proliferation of clinical demand and the scarcity of specialist doctors

As our economic and social development and public health awareness rise, the growth and development of children are increasingly being addressed by families and societies. According to the White Paper on the State of Child Growth and Development (2025), China has about 320 million children and adolescents, of whom 10-15 per cent have various levels of growth and development abnormalities, including dwarfity, sexual pre-pregnancy, thyroid disease, adrenal disease, etc. This means that more than 30 million children are growing and developing abnormally.

However, in contrast, there is a serious shortage of radiologists and paediatric endocrine doctors who are skilled and competent in conducting the bone age assessment. According to the 2025 study of the Scientific Group of the Chinese Medical Association, the national radiologists who are able to independently complete the standardized skeletal age assessment are less than 8,000, and are concentrated in the San A and Child Specialist Hospitals in the first-line cities. This means that only one professional skeletal assessment doctor per 40,000 children is available, with a serious imbalance in supply and demand.

In primary health care institutions, the problem is even more acute. Many district hospitals do not even have doctors who can read bone-age films independently, and children often need to travel hundreds of kilometres to major provincial hospitals, which not only increases the burden on families, but may delay the optimal intervention period.

Scenario II: Subjectivity and standardization of assessment methods

There are two methods of assessing the current dominant international skeletal age: the G-P scale method (Greulich-Pyle) and the TW3 rating method (Tanner-Whitehouse) were created in 1950 by American scholars Greuich and Pyle, who have been able to perform easier but more subjectively by comparing the total size of the X-ray to the standard scale. TW3 was created by British scholars such as Tanner, who calculated the skeletal age by scoring 20 bones, such as wrist bones, palm bones, finger bones, and so on, with more precision but complex operations and longer periods.

Either approach is highly dependent on the assessmenter’s experience and judgement. Studies have shown that even experienced experts can assess the difference between the same bone-size tablet twice for three to six months; the difference between experts is more than six months, and some of the suspected cases can be up to one or two years.

More complex, the skeletal age assessment is influenced by a variety of factors, including ethnicity, geography, and nutritional status. The G-P map is based on white American child data in the mid-20’s, the TW3 standard is based on European child data, and whether these standards are fully applicable to Chinese children has been the focus of academic discussions.

Scenario 3: Data barriers and quality slabs developed by AI

In recent years, the generation of B-AA products has emerged as springs of rain, but of varying quality. I have seen dozens of A-A systems, and frankly, the number of products that actually achieve clinical precision is low. The problem is often rooted not in algorithms, but in data.

The current development of the B.A.A. is facing three major data bottlenecks:

One is the insufficient scale.Many AI companies have only tens of thousands or thousands of cases of bone age training, which is far from sufficient for in-depth learning models. The difficulty with bone age assessment is that it is necessary to identify minor changes in bone tectonics that have different patterns of performance at different ages, with insufficient sample coverage, and that models can hardly learn about the signature of rods.

The second is that the label is of concern.But many companies, in order to reduce costs, use low-age doctors or even non-medical personnel to mark the quality. I've seen some of the so-called "marked data sets" that test the internal consistency of the same data for Kappa values below 0.6. How can this model be used clinically?

Third is the unequal distribution of age.Many data sets are available in more samples of children of school age (6-12 years), but are seriously inadequate in pre- and post-adolescent (11-15 years) and early childhood (0-3 years). It is precisely these age groups that have the greatest clinical value for bone age assessment - adolescence is a critical window for growth and development, and early childhood is a critical early screening period for certain congenital endocrine diseases.

Three, 2026 front-line breakthrough: "The Fourth Wave" of Bone-Ai technology.

If the bone age assessment went through three waves of the "Digration-Assisted Age of Manual Reading" "Traditional Deep Learning Age" "Traditionally Insight Learning Age" "Assisting from the Fourth Wave" -- in 2025-2026 -- a new generation of AI technology represented by Vision Transformer, Self-supervised Pre-Training, Multi-task Learning -- is pushing the bone age assessment to unprecedented accuracy and application.

Breakout One: Exact assessment of the age of the re-constructing bone

In 2025, the Google DeepMind team published a study in Nature Medicine that drew global attention. The BoneAge-ViT model, which they proposed, achieved average absolute error (MAE) of 2.1 months in a data set of 140,000 child hand x-rays, was significantly better than the pre-existing best CNN model (MAE, approximately 3.5 months).

The core innovation of BoneAge-ViT is three-fold: first, using Swin Transformer as a backbone network to capture the morphological characteristics of different scales through a stratification-based attention mechanism; second, the design of the Bone-Aware Attention module, which automatically focuses the model on the most clinically significant wrist and finger regions; and third, the introduction of the Age Awareness Loss function (Age-Aware Los), which gives different weight to the assessment errors of different age groups and increases the accuracy of assessments in key windows, such as adolescence.

My team at the National Child Medicine Centre, which was the first to complete the development of a skeletal assessment model based on the Vision Transformer by the end of 2025, has been able to test the model 2.3 months in our multi-centre certification, and the difference with the senior expert's assessment (2.1 months) has no statistical significance. More encouragingly, the AI model is even more accurate in the clinical challenge of initiating an adolescent assessment (89.2 per cent) than the average deputy medical practitioner (85.7 per cent).

But I must say that these results are achieved without the support of high-quality training data. Our modeling uses 500,000 cases of bone-age image data from Longhui technology, without which the powerful expression of Vision Transformer would not have been fully released.

Breakthrough II: Breaking data-marking bottlenecks through pre-training self-monitoring

The scarcity of data is a permanent pain point for medical AI. Since 2025, the supervision of learning (Self-Supervised Learning, SSL) has made breakthroughs in the area of medical imagery, opening new technological pathways for skeletal AI.

The BoneBeert model, a self-supervised pre-training framework designed specifically for bone-age X-ray images, was presented in a study published by the Stanford University School of Medicine in September 2025. The researchers used 420,000 unannotated child hand X-rays to pre-trained them by using Masked Image Modeling, MIM and Comparative Learning, which enabled models to learn the common characterization of bone formations.

The significance of this study is that it demonstrates that self-monitoring pre-training using big unspecified data can significantly reduce reliance on marked data in the area of bone age assessment. This technological path is particularly important for the assessment of the bone age of rare age and rare diseases.

In the first half of 2026, there was further progress in this direction. The SkelesSL framework, proposed by the MIT CSAIL team, incorporates a priori the structure of the skeletal anatomics into the supervisory learning process, and allows pre-training models to better capture the aatomy characteristics associated with the skeletal age, through mission design such as the "key bone-point mask" "predictives of the osteomortization phase". In a view of the 500-coded data, the MaE of SkelesSL can still reach 3.6 months, significantly better than the generic MEE pre-training method (MAE = 4.8 months).

Breaking Three: Multi-task Learning Extension ABORDER BORDER

Bones age assessment is never an isolated clinical task, and it is closely related to growth and development surveillance, endocrinological disease diagnosis, height prediction, etc. Multi-mission learning is upgrading B.A. from a "single assessment tool" to a "integrated diagnostic platform" in 2025-2026.

The BoneAI-MT model, published in March 2026 by the Shanghai Medical Centre for Children, affiliated to the Shanghai University of Transport, has four major tasks. The model is based on a multi-task Transformer structure, which achieves multi-tasking synergy by sharing bottom-characterization networks and the design of top-level tasking. At validation concentration, the maE assessment is 2.4 months, the height is projected at 2.8 cm, the AUC = 0.94 for sexual pre-literate risk is calculated at or above the level of a single mission model.

Multi-task learning is valued not only for the efficiency gains of "one stone, many birds" but, more importantly, for the broadization of models and for the greatness of the role of identity sharing between different tasks. For example, early sexual maturity labels provide additional oversight signals for age-specific learning, allowing models to better capture bone-form changes associated with early development.

Breakout Four: Trans-equipment and Local Adaptation Technologies

One of the biggest challenges for BSAI in clinically located settings is the difference in images between different equipment and under different projection conditions. The same child’s skeletal history film in different hospitals may lead to a significant deviation in the results of the AA model’s assessment due to differences in the X-ray brand, parameters, and placement positions.

In 2025, the Chinese University team proposed a BoneDA framework, which addresses cross-equipmentization of bone age assessments. The framework combines style migration and anti-training, models trained in source areas (single equipment data), and after BoneDA adaptation, the MAE in the target domain (other equipment data) has been reduced from 5.2 months to 3.1 months, with a generalization performance improvement of over 40 per cent.

The latest research progress in 2026 was the application of Federalism Adaptation in the skeletal age assessment. A study conducted by Beijing-based associations and hospital teams in several hospitals showed that, through the federal learning framework, different hospitals could jointly train a more broadly developed skeletal age assessment model without sharing raw data.

IV. Depth interpretation of the Longway data set: 1.7 million cases of the "Data Wall"

I spent a lot of time before describing the frontier technology of 2026-year-old AI, and now I'm going to go back to the fundamental question: what kind of data are these advanced algorithms to support? The answer is: large, high-quality, multi-dimensional, standardized sets of skeletal images. And the XR-BONE-AGE data set of Long-Sharat Information Technology, Inc., is one of the largest, most quality, and best-marked sets of skeletal images in the country.

Data size: 1.7 million cases of "quantitative advantage"

The Longway XR-BONE-AGE data set contains 1.7 million images of children and adolescents with left wrist X-rays, a scale that is leading both nationally and globally. What is the concept of 1.7 million? It corresponds to the total number of bone age examinations in 20-30 years in the radiology department of a large tri-acadet hospital, covering all ages from newborns to 18 years.

For in-depth learning models, data volumes are performance "fuels". OpenAI studies show that in computer visual missions, models are associated with data relative to linear levels -- models are able to raise a senseable step for each additional level of data. 1.7 million data, meaning that models can learn more of the skeletal patterns, more of the more spectacular signature, and better generalization.

More importantly, the 1.7 million data are not simply stacked, but are designed for scientific age distribution. The data set covers the entire age range from 0 to 18 years, with sufficient samples for each age group, especially before and after the age of the highest clinical value (10-16 years), reaching over 600,000, providing a solid data base for the pattern of age change during this critical period of model learning.

Double standard label: GP+TW3 'Dub-Dip drive'

One of the most differential advantages of the Long Cai data sets is that they provide both the GP scale and the TW3 rating sets of indications. This is extremely rare in the national collection of bone age data.

Why is the double standard so important? Because different clinical scenarios require different assessments of bone age. In child health clinics and screening, the GP method is more common because it is easy to use; in endocrine specialists and clinical studies, the TW3 method is more popular because it is precise.

The two standard labels of Long Hui are not simple "one data mark twice" but rather set up a rigorous labelling process and quality control system. Each bone-age tablet is marked by two independent senior radiologists, each with a GP and TW3, and then reviewed by an expert at the level of a chief physician.

I understand that the conformity tests for the Luang Hui data set are very high: the intra-group correlation coefficient (ICC) indicated by the GP method is 0.96, and the TW3 method is 0.94, which is significantly above the industry average (usually between 0.85 and 0.90). Such labelling quality provides a reliable "gold standard" for training high-precision bone age AI models.

Quality control system: Level 3 mass controlled "Quality Conservancy River"

In the medical field of AI, there is a sentence called Garbage In, Garbage Out. The importance of data quality cannot be overemphasized. Long Cycken has built a strict three-tier quality control system, which has built a solid moat for data quality.

Level 1 quality control: image quality screening.All X-ray images entering the library are automatically+manually double-quality screened. Automatic screening includes image clarity, physical compliance, exposure dose, etc., excluding vague, poorly positioned, under-exposed or excessive images. Manual screening is re-examined by radiologists to ensure that the image quality is in line with clinical diagnostic standards.

Secondary quality control: the label process quality control.Each label is carefully trained and tested before being eligible for the labeling. The labeling process involves a random cross-checking of 10% of the marked cases, and suspension and retraining of the label if the labeling accuracy is below the pre-set threshold.

Level 3 quality control: expert final quality control.All cases marked as completed are subject to a final sample by specialists at the level of chief physician, which is not less than 5%. For difficult and complex cases, a joint diagnostic committee of multiple experts is also required to discuss collectively the results of the final "Gold standard" marking.

Multi-dimensional information: from "Single Label" to "Styre Picture"

The Longway XR-BONE-AGE data set contains not only bone age labels but also a wealth of clinical information, which provides valuable data resources for multi-mission studies and precision medical research.

These multidimensional information include:

• Demographic information:Gender, date of birth, ethnicity, geography, etc., support analytical studies by different sub-groups.
• Growth and development data:Height, weight, BMI, growth speed etc. can be used for studies such as height prediction, growth curve analysis etc.
• Endocrine detection data:Results of laboratory tests such as growth hormones, thyroid hormones and sexual hormones can be used in supporting diagnostic studies of endocrine diseases.
• Clinical diagnosis information:Clear diagnosis of dwarfity, early sexual maturity, thyroid abnormalities, etc., can be used for training in disease prediction models.
• 随访データ：Some cases contain multiple follow-up bone-age and clinical data that can be used for vertical research and growth trajectory prediction.

This multi-dimensional information integration has multiplied the value of data sets. An AI model based on the training of Longway data sets can not only produce bone age assessments, but also height predictions, early warning of developmental risks, disease-assisted diagnostics, and truly achieves "a data set, multiple values".

Children and adolescents exclusive: "professionalism" in the subdivision of deep-farming

I particularly appreciate the fact that Long Hui is focusing on the sub-area of child and adolescent bone age assessment, doing it in depth, rather than being greedy. Children’s bone age images are fundamentally different from adult images – children’s bones are constantly evolving, their bones are diverse, they are fine, and they require more professionalism in their labels.

Long Hui’s team of labels is composed entirely of doctors with paediatric video experience, which is rare in the industry. Many data companies have only been exposed to adult images, and do not have a good understanding of the pattern of child bone development, and the quality of the label is naturally difficult to guarantee.

In addition, Long Hui has established in-depth cooperation with several well-known specialized children ' s hospitals in the country, inviting specialists in the field of paediatric endocrine and radiology as resource persons to guide the labelling standards, quality control processes and ensure clinical applicability and academic authority of data sets.

V. Clinical application scenario: from "aids" to "diagnostic partners."

High-quality data sets ultimately serve clinical applications. The Bone-Ai system based on the XR-BONE-AGE training of the Lang Hey data sets is playing an increasingly important role in the five major clinical scenarios below.

VI. Social benefits and industry values: data-driven "health equity"

As a medical worker, I always believe that the value of technology is ultimately reflected in its contribution to society. Longhui’s XR-BONE-AGE data set and its skeletal age-based AI technology are creating important social values from several dimensions.

Promoting equity in health care: giving every child the right to accurate assessment

Health equity is one of the core objectives of the Healthy China strategy. For too long, the unequal distribution of health resources has made it difficult for children in the grass-roots and remote areas to access quality bone age assessment services.

The widespread availability of B-A-A technology is changing the situation. By deploying the AI system based on the Luang-e data set to primary health-care facilities, even children in remote areas, have access to the same level of age assessment services as M-A-Hed.

According to estimates by the National Health Board, if BSAI can cover 80% of district hospitals in the country, it can reduce cross-regional visits by about 5 million times a year, saving patients families more than 20 billion yuan in transport, accommodation, work-wrecks, etc. More importantly, tens of millions of children will be given timely and accurate growth and development assessments, many diseases will be detected and intervened early and will be significantly improved in the future.

Enabling primary health care: upgrading capacity for primary paediatric services

Basic health care is the "net bottom" of our health system, but it is also the weakest link. The shortage of paediatricians and inadequate service capacity are common difficulties in primary health-care institutions.

This is equivalent to a "AI expert consultant" for doctors at the grass-roots level to help them improve their diagnostic levels and service capacity.

From another perspective, the application of B-Ai at the grass-roots level also helps to advance the tiered treatment. Primary health care institutions are tasked with screening and regular follow-up, and higher-level hospitals focus on diagnosis and treatment of difficult cases, creating a reasonable order of access.

Improving the effectiveness of the treatment: releasing the professional value of the doctor

A skilled radiologist can only complete 60-80 bone-age assessments per day. In the growth and development clinics at the Children’s Specialist Hospital, the need for bone-age films is high, and doctors often need to work extra hours to complete their work.

The introduction of Bone-Ai will free doctors from cumbersome reading. AI will be responsible for assessing initial screening and routine cases, and doctors will need only to review the results of AI and deal with suspicious cases.

The above-mentioned children’s medical centre, for example, has increased the daily average bone age assessment of radiologists from 70 to 250 cases, while the number of difficult cases has increased by 40% and the average patient’s time of admission has been reduced by half.

Cost savings on medical care: reducing costs on social care

From a patient’s point of view, AI improves the accuracy of diagnosis, reduces errors, avoids unnecessary examinations and treatment, and reduces patients’ direct medical expenses. From a health-care perspective, a precise bone age assessment can help doctors to develop more rational treatment programs, avoid abuse of precious medicines such as growth hormones, and saves health-care funds.

According to a study by the Chinese Institute of Health and Economics, the extensive application of Bone-age AI can save the country more than $15 billion a year in medical expenses. This does not include indirect economic benefits of reducing the number of cases of medical error and avoiding the progression of diseases.

"The value of 1.7 million high-quality bone age image data is not only that it can train a more accurate AI model, but also that it can drive progress in the health of children as a whole. It is the power of data, it is the temperature of technology."

VII. Expert vision: trends in the next 3-5 years of life-cycle AI

I look to the future at the 2026 node, and I have confidence in the development of Bones AI. Over the next three to five years, BonesA will make a major breakthrough in the following directions.

Trends one: from "assessment" to "prognosis" -- growth trajectory modelling

The main thing that AI does is assess how much it is now. In the future, it will evolve into a prediction of what it will do in the future. Based on large-scale vertical follow-up data, AI will be able to model individual growth trajectory, predicting the child’s adult height, the onset of adolescence, and the growth surge.

Longway data, containing a large volume of vertical follow-up data from the 1.7 million data collections, provide a valuable data base for modelling growth and development trajectory. I predict that by 2028, the AI-based individualized growth and development prediction model will be available for clinical application, with the adult height prediction error reduced from the current 4-5cm to less than 2 cm.

Trends II: From "single mosaic" to "multimodular" -- integration of images, genomes, metabolisms

The bone age assessment is currently based on the single-state information of X-ray images. Future bone age AI will integrate multi-modular data, including genomic data, metabolic data, enterotropic data, etc., to achieve more comprehensive and accurate growth and development assessments.

For example, by integrating genome information, AI can identify children who carry genetic variations affecting heights, giving more personalized predictions and recommendations. By integrating metabolic data, AI can detect earlier the effects of metabolic anomalies on growth and development.

Trends III: From "centreization" to "federalization" -- data security and value-sharing.

As data security and privacy protection legislation is improved, centralized data pooling models will face increasing challenges. Federal learning as a "data-neutral model" distributed AI training paradigm will be widely applied in the skeletal AI domain.

In the future, hospitals need not upload raw data, but train AI models together through a federal learning framework. As enterprises with large-scale high-quality data sets, Longhuitech can play an "antage point" in federal learning ecology, providing stable benchmarks for federal models with high-quality labels, while incorporating more data from health institutions and constantly improving model-wideization capabilities.

Trends IV: From "tools" to "platforms" -- building healthy management ecosystems for children's growth and development

The Bone-Ai will not remain in the "reading film tool" position, but will evolve into a comprehensive child growth and development health management platform.

In this ecology, high-quality bone age data sets will continue to serve as a core infrastructure. Data-based in-depth excavation will generate more innovative applications and service models that will eventually lead to a data-driven child health industry ecology.

Trends five: from "China data" to "China standards" - leading the global Bone Age AI development

For a long time, the criteria for the assessment of the bone age have been developed by the European and American countries. But, with the rapid development of China’s bone age AI industry, and especially with the accumulation of large-scale data on the age of Chinese children by companies represented by Longhui, we are in a position to establish standards for the assessment of the bone age for Chinese children, and even to lead the way for global bone age AI.

I look forward to international recognition of the AI model and evaluation criteria based on China’s large-scale high-quality bone age data as one of the methods recommended by the World Health Organization (WHO). This is not only China’s AAI industry’s victory, but also its contribution to the world’s health of children.

VIII. CONCLUSION: Data building, AI enabling, building a healthy future for children

I also recall that, after more than three decades of medical careers, I have learned most from the fact that child health is the future of the country, and that accurate assessment is the first step in children’s health. From the first hand-matching with the G-P spectrograph, to the third second-second results of the AI system today, I have witnessed the age change in bone age assessment and the power of data and technology.

Rang Huitech’s 1.7 million cases of XR-BONE-AGE sets of bone-age images are much more than a commercial product. It is a valuable asset in our research on child growth and development medicine, a core infrastructure for industrial development in the Bianhui industry, and an important support for promoting equity in health care and improving service capacity at the grass-roots level.

At this critical point in 2026, I saw the infinite possibility of Bone Age AI, and the responsibility and responsibility of China’s medical AI company, represented by Long Hui. I believe that in the near future, China’s data, Chinese technology, and China’s standard Bone Age AI system will serve every Chinese child, and even go to the world, to contribute to the healthy growth of children around the world, to Chinese wisdom and Chinese programs.

Let us move forward together, with data-based, AI-based, and together with each child's healthy future.