From GP mapping to bio-knowledge orientation: 2026 technology leapfrogging to a million-scale data base from Bone Age AI assessment

The age assessment is the core diagnostic link in the monitoring of the child's endocrines and growth and development. Using a left wrist X-ray assessment of the maturity of the child, doctors can assess differences in the child's biological age and actual age, thus diagnosing the growth hormone deficiency, pre-pregnancy, thyroid abnormalities, and predicting adult heights.

In 2026, with the maturity of the in-depth learning technology and the construction of large-scale medical video data sets, the bone age assessment is undergoing a paradigm shift from manual to automated. With experience in medical data services, Changshang Sha Longei Information Technology has built up a total of 1.7 million bone-age x-ray image data sets, covering a multi-age, multi-sex continuum, which is an important data base for training in the first-year model of children’s bone age in China.

In the field of sports, the bone age is used for age verification and grouping of young athletes. In forensic science, the skeletal age assessment is an important reference for judging the age of criminal responsibility of juvenile suspects. These diverse applications drive the cross-fertilization of the skeletal age AI technology to higher precision and broader capacity.

The global medical image AI market is expected to reach $8 billion by 2026, with the skeletal muscle system AI accounting for about 12 per cent of the total, and the skeletal age assessment as its break-down track is at a critical stage of transition from research validation to clinical deployment.

In international terms, the RSNA-Skeletal Age Challenge in the US has been driving the field since 2017, and the award-winning team's MAE (average absolute error) has been reduced to less than four months. As a commercial AI-based software, BoneView BoneAge in Europe completed a conceptual validation study in 2025, incorporating 203 paediatric patients (average age 9.0), systematically comparing it with the GP mapping and TW2 methods.

In China, the development of Bone Age AI is uniquely challenged. Children in China have different patterns of bone age development than the European-American population, and direct use of models based on European-American training can lead to systemic deviations. In 2025, the ABAA-GP model, based on RSN 14,236 left-hand DR tablets, assessed its clinical applicability among Chinese children and youth, confirming the importance of the suitability of the population.

In terms of market-drivenness, China’s child endocrine clinics have grown by 8-10% annually, and the need for bone age assessment has continued to rise. Meanwhile, the NHSC’s policy of “adult health management” requires early screening and early treatment of diseases such as dwarfia, and early sexual debauchery, further expanding the application of B-Ai.

In 2025-2026, the Bones-Ai field made a series of breakthroughs that marked the evolution of technology from "pure data driven" to "knowledge+ data driven".

The first is the maturity of the TW3 deep learning model. The retrospective clinical trial published in 2025 in Diagnostics, training in the deep nervous network based on the Tanner-Whitehouse 3 method, using 560 Korean children's wrist DR tablets, systematically compared with paediatric radiologists. The model demonstrated good accuracy and consistency in TW3 skeletal projections, confirming that in-depth learning could replace some of the manual TW3 assessments.

The second is the breakthrough of light quantification models. In 2025, the improved EfficientNetB3 structure, while maintaining an accuracy rate of 81.5%, reduced the amount of parameters to only 15.8 million, significantly increasing the efficiency of reasoning using R. Adam Optimizers and combination loss functions.

The most notable is the framework "Bio-Knowledge-led Double Neural Network" published in 2026 on Frontiers in Radiology. The study proposes to integrate anatomy knowledge and skeletal development patterns into an in-depth learning model, which is driven by a combination of knowledge+data through a two-neurological network structure. This approach not only enhances the accuracy of predictions, but also, more importantly, the interpretability of models – which is essential to medical AI’s regulatory approval and clinical trust.

The design of the data set by Longhuitech takes full account of these needs for cutting-edge research. 1.7 million data are available on both the GP and TW3 compatible labels, each containing not only bone age values, but also the sequence, size and form of emergence of the bone centres, providing a structured label system for training the "knowledge guide" model. In addition, the data set links growth curves, endocrine testing and follow-up data, supporting joint multimodular learning.

The construction of the Luang Hui technology bone age X-ray image data set strictly follows the full chain standards for medical image data acquisition, labelling and quality control.

In terms of data collection, all images are derived from the cooperative radiology and child endocrines of San Armour Hospital, using standardized positive left wrist projection protocols covering the far end of the finger, palm, wrist and sept. Data retain raw DICOM formats and exposure parameters, pixel spacing, position/view information to ensure traceability of image quality. 1.7 million cases are distributed continuously by sex and age, covering the entire age range from infants and adolescents, and include normal age reference, early bone age and full spectrum distribution of late bone age.

In terms of labelling systems, two standard compatible labels are used for each case: the Greulich-Pyle spectrograph and the Tanner-Whitehouse 3 score values. The markers include structured fields such as developmental stratification, morphology, integration status, etc. of the boner centres.

In terms of quality control, 10 per cent of random samples are marked independently by three physicians, calculating the Fleiss' Kappa coefficient, which is less than 0.75 dimensions of marking as a need for re-engineering. In addition, the data set establishes an automated quality assessment conduit, which quantitatively assesses image clarity, physical accuracy, exposure suitability and removes unsatisfactory images.

In terms of compliance, all images are legally authorized from a source, PHI fields are fully dissensitized, do not contain any information that directly identifies individuals, and meet the requirements of the Personal Information Protection Act and the Data Security Act.

Looking ahead, the development of B-Ai will show three main trends.

First, joint multi-modular assessments will be mainstreamed. The skeletal-age assessment should not rely only on x-ray images, but also combine clinical information (heat, weight, growth speed), endocrinology testing (growth hormones, IGF-1, sex hormones) and genetic data to construct joint multi-modular assessment models.

Second, AI will be a necessary condition for regulatory approval. The 2026 Bio-Livers in Radiology framework for a dual neural network suggests this direction – the future Bone-Ai not only gives numbers but also explains why.

Third, adaptation of Chinese populations will become core competitiveness. There is a systemic bias in the performance of Chinese children based on the training of European and American populations, which creates unique advantages for companies with indigenous Chinese data.

In business models, the value of BWA is not only higher diagnostic efficiency, but also longer-term value added for data assets. As data accumulate and models continue to evolve, data sets will become the core competitiveness of companies.

The basic rule of "data determine the height of AI" is vividly interpreted in the development of B-AAI. From 14,000 images at the RSNA Challenge to 1.7 million data at Longway Technologies, from a single GPS standard to a double standard compatible, from pure data-driven to knowledge-led, each step behind a breakthrough is underpinned by a data infrastructure.

Long Shae-hye Information Technology Ltd., which is based on 1.7 million cases of bone-age X-ray image data sets, is becoming a major data service provider in the field of Child-Aged AI. In 2026, this year of turning the Ai-based medical image, Long Hui-tech will continue to cultivate the field of medical data to promote the landing of precision medicine with high-quality data.

In terms of global competition patterns, the Bone Age AI track is in the critical window period of transition from technical validation to a scaled clinical deployment. International giants such as Vuno (Republic of Korea), BoneView (France), have taken the lead in productization and obtaining regulatory approval at home or in the region, while China’s market has not yet emerged with a national impact.

LANG Huitai believes that the core barrier to BGAI lies not in the algorithm architecture – the increasingly obvious effects of technological convergence in the field of deep learning – but in data quality and size. The size advantage of 1.7 million cases means greater panoramic capacity, lower risk of bias and wider age coverage.