China Urges Standardized AI Medical Imaging Databases to Unlock Data Potential
In the rapidly evolving landscape of artificial intelligence (AI) in healthcare, medical imaging stands out as one of the most promising domains for real-world application. With the exponential growth of clinical imaging data across Chinese hospitals, driven by advancements in digital infrastructure and diagnostic technologies, the potential for AI-driven breakthroughs in radiology, oncology, and precision medicine has never been greater. Yet, despite this data abundance, a critical bottleneck persists: the lack of standardized, high-quality, and ethically governed databases necessary to train, validate, and deploy robust AI models.
A recent commentary published in the Medical Journal of Peking Union Medical College Hospital underscores the urgency of establishing a national framework for standardized medical imaging databases in China. Authored by Zhenwei Shi and Zaiyi Liu from the Department of Radiology at Guangdong Provincial People’s Hospital, the article calls for a coordinated national effort to transform raw imaging data into scientifically managed, reusable, and interoperable resources that can fuel the next generation of AI innovation in medicine.
The central argument is clear: while China produces vast amounts of medical imaging data annually—estimated to grow at a rate of 30% per year—the majority of this data remains siloed, inconsistently annotated, and inaccessible for large-scale research. This disconnect between data volume and data utility severely limits the development and generalizability of AI models, particularly those intended for clinical deployment. As the authors point out, many Chinese researchers still rely heavily on international public datasets such as The Cancer Imaging Archive (TCIA) and The Cancer Genome Atlas (TCGA), which, while valuable, may not fully represent the epidemiological, genetic, and phenotypic diversity of the Chinese population.
The reliance on foreign datasets introduces several challenges. First, there are inherent differences in patient demographics, disease prevalence, imaging protocols, and equipment vendors, all of which can affect the performance of AI algorithms when applied locally. Second, the absence of a unified domestic data infrastructure hampers reproducibility and collaboration among Chinese institutions. Third, and perhaps most critically, the lack of standardized metadata and annotation practices makes it difficult to compare model performance across studies or to integrate AI tools into clinical workflows with confidence.
To address these issues, Shi and Liu advocate for the adoption of the FAIR data principles—Findable, Accessible, Interoperable, and Reusable—as a foundational framework for building a next-generation medical imaging data ecosystem in China. Introduced in 2016 by the FORCE11 scholarly community, the FAIR principles were designed to enhance the value of scientific data by ensuring that it can be effectively discovered, accessed, shared, and reused by both humans and machines.
In the context of medical imaging, FAIR compliance means more than just storing DICOM files in a central repository. It requires a systematic approach to data curation that includes standardized metadata schemas, controlled vocabularies, semantic annotations, and transparent data governance policies. For example, a FAIR-compliant imaging dataset would not only include the raw scans but also detailed information about the imaging protocol, patient demographics (with appropriate privacy safeguards), radiological findings, pathology reports, genomic data (where available), and expert annotations—all structured in a machine-readable format.
One of the key innovations highlighted in the article is the use of DICOM-SEGMENTATION files, which allow for the structured representation of image annotations such as tumor contours or organ segmentations. By encoding these annotations in a standardized format, researchers can ensure that AI models trained on one dataset can be meaningfully compared with those trained on another, even if they come from different institutions or regions. This level of interoperability is essential for multi-center validation studies and for the eventual regulatory approval of AI-based diagnostic tools.
The authors emphasize that achieving FAIR compliance is not merely a technical challenge but also a socio-organizational one. It requires alignment across multiple stakeholders, including clinicians, data scientists, hospital administrators, ethics committees, and policymakers. In China, where healthcare data is subject to strict privacy regulations and where institutional data sharing has historically been limited, building trust and establishing clear guidelines for data ownership, access rights, and usage permissions will be crucial.
To this end, the commentary references China’s 2018 Scientific Data Management Measures, a policy initiative aimed at improving data stewardship and promoting open science. While this policy provides a legal and administrative foundation for data sharing, its implementation in the medical domain has been uneven. Many researchers remain hesitant to contribute their data due to concerns about intellectual property, patient confidentiality, and the lack of recognition for data curation efforts.
Shi and Liu argue that incentives must be created to encourage participation in data sharing initiatives. This could include formal credit for data contributors in publications, integration of data management into academic evaluation systems, and the establishment of national data commons with tiered access levels based on user credentials and research purpose. They also call for the development of a Chinese-specific medical ontology—a standardized vocabulary for describing diseases, imaging findings, and clinical procedures—that can serve as the semantic backbone of future databases.
Such an ontology would go beyond simple keyword tagging. It would enable advanced querying capabilities, allowing researchers to search for all cases of, say, “adenocarcinoma of the lung with ground-glass opacity on CT” across multiple hospitals, even if the original reports used slightly different terminology. This level of semantic interoperability is essential for powering AI models that rely on large, diverse, and accurately labeled datasets.
The commentary also highlights the importance of technical infrastructure. Building a national AI-ready imaging database will require significant investment in computing resources, cloud storage, and specialized software for data anonymization, quality control, and annotation management. The authors note that real-time data transformation pipelines will be needed to convert raw clinical data into FAIR-compliant formats without placing an undue burden on frontline medical staff.
One of the most compelling aspects of their vision is the emphasis on sustainability. Rather than treating data collection as a one-off project, they propose a continuous, dynamic system in which new data is routinely ingested, curated, and made available to the research community. This would create a feedback loop in which AI models are continuously refined and validated against fresh, real-world data, leading to improved accuracy and clinical relevance over time.
The implications of such a system extend far beyond academic research. A standardized, FAIR-compliant imaging database could accelerate the development of AI-powered tools for early cancer detection, treatment response monitoring, and personalized therapy planning. It could also support regulatory agencies in evaluating the safety and efficacy of AI-based medical devices by providing benchmark datasets for performance testing.
Moreover, a national data infrastructure could position China as a global leader in AI-driven healthcare innovation. By creating a large, diverse, and well-annotated dataset representative of the Asian population, Chinese researchers could address health disparities that are often overlooked in Western-centric datasets. This could lead to the development of AI models that are more equitable and effective for non-Caucasian populations.
However, the path forward is not without obstacles. Data privacy remains a paramount concern, especially given the sensitive nature of medical images and the potential for re-identification even in anonymized datasets. The authors stress the need for robust de-identification techniques, secure data access protocols, and ongoing ethical oversight to ensure that patient rights are protected.
They also acknowledge the challenge of data heterogeneity. Medical images in China are generated using a wide range of equipment from different manufacturers, each with its own proprietary formats and calibration standards. Harmonizing these differences—through standardized acquisition protocols or post-processing normalization techniques—will be essential for ensuring data quality and consistency.
Another hurdle is the shortage of trained personnel capable of curating and annotating medical images at scale. Expert radiologists are already overburdened with clinical duties, and manual annotation of complex imaging findings is time-consuming and prone to inter-observer variability. The authors suggest that semi-automated annotation tools, powered by AI themselves, could help alleviate this burden while maintaining high accuracy.
Despite these challenges, the momentum for change is growing. The Chinese government has increasingly recognized the strategic importance of AI in healthcare, as evidenced by its inclusion in national development plans and funding priorities. The National Natural Science Foundation of China and the National Science Fund for Distinguished Young Scholars have already supported research in this area, including the work of Shi and Liu.
The authors conclude with a call to action for the broader scientific community. They envision a future in which Chinese medical institutions collaborate to build a shared, standardized, and FAIR-compliant imaging data ecosystem—one that not only advances AI research but also improves patient outcomes through data-driven medicine. This will require sustained investment, interdisciplinary collaboration, and a cultural shift toward open science and data stewardship.
Ultimately, the success of AI in medical imaging will depend not on the sophistication of algorithms, but on the quality and accessibility of the data they are trained on. As Shi and Liu aptly point out, “data is the new oil” in the era of AI, but unlike oil, data gains value through sharing and reuse. By embracing the FAIR principles and building a national infrastructure for standardized medical imaging databases, China has the opportunity to unlock the full potential of its clinical data and lead the world in the responsible and equitable application of AI in healthcare.
The vision laid out in this commentary is both pragmatic and ambitious. It recognizes the complexities of real-world medical data while offering a clear roadmap for overcoming them. If implemented effectively, the proposed framework could transform the way medical imaging research is conducted in China, fostering innovation, enhancing reproducibility, and ultimately improving the lives of millions of patients.
As AI continues to reshape the future of medicine, the importance of data standardization cannot be overstated. The work of Zhenwei Shi and Zaiyi Liu serves as a timely reminder that technological progress must be grounded in sound data governance. Their call for a national effort to build FAIR-compliant medical imaging databases is not just a technical recommendation—it is a strategic imperative for the future of healthcare in China and beyond.
Medical Journal of Peking Union Medical College Hospital
Zhenwei Shi, Zaiyi Liu, Guangdong Provincial People’s Hospital
DOI: 10.12290/xhyxzz.2021-0507