How can medical AI move from "accurate diagnosis" to "practical application"? DIAGENS-B (02526) and Anzhen'er explore a new model of regional collaboration.

Medical imaging AI is entering a development stage different from its early days. In the past, the market focused on the accuracy of models on test sets; today, the industry is more concerned with whether models can enter real medical workflows, whether they can run stably across different hospitals and devices, and whether they can continuously generate service revenue under the premise of data compliance. On September 11, DIAGENS-B (02526) issued a voluntary announcement stating that the company signed a cooperation framework agreement with Zhejiang Anzhen'er Medical Artificial Intelligence Technology Co., Ltd., with a cooperation period of three years. The two parties plan to jointly build a Zhejiang provincial-level "digital intelligent imaging" capability foundation around pre-diagnosis consultation, imaging preliminary screening, and report interpretation, and explore an imaging service model of "examination at the grassroots level, recognition in the cloud, and review at the provincial level." As a Zhejiang provincial-level medical health AI infrastructure platform, Anzhen'er also operates the National Artificial Intelligence Scalable Application Pilot Base (Medical). The announcement also mentioned that the two parties will jointly develop high-quality medical datasets, AI training corpora, and scientific research and data analysis products, and explore a model service model billed by Token usage. On the surface, this is a business cooperation between enterprises; from an industry perspective, what is more noteworthy is that Diagens' medical imaging model capabilities may be placed into a collaborative network covering grassroots medical institutions, regional cloud, and provincial-level expert review. Four key issues in the development of medical AI This model corresponds precisely to four key issues currently being explored in international medical AI research. First, "reducing burden and triaging" for doctors: How does AI redistribute doctors' work? The entry of medical AI into clinical practice does not mean completely replacing doctors. A more realistic path is to let AI undertake preliminary screening, triage, quantitative analysis, and risk alerts, while handing complex cases and final judgments to professionals. Medical AI has never been intended to replace doctors, but to serve as doctors' "efficiency assistant": taking over time-consuming, repetitive work such as initial image reading, data organization, and quantitative measurement, so that doctors can free up energy to focus on the links that most require professional judgment, such as complex case assessment and final diagnosis and treatment decisions. This is equivalent to redeploying medical resources. The Swedish MASAI study published in The Lancet Oncology in 2023 is very persuasive: this breast cancer screening study covering nearly 80,000 women showed that after AI-assisted image reading, doctors' image reading workload directly decreased by 36.1%, while the cancer detection rate relatively increased by 28%, and the false positive rate was basically the same as the traditional process. Another study on grassroots diabetic retinopathy screening also confirmed that the AI system achieved a sensitivity of 87.2% and a specificity of 90.7%, which can greatly fill the capability gap at the grassroots level. The tiered imaging service model built by Diagens and Anzhen'er is precisely a benchmark for the localized implementation of this international cutting-edge concept. Grassroots hospitals are responsible for taking images and collecting data, cloud-based AI first completes preliminary imaging screening and draft reports, and complex cases are then transferred to provincial-level experts for review and verification. In this way, Diagens' AI is not just a piece of software installed in hospitals, but is deeply embedded in the imaging diagnosis and treatment process of the entire region, becoming a core supporting link in the tiered diagnosis and treatment system. Second, cross-institutional generalization validation: Can the model run stably across hospitals? Medical AI has a recognized industry challenge: a model trained very well in Hospital A may see its accuracy discounted when moved to Hospital B, with equipment from a different brand, and facing different patient populations. A high score at a single center is not real capability; only stable performance across institutions, devices, and populations can be considered truly deployable. A 2020 multinational study in Nature confirmed this: when the same AI system was validated using breast cancer screening data from the United Kingdom and the United States respectively, the results differed significantly. This also shows that testing at only one hospital cannot represent the real clinical level. The authoritative industry guidelines published in Nature Medicine in 2022 also clearly stated: to judge whether medical AI is good, one cannot only report beautiful laboratory data; one must look at whether it is safe in the real medical process, whether doctors find it smooth to use, what the missed diagnosis and misdiagnosis rates are, and whether clinical practice accepts it. Anzhen'er's collaborative network covering medical institutions across the province provides a natural multi-center validation field for Diagens' models. Hospitals at different levels, equipment from different brands, and patient populations with different characteristics constitute the most realistic "clinical examination room." The two parties will continuously validate and iterate models within this system, making AI more accurate and stable with use, and laying a solid clinical foundation for subsequent large-scale promotion. Third, balancing safety and efficiency: Can data-secure collaboration be achieved? Diagens Technology's announcement proposing "recognition in the cloud" does not mean that medical data will be uploaded centrally without distinction. When mentioning "cloud recognition," many people's first reaction is data security: Can patient privacy be protected? Can hospital data be transferred at will? This is also a common issue for global medical AI: it is necessary for multiple hospitals to jointly train models, but raw sensitive data cannot be arbitrarily centralized. The scientific research community has long had mature solutions: for example, "federated learning" technology, which in plain terms means "data stays put, models move" - raw data from each hospital does not need to leave the hospital, the model is trained separately at each node, and then the training results are aggregated and optimized. A 2020 multi-institutional brain tumor study in Scientific Reports confirmed that models trained in this way are almost no different in effect from training with centralized data, but with a much higher level of privacy security. In the cooperation between Diagens and Anzhen'er, the two parties will explore a hybrid data collaboration architecture that balances service efficiency and data security. For imaging service scenarios that require real-time results, efficient inference will be completed in the cloud within a compliant framework; for model training and iteration, technical paths such as federated learning and secure aggregation will be explored, supported by strict permission control and audit mechanisms, maximizing the value of data collaboration while safeguarding the bottom line of data security and privacy. Fourth, standardization of data assets: How to activate the long-term value of data? Many people think that with medical data, the more images stored, the more valuable it is. Actually, that is not the case. Scattered raw imaging files are only "raw materials." Only data that has undergone standardized governance, compliant authorization, and can be repeatedly reused is a truly valuable "asset." Benchmarking against international cutting-edge practice, the MIDRC (Medical Imaging and Data Resource Center) led by the U.S. NIBIB has formed an industry consensus: a high-quality medical imaging data infrastructure is a complete system covering unified data standards, de-identification processing, source traceability, access control, quality annotation, and standardized evaluation, rather than mere data accumulation. Its core is not how many images have been collected, but that a set of rules has been established for "how data is collected, how it is managed, how it is used, and how it is evaluated," allowing data to flow safely and be reused repeatedly. This is precisely the core direction of the two cooperating parties' exploration of "medical data assetization." Diagens and Anzhen'er will jointly develop high-quality medical datasets, AI training corpora, and scientific research analysis products. In essence, this is processing scattered imaging data into standardized, compliant, and reusable "data products." Diagens' previously accumulated 28.95 million+ high-quality annotated images and cooperation network of 99 hospitals have just laid a deep foundation for this, promoting the upgrade of medical data from "stored files" to assets that can generate value. Three layers of progressive value, opening the long-term growth ceiling If the cooperation is fully implemented, it will bring Diagens three layers of progressive commercial value and is expected to open entirely new growth boundaries. Relying on Zhejiang's massive outpatient and imaging examination resources, this model has a potential market space of RMB 10 billion. First, imaging services will be implemented, with more flexible pay-per-use. Scenarios such as pre-diagnosis consultation, imaging preliminary screening, and report interpretation will generate continuous demand for model invocation. The Token-based service model enables on-demand output of model capabilities, equivalent to "paying for what is used." Hospitals do not need to invest a large one-time cost, greatly lowering the threshold for use and better fitting the procurement habits of medical institutions. This makes large-scale penetration easier and brings continuous and stable service revenue. Second, regional platform reuse can amplify the boundaries of capability growth. As connected medical institutions and imaging tasks continue to expand, the company's full-chain capabilities in data governance, model training, deployment and operations, and feedback iteration will achieve cross-scenario reuse. Previously, projects were done one hospital at a time; now, implementation is happening in batches across an entire provincial platform, equivalent to upgrading from a single-project model to a platform replication model. The company's model service revenue in the first half of 2026 increased by 101.1% year-on-year, accounting for 86.9%, and its deep accumulation of 158 specialty models implemented cumulatively will achieve value amplification within the provincial platform. Third, the cooperation will also extend from clinical services to scientific research innovation. Problems in real clinical scenarios will continuously transform into high-quality datasets, specialty disease models, and scientific research tools, promoting the deep implementation of AI4S in the field of medical imaging, expanding the value boundary of medical AI from assisted diagnosis and treatment to supporting scientific research, and opening a longer-term value track. Summary Competition in medical imaging AI is shifting from "who can build a model" to "who can make models run safely, stably, and continuously in real medical networks." If the cooperation between Diagens and Anzhen'er can be gradually realized along the four lines of clinical workflow, multi-center validation, data governance, and commercial services, its significance will be more than adding one cooperation project. It will provide a new practical sample for medical imaging AI to move from technical capability to regional-level productivity. In the future, as the cooperation is gradually implemented, the company's long-term platform value as a Hong Kong-listed medical imaging AI leader will continue to be released.