How can medical AI move from “seeing accurately” to “being usable”? Deshi-B (02526) explores a new model of regional collaboration with Anji

Medical imaging AI is entering a different stage of development than it was in the early days.

In the past, the market focused on the accuracy of the model in the test set; today, the industry is more concerned about whether the model can enter the actual medical process, whether it can operate stably between different hospitals and equipment, and whether it can continue to generate service revenue under the premise of data compliance.

The Zhitong Finance App learned that on September 11, Deshi-B (02526) issued a voluntary announcement stating that the company signed a cooperation framework agreement with Zhejiang Anjin Pediatric Artificial Intelligence Technology Co., Ltd., for a period of three years. The two sides plan to jointly establish a “digital imaging” capability foundation at the Zhejiang level around pre-diagnosis consultation, initial image screening, and report interpretation, and explore an imaging service model of “inspection at the grassroots level, recognition in the cloud, and review at the provincial level”.

As a healthcare AI infrastructure platform in Zhejiang Province, Anjianer also operates a national artificial intelligence large-scale application pilot base (healthcare). The announcement also mentioned that the two sides will jointly develop high-quality medical data sets, artificial intelligence training materials, scientific research and data analysis products, and explore a model service model that charges based on token usage.

On the face of it, this is a business cooperation between enterprises; from an industrial perspective, what is more noteworthy is that Deshi's medical imaging modelling capabilities may be incorporated into a collaborative network covering primary medical institutions, regional clouds, and provincial expert reviews.

Four key issues in the development of medical AI

And this model corresponds to the four key issues currently being explored in international medical AI research.

1. “Reducing the Burden and Diverting the Burden” for Doctors: How Can AI Redistribute Doctors' Jobs?

The entry of medical AI into the clinic does not mean that it will completely replace doctors. A more realistic approach is to let AI take on initial screening, triage, quantitative analysis and risk reminders, and hand over complex cases and final judgments to professionals.

Medical AI has never been meant to replace doctors, but rather as an “efficiency assistant” for doctors: taking over the time-consuming and repetitive tasks of reading movies, sorting out data, and quantitative measurement, freeing up doctors' energy to specialize in complex case investigation and final diagnosis and treatment decisions, which require the most professional judgment, which is equivalent to rearranging “arranging forces” for medical resources.

The Swedish MASAI study published in “The Lancet Oncology” in 2023 is very persuasive: this breast cancer screening study, which covered nearly 80,000 women, showed that after AI-assisted film reading, doctors' workload directly decreased by 36.1%, but the cancer detection rate increased relatively by 28%. The false positive rate was basically the same as the traditional process. Another study on screening for diabetic retinopathy at the primary level also confirmed that the AI system was 87.2% sensitive and 90.7% specific, which could greatly fill the grassroots capacity gap.

The hierarchical imaging service model created by Deshi and Safety Clinic is a benchmark for the localization implementation of this cutting-edge international concept. Primary hospitals are responsible for filming and data collection. Cloud AI first completes the initial screening of images and the first draft of the report, then transfers complex cases to provincial experts for review and inspection. In this way, Deshi's AI is not just a software installed in hospitals, but is deeply embedded in the imaging diagnosis and treatment process throughout the region, becoming the core support link in the hierarchical diagnosis and treatment system.

2. Cross-agency generalization verification: Can the model run stably across hospitals?

Medical AI has a recognized industry problem: if a model is well trained at Hospital A, the accuracy rate may be reduced if it is replaced by Hospital B, a different brand of equipment, and facing a different patient population. A high score in a single center is not a real skill; it is stable across institutions, devices, and people, so I dare say it can actually land.

A multinational study in “Nature” in 2020 confirmed this: the same AI system was tested using breast cancer screening data from the UK and the US, respectively, with significant differences in results. This also shows that testing at only one hospital does not represent the actual clinical level. The authoritative industry guide published by “Nature Medicine” in 2022 also clearly states that judging whether medical AI is good or bad should not only report beautiful data from the laboratory, but also depends on how unsafe it is in the actual medical treatment process, how difficult it is for doctors to use it, what is the rate of missed diagnosis and misdiagnosis, and whether it is clinically accepted.

An Clinic covers a collaborative network of medical institutions throughout the province, providing a natural multi-center verification field for the Desi model. Hospitals at different levels, different brands of equipment, and patient groups with different characteristics make up the most realistic “clinical examination room”. The two sides will continue to verify and iterate the model in this system to make AI more accurate and stable the more it is used, and establish a solid clinical foundation for subsequent large-scale promotion.

3. Security and efficiency: Can data security collaboration be achieved?

Deshi Technology's announcement proposed “recognition in the cloud”, which does not mean that medical data will be uploaded centrally without discrimination.

When it comes to “cloud recognition,” the first reaction of many people is data security: can patient privacy be guaranteed? Can hospital data be freely distributed? This is also a common issue for global medical AI: it is necessary for multiple hospitals to jointly train models, and raw sensitive data cannot be concentrated at will.

The scientific research community has long had mature solutions: for example, “federal learning” technology, colloquially speaking, “the data doesn't move the model” — the original 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 summarized and optimized. A multi-agency brain tumor study in “Scientific Reports” in 2020 confirmed that the model trained using this method has almost no difference in effectiveness from training by concentrating data, but the privacy security factor is much higher.

With the cooperation between Deshui and Security Clinic, the two sides will explore a hybrid data collaboration architecture that takes into account service efficiency and data security. For image service scenarios that require real-time results, efficient inference is completed through the cloud within a compliance framework; for model training iterative processes, technical paths such as federal learning and security aggregation are explored, supporting strict permission control and audit mechanisms to maximize the value of data collaboration on the premise of maintaining the bottom line of data security and privacy.

4. Standardization of data assets: How to activate the long-term value of data?

Many people think that medical data means that the more movies stored, the more valuable they are; in fact, this is not the case. Scattered original image files are only “raw materials”. Data that has undergone standardized management, compliance authorization, and can be reused over and over again is a truly valuable “asset.”

Benchmarking cutting-edge international practices, MIDRC (Medical Imaging and Data Resource Center), led by NIBIB in the US, has formed an industry consensus: high-quality medical imaging data infrastructure is a complete system covering unified data standards, de-identification processing, source traceability, access control, quality labeling and standardized evaluation, rather than simply data compilation. Its core is not how many images have been saved, but rather a set of rules for “how to collect, manage, use, and evaluate data” so that data can flow safely and be reused over and over again.

This is the core direction of the two partners exploring the “capitalization of medical data.” Deshi and Anji will jointly develop high-quality medical data sets, AI training materials, and scientific research analysis products. Essentially, they process scattered image data into standard, compliant, and reusable “data products.” Deshi's previous accumulation of 28.95 million+ high-quality labeling images and a cooperative network of 99 hospitals just laid a deep foundation for this matter, promoting the upgrading of medical data from “stored documents” to value-generating assets.

Three levels of value progression, opening the ceiling for long-term growth

If the cooperation is fully implemented, it will bring three levels of progressive commercial value to Deshi and is expected to open up new growth boundaries. Relying on Zhejiang's massive outpatient and imaging resources, this model has a potential market of 10 billion imaginative space.

First, the video service has been implemented, and pay-as-you-go is more flexible. Scenarios such as pre-diagnosis consultation, preliminary image screening, and report interpretation will create continuous model call requirements. The service model of token measurement enables on-demand output of model capabilities, which is equivalent to “how much to use”. Hospitals do not need to invest a large amount of money at once, drastically reducing the threshold of use. It is also more in line with the procurement habits of medical institutions, making it easier to penetrate on a large scale, and bring continuous and stable service revenue.

Second, the reuse of regional platforms can expand the boundaries of capacity growth. With the continuous expansion of access to medical institutions and imaging tasks, the company's full-link capabilities in data governance, model training, deployment, operation and maintenance, and feedback iteration will be reused across scenarios. In the past, one hospital did the project, but now the entire provincial platform has been implemented in batches, which is equivalent to upgrading the 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%. The company's deep accumulation of 158 specialist models will achieve value amplification in provincial platforms.

Third, the cooperation will also extend from clinical services to the field of scientific research and innovation. Problems in real clinical scenarios will continue to be transformed into high-quality data sets, specialty models and scientific research tools, promote the deep implementation of AI4S in the medical imaging field, expand the value boundary of medical AI from auxiliary diagnosis and treatment to assisting scientific research, and open up a longer-term value circuit.

Summarize

The competition for medical imaging AI is shifting from “who can make a model” to “who can make the model run safely, stably, and continuously in a real medical network”.

If the collaboration between Deshi and Anji can be gradually realized along the four lines of clinical workflow, multi-center verification, data governance, and commercial service, its significance will not only be to add a cooperative project, but also to provide a new practical sample for medical imaging AI to move from technical capability to regional productivity.

In the future, with the gradual implementation of cooperation, the company's long-term platform value as the leading medical imaging AI leader in the Hong Kong stock market will continue to be unleashed.