Innovative Mechanism Takes Shape: Conference Lays New Foundation for AI in Medical Imaging
Medical imaging, as one of the core components of clinical diagnosis, is currently facing multiple urgent demands, including enhancing capabilities at primary care levels, optimizing resource allocation, and alleviating talent shortages.
The development of AI technology has provided a direction for solving these problems. In recent years, the national government has introduced multiple policies to promote the integrated innovation of AI technologies. This August, the State Council issued the "Opinions on Deepening the Implementation of the 'AI Plus' Action." Regarding public welfare, the "Opinions" emphasize "exploring and promoting high-level resident health assistants accessible to all, orderly advancing the application of artificial intelligence in scenarios such as assisted diagnosis and treatment, health management, and medical insurance services, and significantly enhancing the capacity and efficiency of primary healthcare services." The intensive rollout of these policies has provided clear guidance for the integrated innovation of artificial intelligence and medical imaging, further driving the technological translation and expansion of application scenarios for AI technology in the field of medical imaging.
However, as the industry accelerates its expansion, it is worth noting that previous AI-based medical imaging products focused on single diseases in specific anatomical regions have provided limited clinical support.The industry urgently needs to establish an innovative mechanism—guiding the exploration of AI-based medical imaging products through pioneering innovative products, and further ensuring the feasibility of innovation by fostering collaborative mechanisms in data construction and industrial development.
At the “AI Empowerment, Data Leading the Future” High-Level Summit on Innovation Across the Entire Intelligent Medical Imaging Industry Chain—guided by the Beijing Municipal Bureau of Economy and Information Technology and the Administrative Committee of Beijing Economic-Technological Development Area, hosted by Beijing Yizhuang Smart City Research Institute Group Co., Ltd., and co-organized by Beijing Data Pioneer Zone Service Co., Ltd., Beijing International Computing Power Service Co., Ltd., Yimai Yangguang Group Co., Ltd., Yinghe Yimai Intelligent Technology Co., Ltd., and other entities—VCBeat observed that a mechanism featuring “innovation-driven product leadership, innovation-backed data-supported R&D, and innovation-alliance collaborative development” is taking shape.
The multiple blockbuster achievements released at the conference all signal to the outside world that the field of medical imaging AI is entering a new era of transformation.
In China, the field of clinical medical imaging faces two core pain points. According to LeadLeo Research Institute, first, there is a significant shortage of radiologists. Currently, China has only one radiologist per 70,000 people, whereas the United States has one per 2,000 people. This shortage, compounded by diverse diagnostic demands, results in extremely low diagnostic efficiency and an urgent need for innovation in service models. Second, the misdiagnosis rate is exceedingly high. In the United States, there are 12 million medical imaging misdiagnoses annually, while in China, this figure reaches as high as 57 million. The misdiagnosis rate for medical imaging in China is significantly higher than that in the United States.
In fact, in addition to the high number of misdiagnoses, the rate of missed diagnoses in medical imaging is also relatively high. Taking lung cancer as an example, the article “Analysis of Common Causes of Missed CT Diagnoses of Lung Cancer and Imaging Manifestations,” published in the December 2024 issue of the Journal of Practical Radiology, categorized missed CT diagnoses among the lung cancer patients included in the study into six types. Among these, Type IV refers to cases where lung cancer is accompanied by a large amount of pleural effusion, with the tumor embedded within atelectatic lung tissue; such cases are prone to being missed due to incomplete observation, with the missed diagnosis rate reaching as high as 27.18%.
There is a substantial clinical demand for high-quality medical imaging AI products that can enhance labor efficiency and reduce rates of misdiagnosis and missed diagnosis. However, existing medical imaging AI solutions fail to adequately meet these clinical needs. This shortfall stems from the design philosophy of previous generations of medical imaging AI, which have predominantly focused on “single-disease, single-lesion” detection. For instance, in pulmonary imaging, current AI tools typically detect only a single lung nodule or case of pneumonia per analysis. In contrast, clinicians require comprehensive auxiliary diagnostic solutions capable of identifying all potential abnormalities within the region of interest during a single examination.
At the conference, Yinghe Yimai, in collaboration with West China Hospital of Sichuan University and Huawei’s Healthcare Corps, officially released the core achievements of the “Hetu Initiative”: the AIR (AI-based Radiology) tool for pathway-level auxiliary diagnosis of chest CT scans.AIR distinguishes itself from previous medical imaging AI products by adopting a multi-disease approach covering various lesion sites and starting with “pathway-level assisted diagnosis,” thereby fundamentally addressing the mismatch between earlier medical imaging products and clinical needs—
First, a leapfrog improvement in diagnostic efficiency. Unlike single-disease medical imaging AI, AIR can detect 31 common abnormalities—including pulmonary nodules, pneumonia, rib fractures, and hiatal hernia—from a single non-contrast chest CT scan. Furthermore, AIR automatically generates structured reports that precisely annotate the location, morphological features, and severity of detected lesions, significantly assisting radiologists in image interpretation. Additionally, AIR features an intelligent “Smart Eye” viewer that automatically links report descriptions to corresponding lesions when reviewing physicians click on them in the images.
Second, the detection accuracy has been significantly improved. AIR has largely addressed the issue of missed diagnoses in manual interpretation. Currently, AIR covers more than 19 AI Sub-Units (AISUs), with a lesion coverage rate of 75%. Relevant studies have shown that in preliminary tests involving 100 non-contrast chest CT scans from six medical institutions, AIR significantly outperformed human radiologists. Moreover, AIR can successfully detect lesions that are prone to being missed by humans, such as early-stage pulmonary space-occupying lesions, occult rib fractures, and esophageal thickening.
Yinghe Yimai’s AIR product for assisted diagnosis in non-contrast chest CT scans covers the entire examination workflow, from “image analysis” to “report generation,” fully meeting physicians’ core need for “efficient and accurate report writing.”
Current data also demonstrate its tangible effectiveness. With the assistance of AIR, the average time required for manual report writing was reduced from 8.4 minutes to 7 minutes, representing a 16.7% improvement in efficiency; the average time for manual review and publication of reports was reduced from 4.3 minutes to 3.4 minutes, marking a 20.9% efficiency gain. Physicians participating in the evaluation noted that, empowered by AIR, they could devote more attention to the assessment of complex lesions, thereby significantly enhancing their work efficiency.
According to reports, following the implementation of AIR’s innovative products—the core achievement of the “Hetu Plan”—Yinghe Yimai will also launch the “Luoshu Plan.” This initiative aims to collaborate with multidisciplinary experts to build a multimodal dataset integrating “imaging + clinical” data, further advancing medical imaging AI from “assisted diagnosis” to “assisted diagnosis and treatment.” This move is expected to lead new directions for industry development.
As the flagship product AIR continues to achieve breakthroughs and undergo continuous refinement, it is also charting the course for the future development of the AI in medical imaging industry. The sector is poised to advance beyond the era of single-disease and single-lesion analysis into a new phase characterized by comprehensive diagnosis of multiple pathologies within specific anatomical regions.
The implementation of innovative AI products in medical imaging has always been guided by a key logic: informatization is the source of data, and data is the core fuel for AI development.
Behind the creation of Yinghe Yimai AIR’s innovative products lie two critical factors. First is the support of massive datasets. This data foundation includes pre-training on tens of millions of proprietary medical imaging cases from its incubator, Yimai Yangguang, injecting vast amounts of clinical prior knowledge, as well as 100,000 meticulously annotated cases exclusive to the “Hetu Project.” Second is the dual support of “medical authority + technological infrastructure.” Professors from top-tier Grade A tertiary hospitals, such as West China Hospital, participated throughout the process of establishing standards. They deconstructed over 3,000 clinical reports to develop a “Core Definition Table,” ensuring the product aligns seamlessly with physicians’ workflows. Meanwhile, Huawei’s Healthcare Business Unit provided the AI computing, storage, and networking infrastructure. Through XPU pooling and intelligent scheduling technologies, it increased concurrent imaging inference tasks by 30% and supported “zero-code development,” enabling hospitals to rapidly adapt to localized needs.
It is worth noting that Yimai Yangguang and Yinghe Yimai have gradually developed a unique incubation twin mechanism over their long-term growth. Yimai Yangguang’s sustained focus on the informatization of imaging departments has facilitated the accumulation of imaging data for the development of innovative products. Yinghe Yimai has further driven the upgrading and iteration of AI-based medical imaging products, which in turn feeds back into the healthy development of imaging departments.
This “incubation twin” mechanism, characterized by “a continuous source of fresh data and ample, high-quality fuel,” essentially establishes a closed loop from informatization to data, and then from data to AI capabilities. This is what enables Yinghe Yimai AIR to rapidly respond to clinical needs and achieve efficient product development and deployment.
As a core element, data will play a pivotal role in the future development of innovative AI-powered medical imaging products. However,Medical imaging data has historically faced multiple challenges. On one hand, the industry lacks standardized, high-quality training datasets. Although there is an abundance of medical imaging data, most sources are unverified, and acquiring high-quality data from leading hospitals remains challenging. Furthermore, medical imaging datasets often suffer from limited disease diversity and severe data scarcity for certain conditions. On the other hand, the industry lacks unified standards for training datasets. The varying standards adopted by different enterprises hinder cross-industry communication and collaboration. Additionally, unclear data sources raise concerns regarding regulatory compliance in data usage. These issues have significantly impeded the translation of medical imaging AI from theoretical research into clinical practice.
Today, the industry is attempting to address this issue, and the advancement of corpora constitutes a crucial component of these efforts.
At the conference, the “High-Quality Medical Imaging Corpus” (hereinafter referred to as the “Corpus”) was officially released. The industry is expected to leverage this Corpus to accelerate the development of innovative AI models for medical imaging. Notably, Yimai Yangguang and Yinghe Yimai were deeply involved in the construction of the Corpus, providing substantial support with extensive medical imaging data.Currently, the corpus has established a data system characterized by high coverage, high availability, high trustworthiness, and high security, centered on four core dimensions: “scale, diversity, quality, and compliance,” to further enhance the generalization capability and robustness of AI models.
In terms of data scale, the corpus comprises over ten million cases, covering multi-modal data across all disease types. As of August 4, 2025, the total number of imaging studies in the corpus reached 12,116,964. It aggregates data from multiple regions and various scanner manufacturers, encompassing all disease categories and mainstream imaging modalities such as CT, MR, DR, and PET/MR. The dataset covers 11 major anatomical regions, including the chest, head, abdomen, and spine, thereby meeting the AI research and development needs across diverse disease areas.
In terms of data diversity, the corpus utilizes multidimensional data to fully accommodate the needs of different R&D stages. Currently, it features comprehensive sample dimensions, covering various age groups, gender ratios, and classifications of positive and negative samples. It can sufficiently meet the data requirements for different stages, such as early-stage algorithm exploration (which requires small batches of precisely annotated data) and mature model optimization (which requires large-scale generalized data). Furthermore, up to 50,000 cases of publicly annotatable data provide convenient access for research institutions and small and medium-sized enterprises.
In terms of quality, rigorous quality control processes and data de-identification ensure high data credibility and usability while safeguarding patient privacy. Regarding compliance, data sources are fully traceable throughout the entire process, ensuring that data compliance is verifiable and auditable.
Meanwhile, the corpus offers data dashboards and a data sandbox feature, further empowering developers to assess data suitability and providing a secure, compliant environment for data usage, thereby reducing barriers to data application.
For AI companies, application research and development based on clinical data are often constrained by collaborations with a limited number of hospitals, which may lead to regional limitations in algorithm performance and hinder widespread adoption. The release of this corpus, building upon the foundation of a “data warehouse,” marks the first step in constructing a prototype of an “enablement platform” for the full-chain development of medical AI. This initiative establishes the “new infrastructure” for medical imaging AI, covering four core scenarios: “scientific research, technology, industry, and clinical practice.” It will comprehensively support the entire process from academic research-level exploration at universities and institutes to enterprise-level clinical applications, providing precise, all-around enablement for frontier academic exploration, model development, product commercialization, and clinical deployment.
AIR’s innovative products are pioneering new directions for the industry, while corpus construction supports innovative R&D from multiple perspectives, including data and efficiency. However, a critical element is still lacking in the process of ensuring the commercialization of innovative products and the establishment of innovation mechanisms—collaboration.
Countless historical facts have demonstrated that collaboration between academic institutions and the industry is an indispensable component for achieving virtuous, sustainable development within the sector. A lack of information sharing among academic institutions, enterprises, and other stakeholders, coupled with insufficient industry-wide coordination, may lead to redundant and ineffective infrastructure development by practitioners, resulting in a waste of resources.
The results fall into two categories:On the one hand, due to a lack of synergy, startups struggle to meet the full-spectrum needs of hospitals and find it difficult to develop truly effective intelligent solutions, which may hinder their growth. On the other hand, medical imaging AI products remain isolated from one another and lack unified standards, making it difficult for them to be integrated into clinicians’ workflows and limiting their utility to specific scenarios. As a result, high-quality medical imaging AI products are hard to come by in clinical practice.
Today, this situation may see significant improvement.At the conference, the “Medical Imaging Intelligence Alliance (MIIA)” (hereinafter referred to as the “Alliance”) was officially established.The Alliance aims to build a community of shared destiny, centered on ecological win-win cooperation and dedicated to enhancing physicians’ user experience, thereby helping medical imaging AI products break free from their “isolated” predicament.Currently, the first batch of members, including Beijing Yizhuang Smart City Research Institute, Vmai Yangguang, Yinghe Yimai, Alibaba Cloud Computing, MaiFlow Technology, Yizhun Intelligent, Yingwei Medical, Ziwei Dixing, and Everbright Hongda, have completed their signing, laying the foundation for the alliance’s subsequent development.
To address the issues arising from the lack of top-level design in the current industry, the alliance is implementing a series of measures:
On one hand, the alliance is working to unite all members in jointly formulating and implementing the “MIIA Alliance Integration Standard,” achieving comprehensive standardization across data interfaces, algorithm encapsulation, result output, and service ports. This initiative aims to eliminate past compatibility challenges, allowing practitioners to focus their efforts on core technological innovation. On the other hand, by building the MIIA platform, the alliance provides industry enterprises with a “plug-and-play” technical foundation, realizing synergistic effects where the whole is greater than the sum of its parts. The platform enables hospitals to freely combine AI modules according to clinical needs, facilitating the construction of customized, end-to-end intelligent workflows. Meanwhile, enterprises can concentrate on their core competencies and reach clinical hospital users through the platform. Furthermore, the alliance is poised to establish a virtuous cycle in which “clinical needs drive technological iteration, and technological breakthroughs feed back into clinical development,” thereby addressing the historical mismatch between technology and clinical demands and better coordinating industrial forces to tackle industry-wide challenges.
Currently, the Alliance is inviting ecosystem partners to join and share the fruits of collaborative innovation. Participating association members will be better positioned to coordinate clinical resources for R&D, contribute to the development of the “MIIA Alliance Integration Standards,” facilitate technology promotion, and leverage shared market resources to accelerate the clinical implementation of innovations.
With the establishment of the alliance, we will continue to refine ecosystem collaboration mechanisms and deepen adaptation to clinical scenarios. By gradually expanding the scope of cooperation, we aim to realize collaborative value in a broader range of medical settings, thereby injecting new momentum into the standardized and ecosystem-driven development of the medical imaging AI industry.
While a single conference may have limited scope, we were delighted to witness the gradual establishment of an innovative mechanism in the field of medical imaging AI at this event. This mechanism is designed to lead industry innovation through the launch of AIR, an innovative medical imaging AI product; to support subsequent R&D efforts by building foundational corpora; and to foster collaborative innovation through the establishment of the Medical Imaging Intelligent Industry Alliance (MIIA). It will play a pivotal role in the future development and application of innovative medical imaging AI products. This initiative may well bring about a true transformation in the field of AI-powered medical imaging, spurring the emergence of more AI solutions that meet clinical needs and, through hospital-based clinical applications, truly benefit a broad patient population.