Challenges and Models of Medical AI Industrialization: Insights from the Roundtable Forum at the First China Medical Imaging AI Conference

From December 15 to 16, the inaugural China Medical Imaging AI Conference, hosted by the China Alliance of Industry, Academia, Research, and Application for Medical Imaging AI (CAIERA, hereinafter referred to as the “Alliance”) and co-organized by the China Health Promotion Foundation, the Radiology Branch of the Shanghai Medical Association, and the Chinese Journal of Computerized Tomography, was grandly held at the Shanghai International Convention Center.

This article is compiled from the insightful remarks made by panelists during a roundtable discussion. Moderated by Chen Shengyu of Philips China Investment Co., Ltd., the session featured discussions among the guests on the industrialization of AI products and the associated processes.

The participating guests included: Ren Haiping, Director of the Optical, Mechanical, and Electrical Medical Device Inspection Laboratory at the National Institutes for Food and Drug Control; Wang Peijun, Vice President of Tongji Hospital Affiliated to Tongji University; Qiao Xin, CEO and Co-founder of Deepwise Healthcare; Fang Cong, Vice President of Yitu Healthcare; and Bian Haifeng, CTO and Founder of Wingtech Medical.

At the outset of the interview, host Chen Shengyu addressed the recent development of the AI industry: “If over the past two years our understanding of the entire sector was centered on exploring how AI could be applied in clinical medical settings, then this year the focus has unequivocally shifted to how AI products can achieve productization and commercialization, whether they can undergo public testing, and whether they can be deployed in real-world clinical scenarios.”

In fact, no company in China has yet obtained a medical device registration certificate for an AI product. In this regard, Ren Haiping stated, “Most AI products are standalone software, with a few being integrated hardware-software solutions; overall, they are at the inspection submission stage with the National Institutes for Food and Drug Control (NIFDC). Although there have been inquiries regarding innovative medical devices involving AI products, the typical commercialization pathway for AI is inspection submission, clinical testing, and regulatory review.”

The difficulty in obtaining evidence for AI products is a major factor constraining the industrialization of AI in the healthcare sector; furthermore, from the hospital perspective, integrating AI into clinical practice presents another significant challenge.

“Specifically, AI products for pulmonary nodule detection are the ones that have played a significant role in clinical practice. Current AI products have limited clinical utility; their accuracy, sensitivity, and specificity often fail to withstand rigorous evaluation, and their robustness is relatively poor,” said Wang Peijun. “Some AI products in hospitals are already applicable in clinical settings, such as those for pulmonary nodules, prostate cancer, and breast cancer diagnosis, while others are still under research and refinement, such as those for Alzheimer’s disease prediction.”

From the perspective of hospitals, there is still a gap between the practicality and stability of AI products and their clinical application. For enterprises, integrating products and algorithms into hospital workflows is a key R&D objective for many AI companies. In this regard, Qiao Xin from Deepwise, Fang Cong from Yitu, and Bian Haifeng from Wingtech shared their respective approaches.

Qiao Xin stated, “First, regarding disease screening, Deepwise has AI products for pulmonary nodules, breast conditions, and cerebral hemorrhage. Currently, more than 200 hospitals in China are trialing our equipment, with some having already integrated it into their clinical workflows. In fact, 30,000 AI reports are generated daily, over 10,000 of which are used for routine health screenings; however, substantial work remains to be done in the differential diagnosis of diseases.”

“Since its inception, Yitu Healthcare has adhered to four core principles: leveraging clinical data for product development, employing professional physicians for data annotation, integrating into clinical workflows, and conducting product feedback and iteration on a weekly basis.” Regarding the integration of AI products into clinical workflows, Fang Cong believes that efforts can be approached from three dimensions: business, operations, and technology.

From a business perspective, AI products must secure approval from all hospital departments and ensure that data is de-identified to prevent any leakage. From a technical perspective, the focus lies on the integration of AI products with the radiology departments of different hospitals. From an operational perspective, it refers to whether the operation of AI products aligns with physicians’ needs and habits, addressing detailed issues such as icon placement, brightness, and flicker frequency.

Bian Haifeng of Yizhan stated, “Since its establishment in 2009, Yizhan has been committed to leveraging the internet and AI to achieve three key objectives: improving quality, enhancing efficiency, and reducing costs. We operate our own imaging centers, provide imaging solutions, and manage a physician group. Our AI efforts primarily focus on digital radiography (DR). For disease detection, we use AI to trace data, thereby identifying limitations in diagnostic technologies such as CT and DR. Furthermore, we employ AI to develop customized algorithms that enable our AI products to generate reports, thus guiding diagnosis and minimizing errors.”

The AI technologies of the aforementioned three companies are closely tied to data. In fact, data is a key driver of AI. Ren Haiping stated, “Establishing closed third-party datasets is crucial. While the National Health Commission, universities, and research institutions are all developing datasets, it is primarily physicians or physician groups that are leading these efforts.” “As long as the dataset quality is controllable, coverage is extensive, clinical phenomena are adequately reflected, and the data meet our needs for product testing or evaluation, they will gain recognition. Data sources should reflect the intended use of the product and real-world scenarios, with the ultimate goal being physician-oriented rather than IT engineer-oriented.”

Physicians are a key force in AI research and development. In this regard, Wang Peijun believes, “Physicians need to be involved in raising questions, defining objectives, organizing data resources, annotating data according to standards, conducting system validation, and ultimately applying the technology in clinical practice—continuously refining it and identifying issues. Throughout this process, companies must heed the requirements of frontline clinical practitioners to establish the goals for their systems or products.”

“But currently, due to the lack of a shared language for academic discussion, there remains a knowledge gap between doctors and IT engineers, resulting in a continued disconnect in their R&D collaboration,” added Wang Peijun.

Generally, the business expansion of artificial intelligence can be pursued through two models along the industry chain: upstream and downstream. The upstream model involves collaborating with medical device manufacturers to integrate software with hardware by embedding products directly into the hardware. The downstream model entails selling software licensing rights or charging for related services to hospital health examination centers and third-party medical imaging centers.

Qiao Xin stated, “Current AI products can take the form of cloud-based or internet-based solutions. They are not limited to AI-assisted image interpretation in radiology but can also be applied in the development of smart hospitals.”

“Currently, all imaging-based clinical assistance products are used in clinical settings through product testing to continuously improve and enhance their performance. Additionally, many companies are simultaneously pursuing approvals from both the CFDA and the FDA,” said Fang Cong.