Recently, the 4th Hong Kong Clinical-driven Robotics and Embodied AI TEchnology Symposium (CREATE Symposium), hosted by the Center for Artificial Intelligence and Robotics (CAIR) of the Hong Kong Innovation Institute of the Chinese Academy of Sciences and co-organized by the Hong Kong Science and Technology Parks Corporation, was grandly held at the Charles K. Kao Auditorium of the Hong Kong Science Park.
As an important international exchange platform bridging clinical medicine and cutting-edge technology, this forum took "The Path to Commercialization of Medical AI" as its theme, bringing together nearly 300 experts, scholars, hospital administrators, authoritative clinical physicians, and industry and capital representatives. Closely focused on clinical pain points in specialty fields such as respiratory intervention, otolaryngology, ultrasound imaging, and sleep medicine, the event featured diverse segments including cutting-edge technology sharing sessions, a hospital president roundtable dialogue, a venture capital roadshow, thematic group discussions, a scientific and technological achievements exhibition, and exploration of embodied intelligence operating room scenarios, accelerating multi-party collaborative innovation among medicine, research, industry, and investment, and injecting strong momentum into the translation of scientific research results into medical practice.
At the opening ceremony, speeches were delivered in succession by CAIR Director Prof. Liu Hongbin, Hong Kong Science and Technology Parks Corporation CEO Wong Bing-hau, China Resources Health General Manager and China Resources Medical President Zhang Chuang, Executive Dean of the Hong Kong Innovation Institute of the Chinese Academy of Sciences Hao Yinxing, Deputy Director of the Institute of Automation of the Chinese Academy of Sciences Prof. Lyu Pin, and CEO of Shenzhen MGI Tech Co., Ltd. Liu Jian.
At this forum, CAIR made a major release of its self-developed multimodal clinical intelligent agent system CARES 4.0. Built on the Harness agent framework, the system is underpinned by a self-developed multimodal medical foundation model covering CT, MRI, ultrasound, endoscopy and EEG, breaking through the limitations of traditional single-point medical AI models. It is deeply embedded in hospital clinical workflows and, acting as an intelligent partner for physicians, covers the full chain of scenarios spanning diagnosis and treatment, department management, and medical research.
At the clinical application level, CARES 4.0 has completed deployment validation at multiple top-tier (Grade A tertiary) hospitals. In diagnostic and treatment scenarios, it enables full-process auxiliary diagnosis from patient chief complaints, imaging examinations, and pathology evaluation through to final confirmed diagnosis. In department management scenarios, it supports natural language interactive queries, automatically aggregates information such as equipment workload and disease distribution, and generates department operation reports with one click. In medical research scenarios, it can conduct multidimensional statistical analysis on clinical specialty disease datasets, train simple machine learning models, and output analysis reports that can be directly polished and used for journal submissions. The system is accompanied by a clinical high-quality data accumulation platform that can automatically clean, quality-check, and structurally consolidate multi-source in-hospital data, laying a solid data foundation for continuous AI iteration.
Currently, the CARES 4.0 official website is open, the framework and general toolkit have been released, and the medical toolkits are being launched one after another. The center's self-developed ultrasound and surgical video open-source models will also be integrated into the platform and made freely available to the public.
CAIR researcher Yi Dong introduces CARES 4.0
Facing the practical challenges of AI's deep integration into the healthcare system and its move toward clinical implementation, this year's forum specially featured a Hong Kong–Shenzhen hospital CEOs roundtable dialogue, inviting IHH Healthcare (North Asia) Regional Chief Executive Officer and Gleneagles Hospital Hong Kong Chief Executive Officer Kenneth Tsang, University of Hong Kong-Shenzhen Hospital President Zhang Wenzhi, CUHK Medical Centre Chief Executive Officer Fung Kin Lai, Guangdong 999 Brain Hospital President Cai Linbo, Shenzhen Luohu People's Hospital Executive President Xiong Yi, The Third Affiliated Hospital of Sun Yat-sen University Lingnan Hospital Management Committee Director Hu Xiquan, and HKU Faculty of Dentistry Assistant Dean (Teaching and Technological Innovation) Lam Yu Heng, among other managers from multiple medical institutions, to participate in exchanges and discussions.
Roundtable panelists shared their views: when private medical institutions introduce high-value AI medical systems, they need to assess return on investment, balancing operational sustainability with public health value; cross-border medical AI applications should strictly observe compliance red lines, explore data interoperability within the legal framework, and coordinate safety, compliance, and operational efficiency; smart hospital construction should focus on restructuring business processes, driving AI integration across the entire chain of diagnosis and treatment, nursing, and management; clinical implementation of AI depends on high-quality data and physician–engineer collaboration, with the core challenge at present being the definition of risks and allocation of liability in clinical scenarios, making it urgent to establish a multi-party risk-sharing mechanism; when promoting AI in primary care, it is advisable to proceed in phases, prioritizing experience-optimization applications, while high-risk AI diagnostic tools require a well-developed supporting safety assurance system; in the field of rehabilitation, AI can empower rehabilitation assessment and plan formulation, reducing therapists' workload; AI will become a fundamental capability in healthcare—dental specialties can leverage AI for precision diagnosis and treatment as well as at-home screening, achieving refined management of oral health.
In the keynote-sharing session, this year's forum brought together a number of the world's leading scholars in medical-engineering innovation, including Professor Sebastien OURSELIN, Fellow of the Royal Academy of Engineering and the Academy of Medical Sciences (UK) and Vice-President (Innovation) of King's College London; Professor Nassir NAVAB, member of Academia Europaea and Director of CAMP at the Technical University of Munich; Professor Yunhui Liu (Liu Yunhui) of Mechanical and Automation Engineering at The Chinese University of Hong Kong; Professor Hongbin Liu (Liu Hongbin), Director of CAIR; Professor Hongen Liao (Liao Hong'en), Chair Professor and Dean of the School of Biomedical Engineering at Shanghai Jiao Tong University; and Professor Dong Ni (Ni Dong) of the School of Artificial Intelligence at Shenzhen University. Standing at the international frontier, the experts delivered in-depth analyses of the development trends of embodied-intelligence medical technology around hot topics such as the medical device and digital health ecosystem, human-robot collaboration in medical robotics, AI-driven surgical robots, clinical agent-based AI systems, imaging-based decision-making for diagnostic and therapeutic robots, and the translation of intelligent ultrasound research into production, sharing the fruits of global technological innovation.
Academician Sebastien OURSELIN shared "From Bench to Bedside to Boardroom - An Innovation Ecosystem for Medical Devices and Digital Health"
Academician Nassir NAVAB shares "Dyadic Partnership(DP): A Missing Link Towards Full Autonomy in Medical Robotics"
Professor Liao Huai, Deputy Director of the Department of Respiratory and Critical Care Medicine at the First Affiliated Hospital of Sun Yat-sen University, Professor Lei Dapeng, Director of the Department of Otolaryngology at Qilu Hospital of Shandong University, Professor Hou Gang, Deputy Director of the Department of Respiratory and Critical Care Medicine at China-Japan Friendship Hospital, Professor Xie Xiaoyan, Director of the Department of Ultrasound Medicine at the First Affiliated Hospital of Sun Yat-sen University, Professor Tao Guowei, Director of the Ultrasound Department at Qilu Hospital of Shandong University, Professor Cui Xinwu, Director of the Department of Ultrasound Imaging at Tongji Hospital Affiliated to Tongji Medical College of Huazhong University of Science and Technology, Professor Cao Yingjuan, Director of the Nursing Department at Qilu Hospital of Shandong University, Wong Wang-hung Michael, Professor in the Department of Surgery and Chief of Cardiothoracic Surgery at the Faculty of Medicine of The Chinese University of Hong Kong, Professor Zhang Bin, Director of the Department of Psychiatry and Psychology (Sleep Medicine Center) at Nanfang Hospital of Southern Medical University, Professor Pan Jiyang, Director of the Sleep Medicine Center at the First Affiliated Hospital of Jinan University, Professor Xing Yingqi, Director of the Vascular Ultrasound Department at Xuanwu Hospital of Capital Medical University, Dr. Mak Hoi-kwan, Deputy Chief and Consultant of the Neurosurgery Department at Queen Elizabeth Hospital in Hong Kong, and Tam Siu-chung, Clinical Assistant Professor in the Department of Surgery of the School of Clinical Medicine at the University of Hong Kong, among others, drew on years of accumulated experience in their respective specialties to exchange frontline clinical practice insights in areas including medical imaging, ultrasound diagnosis and treatment, intelligent nursing, cardiothoracic surgery, sleep medicine, vascular ultrasound, and neurosurgery.
In addition, scholars including Professor Li Zongming, Dean of the School of Biomedical Engineering at Shenzhen University of Technology, Professor Hung-Leung Yam? of the Department of Electronic Engineering at The Chinese University of Hong Kong, Professor Kwok Kai-wai of the Department of Mechanical and Automation Engineering at The Chinese University of Hong Kong, Zhu Zhiguang, tenured Associate Professor in the Department of Mechanical Engineering at the National University of Singapore, and Dou Qi, Associate Professor in the Department of Computer Science and Engineering at The Chinese University of Hong Kong, and Jiang Zhongliang, Assistant Professor in the Department of Mechanical Engineering at the University of Hong Kong, also shared research progress from different dimensions such as AI algorithm development, medical robotics systems, and clinical translation, showcasing from multiple perspectives the diverse pathways through which medical-engineering cross-disciplinary collaboration empowers the development of smart healthcare.
In addition, this forum also featured a dedicated venture capital roadshow session, focusing on the frontier tracks of medical technology, bringing together innovative enterprises, research institutions and investment institutions to build a platform for project exchange and resource matching. Judges helped analyze technical barriers, clinical validation and commercialization strategies, assisting teams in clarifying their path to market and accelerating the industrialization of medical technology achievements.
Multiple hardcore tech innovation projects took part in the roadshow, covering cutting-edge directions such as optical multi-dimensional force sensors, endoscopic surgical robots, novel instruments and auxiliary equipment for minimally invasive surgery, sleep large models, ultrasound robots, surgical robots, surgical navigation systems, ASIC chips and active implantable medical devices, brain-computer interfaces and bionic muscles, surgical planning and simulation, and personalized orthopedics and reconstruction. Investment institutions and industry representatives including CEC Capital, Guangyuan Investment, Jiuguang Capital, Allwinner Technology, Yuexiu Group, Yanghe Investment, Pinehurst Medical Health Fund, CICC Capital, BGI Pinehurst Life Science Fund, Guoke Jiahe, and Jianyuan Tianhua attended the event, engaging in face-to-face in-depth exchanges with startup teams and fully matching investment and financing needs.
Dr. Yu Chen, co-founder of Huali Chuang Science, explains the key technical points of optical multi-dimensional force sensor technology.
As a co-organizer of this forum, the Hong Kong Science and Technology Parks Corporation (HKSTP) organized a themed session titled “Empowering Founders, Enabling Growth: HKSTP's Entrepreneurship Journey from Idea to Commercialization”, focusing on the industrialization of cutting-edge technological achievements. It gave a detailed introduction to the park's life and health technology ecosystem and entrepreneurial community, invited representatives of platform startups to share commercialization insights and practical cases, and held multi-dimensional dialogues with representatives from various fields, exploring how strategic collaboration can empower value creation and business growth, providing vivid practical examples and ecosystem support for medical-enterprise collaboration and the translation of innovation achievements.
In addition, the forum concurrently hosted a Medical Technology Achievements Exhibition, where sci-tech companies including Huali Chuang Science, Endong Medical, Shenzong Technology, Houkai Medical, Meixin Chuangke, and Lejin Purification, along with more than ten startups incubated by the Incu-Bio platform of Hong Kong Science Park, made a concentrated appearance, comprehensively showcasing cutting-edge technologies and innovative achievements in the fields of intelligent healthcare, digital healthcare, and precision medicine.
At the end of the event, attending guests moved to the CAIR Embodied Intelligent Operating Room, where they observed on-site cutting-edge achievements including the CARES multimodal clinical agent system, laryngoscope and bronchoscope diagnostic agents, the AI sleep system, the "Lingyin" ultrasound large model, embodied intelligent handheld ultrasound, an ultrasound scanning robot, a minimally invasive surgery solution for confined spaces, a real-time medical physics simulation platform, low-field MRI for interventional surgery, and MicroNeuro, an embodied intelligent surgical robot for neurosurgery. During the tour, the guests attentively listened to technical introductions from the R&D teams, gaining an in-depth understanding of the core technical features, clinical diagnostic and treatment value, and practical application scenarios of each achievement, and held in-depth discussions on topics such as technical R&D bottlenecks, clinical pilot progress, and paths for industrialization, helping intelligent diagnosis and treatment technologies break through scenario limitations and expand their coverage of applications.
In conversation with CAIR Director Liu Hongbin: CARES 4.0's clinical implementation and the path to commercialization of medical AI
In addition to the above-mentioned achievements releases and showcases, during the forum we also held a conversation with CAIR Director and researcher Liu Hongbin, discussing topics including the capability upgrades of CARES 4.0 compared with previous generations, the clinical presentation of surgical risk warnings, hallucinations and safety boundaries of medical AI, and the path to commercial implementation of medical AI.
The following is a transcript of the conversation, edited by Leifeng.com without altering the original meaning:
Leifeng.com (WeChat account: Leifeng.com): Compared with 3.0, is 4.0 a matter of more functions, or a change in the nature of its capabilities? Can you give a specific example to illustrate this difference?
Liu Hongbin:From CARES 1.0 to 3.0, it was itself a large model. At 3.0, it could only do question answering—this Q&A could be multimodal, for example, answering questions about an image or about a video segment; this is what versions 1.0 to 3.0 addressed. But 3.0 itself was merely a tool; a doctor had to know how to use this tool to complete specific tasks.
With 4.0, we upgraded it into an agent, so it can truly get work done for doctors. Take the example given at this launch event: a clinical department director wants to compile last month's surgery volumes, consumables usage, and patient conditions into a report, to present to hospital leadership or for internal departmental discussion. Previously, this work had to be done by the director's assistant using CARES 3.0; now no assistant is needed—the entire agent system completes it autonomously. During the process, there may be some interaction with the director to confirm his ideas and what type of analysis report should be produced—we have built this kind of interaction in as well.
Leifeng.com: Among the five capabilities released this time, the semantic understanding of surgical videos and risk warnings have drawn the most attention. How are these warnings presented in real surgeries?
Liu Hongbin: Warnings are divided into several aspects. One is spatial warnings about the operating area: when AI identifies that the site being operated on is very close to important blood vessels, arteries, or nerves, it should alert the doctor, indicating that this step may be risky and that they should not touch that location.
Another type is warnings across the time dimension. Surgery is a procedural process—one step must be completed before moving on to the next. If the doctor does not perform the first step adequately—for example, in neurosurgery, a bone window must first be opened to create a channel, with the ultimate goal of completely and thoroughly removing the tumor inside. If the bone window is not opened large enough, the tumor inside may not be removed completely, and by the time you go deeper and try to enlarge it, it will be too late. This kind of warning along the time dimension is also one dimension.
What is mainly in use now is spatial early warning, for example telling the doctor: the area you are operating on is already very close to a very important anatomical structure, and you need to avoid it.
Lei Feng Network: In which hospitals and departments is it currently being used, and to what extent?
Liu Hongbin:For example, at the First Affiliated Hospital of Sun Yat-sen University, we work with Director Liao Huai, who uses this function during bronchoscopy. We provide a prompt: the spot you think is a bronchus actually has a blood vessel underneath. Because some inexperienced doctors see a protrusion on a bronchus and may think a biopsy can be done, but that protrusion is actually part of a blood vessel; if a biopsy is performed, an artery might be punctured, which would be a major surgical risk.
Lei Feng Network: The most sensitive issues in medical AI are hallucination and misdiagnosis risks. Does CARES give doctors 'suggestions' or 'conclusions'? What has the team done in system design to address possible disagreements in judgment?
Liu Hongbin:In the current collaboration between CARES, hospitals, and doctors, it mostly plays the role of an assistant, providing suggestions. But AI will certainly sometimes hallucinate, and the first thing we need to do is minimize hallucinations as much as possible, especially for decision-type suggestions. For example, when a doctor is about to issue a diagnosis, the AI model might recommend a certain diagnostic approach; if an error occurs here, it could mislead the doctor, which is a fairly big risk.
Our current approach is: first, make the AI's entire reasoning strictly follow the reasoning logic of clinical experts—when making a diagnosis, clinical experts refer to guidelines and the latest clinical results from top international journals, and only make a conclusion or recommend a treatment plan after comprehensive judgment. We make the AI strictly follow this process, and turn the information the AI can call upon into structured data, like a knowledge graph, so the AI searches along these nodes for the basis on which to provide suggestions to doctors.
So the final suggestion given to the doctor will always come with its reasoning path: based on a certain recommendation in a certain guideline, or the latest findings in a certain academic paper, a diagnostic opinion is proposed, and then the clinician makes the judgment. In any case, the doctor is the final decision-maker, and this cannot be changed. The most important thing is to enable doctors to integrate the AI's judgment with their own thinking and give the final conclusion.
Lei Feng Network: From hospital pilots to truly entering clinical workflows, which is the hardest hurdle—technology, approval, or doctors' mindset? Which stage have you reached now?
Liu Hongbin:The challenge of medical AI is that it must land products, apply them, and generate sales revenue within a very complex system; it is not a single-point problem. The entire healthcare system has several major stakeholders: patients themselves, doctors, hospitals, and the medical insurance payment side—what these four parties care about is actually different. Only when a product's design can satisfy the demands of all four parties simultaneously can commercialization succeed.
For example, first, technological leadership must be ensured—major domestic hospitals now all want to be leading in technology, which is what the 'national team' advocates; at the same time, you must genuinely solve clinical problems so that doctors find it helpful; for hospitals, using your new technology must bring operational gains and improved efficiency; for medical insurance, your technology must not add burden—ideally, after using it, the overall treatment process saves money rather than adding an extra expense. When designing each product, all these aspects must be considered.
Lei Feng Network: You are now doing the medical-vertical CARES. In your view, which capabilities of general models can be directly transferred over, and which must be rebuilt from scratch in medical scenarios?
Liu Hongbin:Regarding general capabilities, the stronger the model's general capabilities, the better the performance after post-training in the vertical domain plus vertical-domain tools. The base capabilities of a general model are like a cornerstone: the stronger the general capabilities, the higher the capability ceiling of the vertical-domain model. That is the first point.
But in actual application, if you are only solving a scenario in a certain department of a certain hospital, the task content is actually quite localized. Moreover, strong general capability often means the model is too large, making deployment difficult and compute consumption high. So actual deployment is a balance: on one side is model capability, and on the other is whether the task really needs such high capability—or whether a smaller model, with tool design and post-processing, can achieve very good results.
Lei Feng Network: From DAMO Academy to the Chinese Academy of Sciences and then to Hong Kong, your research direction has moved step by step from general multimodality to the medical vertical. How did this journey come about? Was there a particular moment that made you certain you wanted to dive into healthcare?
Liu Hongbin:In fact, we have always been working in the AI healthcare direction, persisting on this path all along. For me, there were two events that exponentially increased my confidence in the application of AI in healthcare.
The first was the release of ChatGPT. Before that, AI models—such as CNN convolutional neural networks—could only count as part of a tool; engineers were needed to use them to build a piece of software, which doctors would then use, and the application was heavily limited by the engineers' abilities—in other words, you could not hand this thing directly to doctors. After ChatGPT appeared, AI could converse with people, becoming interactive, and the original problem of 'doctors not knowing how to use tools' could be solved. Looking at it now, in only about four years, doctors can master AI tools without a long learning curve, which gave us great confidence at the time.
The other was over the past year, when many of our AI systems began to be deployed in hospitals. The team communicated with collaborating physician teams to understand exactly how AI is used in their daily work. We found that every doctor is more or less using AI tools—perhaps not for diagnosis, but for processing literature and daily work, AI has become indispensable. At the same time, we found that as long as the data volume is large enough and the data quality high enough, the model's capability can exceed the average level of experts, which is a commonality across multiple real deployment projects.
So on individual tasks, AI will certainly do better than humans—just like having a car race a person, the car will surely win. Healthcare is evidence-based; as long as the evidence is sufficient and strong enough, people will trust it.
Lei Feng Network: How many hospitals does the team currently have substantive cooperation with? What is the most concentrated piece of improvement feedback from doctors?
Liu Hongbin:On the hospital side, it should be more than 30 now. The fastest-adopted is the laryngoscope diagnosis AI workstation; this one system alone has been deployed in more than 10 hospitals; the bronchoscopy system in more than 4; the ultrasound system also in more than 10. Some hospitals deploy several of our systems at the same time, so all told, more than 30 hospitals should be actually using them.
Most of the feedback is hoping operations can be easier, so the work can be finished without too much thinking—reducing work burden is what they like most. Another concentrated demand is that the model's output and interaction should be more intuitive, with information visible at a glance, and doctors should as far as possible not misunderstand what the AI's results mean. So much of our revision work is actually in human-computer interaction—what form to use to display the AI's results: charts, numbers, or heatmaps.
Lei Feng Network: If the CARES series succeeds five years from now, how will the operating room differ from today's?
Liu Hongbin:The operating room of the future will certainly be an intelligent space of human-machine collaboration.
First, many instruments in the operating room will be robotized, but this will not necessarily take the form of humanoid robots. For example: today's shadowless lamp is a passive instrument, which the doctor must drag and adjust by hand; in the future, the shadowless lamp itself will be a robotic arm that can move by itself, and the doctor will not need to adjust it by hand during the entire surgery—because grabbing by hand itself brings infection risk. The current action of nurses delivering instruments may also in the future be done by robotic arms.
Second, the personnel in the operating room. Healthcare is a rather slowly evolving process, so let's imagine 20 years from now: today a surgery requires a medical team of seven or eight people; in the future, there may be only two or three people in the operating room—the chief surgeon, assisted by one or two nurses or assistants, with the rest of the interaction handled by robotic equipment.
In addition, the hospital of the future may no longer take its current form of isolated departments, but rather be one large intelligent agent: there is a central computing platform, and each department acts like one of its neural branches, with everything happening in operating rooms, consultation rooms, and outpatient clinics aggregated in real time into this "brain."
Leiiphone: If you had to sum up in one sentence what your team has been busy with this past year, what would you say?
Liu Hongbin:We call it real-world research. Over the past few years we built many AI models and AI tools, and over the past year we focused on clinical deployment in various hospitals and collecting feedback from doctors, while also evaluating ourselves whether the models truly helped doctors—because before deployment everything was an assumption; you didn't know whether doctors actually found them useful or not.
What encouraged us greatly is that the vast majority of the models showed value in actual use. There were also some features we hadn't thought of at first, but clinical experts are very creative: you built feature A, and they would think of their pain point with feature B and ask whether A could do it—and we found that it indeed could be transferred over. That is how many tools, evolving together with clinical experts, took on forms that are useful to doctors.
This is an original article by Leiiphone; unauthorized reproduction is prohibited. For details, seethe reprint notice。