A Hong Kong-based research centre under the Chinese Academy of Sciences (CAS) has launched a new artificial intelligence (AI) tool to assist in complex brain surgery, even as the healthcare industry deals with an inadequate number of specialised databases for this procedure.
The Centre for Artificial Intelligence and Robotics (CAIR), the Hong Kong branch of mainland China’s national research institute, on Monday unveiled the CARES Copilot 1.0 AI model to help neurosurgeons provide “more efficient clinical diagnosis and to make better medical judgments based on sufficient references”, said Liu Hongbin, the centre’s executive director, in an interview with the South China Morning Post.
The CARES Copilot 1.0 system has already undergone internal testing across a number of hospitals in Hong Kong and the mainland. It has been deployed in doctors’ workflow to help them “prepare surgical plans and for post-surgery management”, Liu said.
During a live demonstration on Monday at the Hong Kong Science Park, Danny Chan Tat-ming, head of the neurosurgery division at the Department of Surgery in the Chinese University of Hong Kong, showed how the CARES Copilot 1.0 model is capable of generating key information from multiple academic papers, including citations, within seconds to ensure the accuracy of answers. Chan said the tool can achieve an accuracy rate of up to 95%.
The latest initiative by CAIR, which is co-funded by the Hong Kong government’s InnoHK research programme, reflects the efforts of state agencies on the mainland to develop a range of AI innovation alongside private technology companies, in the country’s bid to catch up with advances made by the likes of ChatGPT creator OpenAI.
CARES Copilot 1.0, which was based on Meta Platforms’ Llama 2 large language model (LLM), was trained on various multimodality databases – including text, images, magnetic resonance imaging (MRI), computed tomography (CT) scans and ultrasound imaging – tailored for the medical field. LLMs are the technology used to train ChatGPT and similar generative AI services.
The impact of US tech sanctions, however, led to the CARES Copilot 1.0 being trained on a computing infrastructure powered by Huawei Technologies’ Ascend AI processors, instead of one based on Nvidia’s much sought-after graphics processing units.
CAIR’s Liu said their research team has “put extra effort into adapting Huawei’s system”. They were optimistic that training on a computing platform based on Huawei’s AI chips would “improve faster” in the long run, he said.
Still, Liu said there are currently limitations across the broader healthcare industry in terms of the number of specialised databases for brain surgery procedures. He added that such databases between different hospitals are usually not linked, too.
“In the AI industry, it is common knowledge that the bigger the database sets are the more powerful a model could be trained,” he said.
To overcome that limitation, Liu pointed out that their research team has fed CARES Copilot 1.0 with thousands of medical textbooks, academic papers and international neurosurgical guidelines.
“Clinical decision-making is based on multiple information, such as CT scans, MRI, patients’ physiological testing results and more,” he said “So we want doctors to make better decisions through the combination of multiple sources of information.”
Founded in 2019, CAIR is the first CAS branch outside the mainland. It plays the role of a bridge to connect Hong Kong’s international talent and top academic research with the needs of well-established industries on the mainland, according to Liu.
Their research team is now approaching a number of medical device manufacturers to adapt CAIR’s AI innovation in their products. – South China Morning Post