Microsoft AI published a study in Science showing its AI diagnostic system MAI-DxO achieved 83% accuracy on complex case simulations, far exceeding generalist physicians' 36%. Microsoft AI CEO Mustafa Suleyman called it a step toward “medical superintelligence.” The system coordinates multiple AI agents, and tests with models from OpenAI, Google and others found the best results came from GPT-5 and GPT-5.5.
Siemens Healthineers and NVIDIA announced a partnership to develop next-generation adaptive radiation therapy, integrating agentic and robotic intelligence, accelerated computing, and simulation. Initial directions include agentic AI in treatment planning, patient digital twin simulation, and real-time adaptive computing.
At a forum, He Chao, president of Shanghai MicroPort Medical Robotics, said remote surgery is a core track for domestic robots, and introduced the team's second-generation remote surgery system, the world's first remote surgery industry standard and registration certificate, as well as an autonomous surgery large model developed with Peking University, Fudan University and NVIDIA that completed the world's first animal trial with over 95% success rate.
Alec Tarashansky, Anthropic technical team member and Claude Science development lead, announced he is joining AI healthcare company OpenEvidence. He holds a PhD in bioengineering from Stanford and worked at the Chan Zuckerberg Initiative for three years; he brought the Claude Science prototype to Anthropic and helped develop it for about seven months.
During the 5th Global Digital Trade Expo, DIAGENS hosted a Medical AI Industry Festival in Hangzhou, where Song Ning and Kevin Kelly discussed the future of medical AI. Song compared the shift to the move from printing shops to Office software, introducing DIAGENS' iMedImage foundation model that allows radiologists to train their own AI assistants. Both agreed AI will not replace doctors, as human empathy and judgment remain irreplaceable. Kelly also addressed AI-era education reform, saying companies will become schools and lifelong learning is key.
Neusoft Medical's self-developed NeuSpace multimodal intelligent imaging post-processing platform and its CT-focused AI software NeuSpace.CT have both received NMPA market approval. The platform supports hospital deployment and multi-terminal access through intelligent imaging output, second-level image analysis, quality control and risk alerts, and full-cycle patient services; NeuSpace.CT is powered by a multimodal large-model engine covering 44 diseases, 78 applications and 6 imaging modalities, with a built-in 'Lingyu' bio-cognitive compression engine that can compress 200GB of data to 20GB.
Yuan Huishu said that Peking University Third Hospital and United Imaging Intelligence have officially released the world's first multi-disease MR spine "one-scan multi-report" AI agent, now deployed across the hospital's main campus and four directly affiliated branches. The agent uses large models to generate structured reports covering 76 diseases from a single scan, improving reporting efficiency by 56%. The project was launched in July 2025 and took 18 months of underlying redevelopment.
Beijing Jingyi Medical, in collaboration with BAIC Group and ecosystem partners, officially launched the world's first integrated AI-powered mobile medical vehicle, integrating AI fundus camera from Beijing Tongren Hospital, cardiovascular ultrasound AI from Beijing Anzhen Hospital, and AI pediatrician from Beijing Children's Hospital to support routine diagnosis and treatment at grassroots level.
Two medical AI achievements incubated by the National AI Application Pilot Base (Medical) in Beijing — Beijing Tongren Hospital's AI fundus camera and Beijing Children's Hospital's AI pediatrician — have been put into clinical use at the China-Guinea Friendship Hospital by the 32nd Chinese medical team to Guinea, with localization completed and positive patient feedback.
The AI healthcare industry reaches a critical turning point as iFlytek Medical launches Spark Medical Large Model V3.5 on June 9, trained on domestic computing