Scientists have developed a medical vision model called HealthGPT, which achieves state-of-the-art results in multimodal understanding and generation tasks. This innovative model has several key features that make it highly versatile and efficient in different medical imaging scenarios. For high-efficiency training and iterative updates, HealthGPT employs a technique known as High-efficiency LoRA (H-LoRA). This method uses a parameter-efficient fine-tuning approach, allowing the model to be quickly updated with new medical data, such as diagnostic images from hospitals. Additionally, the model's lightweight architecture makes it suitable for deployment on endpoint devices like smartphones, enhancing its accessibility and utility in real-world applications. In terms of multimodal medical imaging support, HealthGPT can assist physicians in diagnostics, knowledge queries, and patient consultations. At the patient end, the model can facilitate basic disease consultations, potentially reducing healthcare costs by minimizing the need for in-person visits. According to the research team, they have already collaborated with the Second Affiliated Hospital and Yuyao Hospital of Zhejiang University School of Medicine to implement sections of the model for practical medical use, particularly in the image domains of CT scans and nuclear magnetic resonance images, where it aids radiologists in interpreting medical scans. The team hopes to expand the application of HealthGPT to include rare disease diagnosis in the future. They are currently exploring this potential in collaboration with institutions like the Second Affiliated Hospital and other research organizations within Zhejiang University School of Medicine, conducting initial trials to assess feasibility. Looking ahead, the research group plans to continue their exploration in two primary directions. First, they aim to develop larger-scale versions of the HealthGPT model. The current model has parameters ranging from 38 billion to 140 billion, and the team is working on expanding the model's architecture and parameter size to further enhance its capabilities. Second, they are researching the integration of a healthcare agent system. As OpenAI’s CEO Sam Altman noted, model collaboration is the future trend. In practice, many medical conditions do not require large models to address, and smaller models can complement each other effectively. This integrative approach is expected to help create a more sophisticated and multifaceted intelligent system that can provide higher-quality medical services to both healthcare professionals and patients. Zhang Wenhong, head of the research team, stated, “We plan to upgrade our single model to a Health Agent, enabling different roles (such as physicians and patients) to participate in data updates and parameter optimization. Our goal is to build a more advanced and well-rounded intelligent system that will assist medical personnel and patients by delivering higher-quality healthcare services.” References: 1. https://arxiv.org/pdf/2502.09838 2. https://github.com/DCDmllm/HealthGPT Editor/Arrangement: He Longxiang