On April 11, United Imaging Intelligence officially launched the uAI Shadow Intelligence Large Model Base, which is based on a million-level medical database and serves as a base model, which can bring all-dimensional assistance to the development of text, image, and mixed-modal products.
According to reports, through a large amount of data training, the uAI image intelligence model has the general underlying learning skills of medical imaging, and has the ability to quickly migrate to new disease types, which can make full use of the correlation between different modalities and different tasks, and can achieve excellent performance for all kinds of different tasks. Taking the renal artery vascular segmentation task as an example, after testing, the large model only needs 10 training data to achieve the performance level that the traditional small model needs to use 201 data to achieve. In addition, United Imaging Intelligence introduced that the uAI Shadow Intelligence large model is better than the traditional small model in terms of segmentation accuracy on the other 8 organs. This means that the large model base can not only improve product performance, but also provide acceleration for the implementation of new product development, promote model development into a new paradigm, and can meet more multi-dimensional clinical needs more flexibly and quickly. At the beginning of the design, it was proposed that "the training should be large and the deployment should be small", emphasizing that in the actual deployment process, the lightweight model can be split to ensure the accessibility of the model in the hospital and improve the flexibility of the model application. At present, based on the base of this large model, United Imaging Intelligence has achieved innovative products in three aspects: text, image, and hybrid modality. Shen Dinggang, co-CEO of United Imaging Intelligence, told the media that the medical text large language model developed by United Imaging Intelligence supports 70B text input, which is not "large" compared with the long text model that is currently being "rolled" in China, but the small model has the advantage of a small model. The 70B Parameter Text Model is a tool optimized for the medical field, integrating knowledge from a large number of medical textbooks and clinical databases, and achieving a passing standard of 67% on the USMLE (American Medical Licensing Examination), demonstrating its potential for application in the medical field. At present, the text model has been used to assist clinical and scientific research in the Department of Cardiac Surgery and Respiratory Medicine of Zhongshan Hospital Affiliated to Fudan University. Co-founder of United Imaging Intelligence, Zhan Yiqiang, chief operating officer, said that the medical text model developed by United Imaging Intelligence after professional training does not require huge computing power, can be deployed locally in hospitals, and has good information summary performance, which can quickly capture the key information of patient admission records up to several thousand words and import them into the first disease course record, which can become an auxiliary disease tool for doctors to write diagnosis and treatment records。
Source | Big health pie
Join the community
Dots to share
Point Collection