NVIDIA-Powered Imaging AI Platform Enables Traditional Healthcare IT Firm to Accelerate Development of Practical Radiology AI Models

Beijing Smart Medical Technology Co., Ltd. is a developer and service provider of end-to-end imaging Clinical Decision Support Systems (CDSS).

In the collaborative development of sequential imaging AI with the Department of Radiology at Peking University First Hospital, the utilization of NVIDIA GPUs, open-source algorithms, the NVIDIA Clara platform, and advanced AI training platforms and courses has empowered radiologists to independently develop complex imaging diagnostic applications at scale. To date, more than 20 AI models have been developed, and their seamless integration with structured reports has been preliminarily validated.

This case primarily utilizes the Tesla V100 GPU and NVIDIA’s open-source algorithm libraries, drawing reference from the Clara platform.

Beijing SmartRay Medical Technology Co., Ltd. (hereinafter referred to as “SmartRay”), established in 2016, is a developer and service provider of full-process Clinical Decision Support Systems (CDSS) for imaging. Its imaging decision support system covers four key areas of the radiological examination workflow: a knowledge base for clinical imaging orders, a scanning knowledge base for radiology departments, post-processing and the currently prominent field of imaging AI, and a back-end diagnostic knowledge base. In collaboration with the Department of Radiology at Peking University First Hospital, SmartRay has developed a series of sequential imaging AI solutions that can automatically populate measurements and key images into more than 20 types of structured reports, significantly enhancing diagnostic efficiency.

Today, the Department of Medical Imaging faces simultaneous challenges of surging data volumes and insufficient human resource supply. During the image diagnosis process, physicians need to obtain various measurements and key images for analysis by interpreting scans, such as measuring the diameters of masses or obtaining comprehensive tumor burden assessment data. This information is critical for accurate diagnosis; however, in many current application scenarios, there are no available post-processing systems or AI-powered imaging tools to assist physicians in automating these measurements. Faced with massive datasets, relying entirely on manual measurements by physicians would entail an unimaginably heavy workload. There is a strong demand for the development of specialized AI applications with clearly defined purposes.

In medical imaging, structured reporting represents one of the key trends in the development of diagnostic imaging services. The tags within structured reports constitute a core data element for the continuous improvement of scientific research, teaching, and clinical practice guidelines, as well as for other front-end clinical decision support (CDS) systems, including artificial intelligence (AI) applications in imaging. The primary challenge in promoting structured reporting in radiology lies in enhancing the efficiency of report completion.

Therefore, how to apply AI technology to automatically generate structured imaging reports, addressing the surging diagnostic demands and personalized diagnostic requirements in the field of medical imaging, is the core issue that imaging decision support systems need to resolve.

To enable automated measurements in imaging diagnosis and generate structured reports, Semaitui, leveraging NVIDIA Clara, has proposed a decomposition scheme for an AI sequential model based on the dialectical definition of single-disease imaging structured reports.

NVIDIA Clara supports the integration of various external models, functioning as a shared platform rather than a proprietary one for any single enterprise. Furthermore, the AI-Assisted Annotation feature on the Clara Annotation Server leverages existing AI models to pre-annotate data, which annotators can then modify, significantly reducing the labor intensity associated with new annotation tasks.

By drawing on these two concepts from the NVIDIA Clara platform, Semaitrix clearly defines the data dimensions and reasoning logic required for diagnosis in its structured imaging reports for specific diseases. Even when physicians manually select options, clear, standardized, and compliant diagnostic reports are automatically generated. For imaging features that need to be extracted from images, AI automatically populates the relevant fields, thereby reducing physicians’ workload.

In actual business operations, Sematrix organizes these data by anatomical site and feature type, integrating them into several independent AI models. These trained AI models are then sequentially combined to support a single diagnostic report. Data generated by these models can be automatically incorporated into the report, thereby enabling high-quality, automated data acquisition and diagnostic reasoning, ultimately forming an automated diagnosis and treatment plan.

Leveraging NVIDIA V100 GPU hardware and NVIDIA’s open-source algorithms, and drawing on the excellent concepts of Clara, this approach utilizes various existing tools for annotation and model training, significantly lowering the barrier to entry for AI development. Training with automatic mixed precision has increased training speed by 50%.

During the collaboration with the Department of Radiology at Peking University First Hospital, the majority of medical professionals lacked a professional background in AI development. However, through NVIDIA’s online training courses and on-site development workshops organized at the RSNA annual meeting, interested physicians were able to participate in AI technology training. This enabled healthcare workers, even those without a technical background in AI, to quickly get up to speed and engage in the development of AI models.

Currently, Semaiter has collaborated with the Department of Radiology at Peking University First Hospital to develop a suite of sequential AI application models that differ significantly in style from current mainstream AI applications. These models are gradually enabling the automated population of measurement values and key images across more than 20 types of structured reports, thereby substantially improving diagnostic workflow efficiency.

Leveraging the NVIDIA GPU computing platform, the Clara platform, open-source algorithms, and specialized, systematic AI training, radiologists in medical institutions can engage in AI-driven innovation with a low barrier to entry. Without the need to learn algorithms or write code, they can delegate tedious, complex, and repetitive tasks to machines, allowing healthcare professionals to focus their efforts on diagnostic processes and strategic analysis, thereby significantly enhancing diagnostic efficiency.

The incremental growth of medical imaging diagnostic services and the integration of new knowledge inevitably rely on Imaging Clinical Decision Support Systems (CDSS). A decision support system integrates physicians’ expertise with traditional workflows. Engaging radiologists in the development of knowledge application and knowledge mining is crucial for accelerating industry advancement and enhancing the operational and managerial capabilities of healthcare institutions.

Yue Xin, Founder and CEO of Smart Medical, stated, “No single AI company is powerful enough to address all imaging AI challenges, nor can any healthcare institution achieve this independently. Therefore, collaborative innovation and shared outcomes represent the most rational path forward. Leveraging the NVIDIA Clara platform and its open philosophy, Smart Medical shares various AI models and report templates with a wide range of healthcare institutions, accelerating the deployment of practical intelligent reporting solutions.”