Xilinx in collaboration with Spline.AI announced a medical X-ray classification model that promises to improve COVID-19 and pneumonia prediction rates using artificial intelligence (AI) at the edge.

The companies used more than 30,000 verified pneumonia and 500 Covid-19 X-ray images collected by public health and research facilities, including the National Institute of Health, Stanford University, and MIT, to train the model using Amazon Web Services (AWS) SageMaker. From the cloud, the open source model can be deployed to edge devices — such as Xilinx's FPGAs — using AWS IoT Greengrass.

According to Subh Bhattacharya, head of Xilinx's healthcare, medical devices, and sciences division, the model will have an immediate impact on radiologists working in the field.

“Chest X-rays for respiratory disease detection is the most run scan with over half a million scans a day — with two billion a year — and radiologists are stressed and cannot provide enough time to investigate each scan in detail," he explained, in an email response to questions. "Studies have shown a significant number of missed or error in diagnosis [due] to human fatigue — up to 30%. The trained model developed for pneumonia already achieves 94% accuracy and can be improved with bigger datasets and additional training.”

As part of the announcement, the companies are making a reference kit, complete with a Xilinx Zynq UltraScale+ FPGA, available to researchers to speed the development of radiology flows for testing and predicting diseases like COVID-19 in patients.

The FPGA functions as a tensor accelerator, which is capable of running a variety of neural networks, including those used for the classification and detection of diseases. This means that the actual prediction can take place at the edge in multiple geographically disparate locations, and the model can be updated remotely.

However, because the model was developed on the Python programming language, the model can be adapted by researchers to suit a variety of operating environments requirements including mobile, portable, and point-of-care devices.

"As the model can be easily adapted to similar clinical and diagnostic applications, medical equipment makers and health care providers are empowered to swiftly develop future clinical and radiological applications using the reference design kit,” said Kapil Shankar, VP of marketing for Xilinx core markets group, in a statement.

Last week, The Wall Street Journal reported that AMD could buy Xilinx in a deal worth $30 billion. The report claims the deal could happen as early as this week.

The deal would add FGPAs to AMD's line up and allow the chipmaker to more directly compete with Intel — the only other large FPGA manufacturer — in the data center and at the edge.

Xilinx's FPGAs enable a variety of workloads including SDN, virtualized switching, NFV, and AI inference to be offloaded, freeing up the CPU for other applications.