Xilinx, AWS and Spline.AI collaborated on a deep-learning model for medical X-ray classification and disease detection which could be used in clinical settings for pneumonia and COVID-19 predictions.
The team relied on 30,000 curated and labeled pneumonia images and 500 COVID-19 images to train deep learning models, resulting in high accuracy and speed in predictions. This training data is being made available for public research and healthcare bodies including NIH, Stanford University and MIT, Xilinx said in a release on Tuesday.
The functional deep-learning model is being introduced along with a reference design kit available for $1,295. The Xilinx Zynq UltraScale+MPSoC is added to the ZCU104 FPGA which is provided as an edge device, according to a web page description. A Xilinx deep-learning processor unit is integrated into the MPSoC (Multiprocessor System on Chip) which accelerates the convolutional neural network within AWS IoT Greengrass.
Python PYNQ, an open source programming platform was employed allowing clinical researchers to adapt the board for different applications. Models can be developed with mobile, portable or point-of-care situations with the ability to scale up using the cloud. The AI model was trained using Amazon SageMaker and is deployed from cloud to edge via AWS IoT Greengrass, allowing remote learning model updates.
The combined technologies mean that highly accurate clinical diagnostics are possible with the use of low-cost medical appliances, AWS said. Physicians can upload X-ray images to the cloud without the need of a physical medical device, allowing physicians to reach remote locations for care, said Dirk Didascalou, vice president of IoT at AWS.
The system can be deployed for clinical use after verification and regulatory approvals. “It is up to a medical equipment OEM or a clinic or hospital to deploy,” said Subh Bhattacharya, lead on healthcare and medical devices and sciences for Xilinx, via email. “The reference design kit can be used to speed up time to market to develop and deploy other models with different radiological or clinical flow.”