Detecting pneumonia in chest x-rays using AI inferencing at the edge

Subh Bhattacharya Lead, Healthcare, Medical Devices & Sciences at Xilinx

The use of artificial intelligence (AI), specifically machine learning (ML), is fast becoming a transformational force in healthcare. Using AI and various image processing techniques within radiological modalities like X-rays, ultrasound and CT scans can lead to better diagnosis and better patient outcomes. Additionally, use of AI can lead to increased operational efficiencies, and significant cost reduction in healthcare procedures.

Chest X-rays used to detect respiratory diseases, like pneumonia, are the most used radiological procedure with over two billion scans performed worldwide every year, that’s 548,000 scans everyday. Such a huge quantity of scans imposes a heavy load on radiologists and taxes the efficiency of the workflow. Some studies show ML, Deep Neural Network (DNN) and Convolutional Neural Networks (CNN) methods can outperform radiologists in speed and accuracy, particularly under stressful conditions during a fast decision-making process where the human error rate could be quite high. Aiding the decision-making process with ML methods can improve the quality of the results, providing the radiologists and specialists an additional tool.

Healthcare companies are also now looking for effective point-of-care solutions to provide cost-effective and faster diagnosis and treatment in the field or in locations away from large hospitals. As a result, there’s rising demand to perform accurate image inferences to efficiently detect respiratory diseases like pneumonia in the field and provide clinical care using small, portable and point-ofcare devices at the edge.

Spline.ai (a partner of Xilinx) has developed a model using curated, labeled images for X-ray classification and disease detection (Figure 1). The model – trained using datasets from the National Institute of Health (NIH), Kaggle and from the likes of Stanford and MIT – can detect pneumonia with greater than 94% accuracy today. The model is then deployed and optimized on Xilinx’s Zynq® UltraScale+™ MPSoC running on the ZCU104 platform acting as an edge device. Xilinx’s Deep Learning Processing Unit (DPU), a soft-IP tensor accelerator that enables low inference latency of less than 10 microseconds. The model will be retrained periodically to improve the accuracy of its results.