Using Artificial Intelligence to Better Diagnose COVID-19 and Other Respiratory Illnesses
Diagnosing a disease is the first step to infection control, but in the early days of the COVID-19 pandemic, medical professionals struggled with this critical measure. Preliminary research indicated that early COVID-19 tests could produce “false negatives” up to 30 percent of the time. That means in any group of people, testing could miss up to one-third of cases.
Testing is key to controlling a pandemic because it helps identify those who may have or have been exposed to a disease in a timely fashion – and the lack of sufficient testing has been cited as reason for the devastating spread of a disease that has killed more than 500,000 and infected more than 28 million people in the U.S. alone.
But potentially life-saving testing advances aren’t just made by medical professionals. Today, they’re also made by data scientists. Researchers at CACI have invented new artificial intelligence (AI)/deep learning techniques to produce a model that diagnoses COVID-19 based on X-Ray images with nearly 97 percent accuracy on a popular academic dataset for COVID-19 detection.
This new technology, along with accepted public health measures, not only has the potential to accelerate the fight against COVID-19, but also against other respiratory illnesses, and perhaps the next pandemic.
This is just one example of many in which CACI is applying Artificial Intelligence – including Machine Learning, Deep Learning, Natural Language Processing and Computer Vision - across our customer base. CACI applies “AI at Scale” to solve enterprise and mission challenges that just a few years ago were thought to be unsolvable. To learn more about CACI’s AI, machine learning, and deep learning capabilities, contact [email protected].
Building the Model
CACI data scientists approached the issue of accurate COVID-19 diagnosis as an image classification problem. CACI has pioneered some of the most advanced techniques in imagery analysis using AI. For example, CACI’s FeatureTrace® uses deep learning to automate the identification of features such as roads, buildings, and lakes contained within satellite imagery.
The research team set up a competition-style image classification challenge, whereby the X-Ray image data was fed through four algorithms to determine and refine the one that performed the best. The algorithms classified each medical image as either displaying COVID-19, pneumonia, or as normal. The most successful algorithm, a “split-attention” residual network, can determine which image features are the most relevant and diagnose the images correctly with almost 97 percent accuracy.
Future Applications
The neural networks CACI used to classify medical scans have potential applications beyond the current pandemic. The same techniques could be used to diagnose other respiratory illnesses more accurately and efficiently. CACI data scientists need only fine-tune and train their existing model on new data to automatically diagnose the illness. CACI also uses many similar techniques to improve the speed and accuracy of imagery analysis, including full-motion video streams.
The neural networks CACI uses, the algorithm tuning, and the robustness of the company’s approach ensures that when applying AI, it’s done with AI assurance and AI ethics in mind. Models must be trusted, data must be protected, and solutions must be resilient for our U.S. Government enterprise and mission customers. It’s these building blocks that enable CACI to rapidly respond to challenges like COVID-19 and enable the company to scale AI solutions and meet other customer challenges.
