Google has spent the last few years focusing on developing some extremely impressive computer vision technology. You can see the effects in the way Pixel phone cameras can produce incredible images with years-old hardware and Google Photos can identify almost any object in your uploaded snapshots. Now, Google is leveraging that machine learning technology to build an “augmented reality” microscope with the Department of Defense (DoD) to help doctors diagnose cancer.
In most cancer diagnoses, someone will at some point look through a microscope at a biopsy sample to confirm the presence of cancerous cells. Even pathologists who have been doing this for years can get it wrong occasionally. With this AI project, Google says its aim is to lower the current rate of diagnostic errors, which hovers around 5 percent.
The DoD might not seem like the ideal partner for this, but the agency has a department called the Defense Innovation Unit (DIU). This division works to scale commercial technology and deploy it throughout the country. Usually, this focuses on national security challenges, but medicine isn’t out of bounds for the DIU.
Google envisions a system running on the open-source TensorFlow platform that would scan the samples along with the pathologist. The AI might never be good enough to take over completely from a human, and that’s not the goal. By providing a second set of eyes, even if they’re synthetic, Google believes it can cut down on the 12 million yearly misdiagnoses.
Pathologists can differentiate between cancerous and non-cancerous samples like these, and Google thinks AI can learn to do it, too. Credit: Richard M. Levenson, University of Iowa, CC4.0
Teaching the AI to spot cancer should be theoretically similar to teaching it to visually identify a tree or a dog. You have to feed the algorithm labeled data, making tweaks until it’s able to reliably recognize the important features. Then, it can do the same with new samples. The system will create an overlay that highlights potential cancer cells, helping pathologists make a more accurate determination.
The data used to build the AI came from public and private sources, all of which was de-identified to preserve patient privacy. Google also notes that all new patient data will remain under the control of the individual hospital or provider. The initial testing will take place at select Defense Health Agency treatment facilities and Veteran’s Affairs hospitals in the US. If the system performs well, it will expand to more US military health care facilities.
Top image credit: Konstantin Kolosov/Pixabay
Now read:
- Machine Learning AI Confirms 50 New Planets
- Some AI Advances in the Last Decade May Have Been Illusory
- Level Up: Nvidia’s GameGAN AI Creates Pac-Man Without an Underlying Game Engine
No comments:
Post a Comment