How the AI Behind Deep Fakes Could Improve Cancer Diagnoses

Not that that negates the damage deep fakes can do to medical science, interpersonal communications - and democracy. JL


Karen Hao reports in MIT Technology Review:

Deep-learning algorithms are excellent at pattern-matching in images; they can be trained to detect different types of cancer in a CT scan, differentiate diseases in MRIs, and identify abnormalities in an x-ray. But because of privacy concerns, researchers often don’t have enough training data. This is where GANs come in: they can synthesize more medical images that are indistinguishable from the real ones, effectively multiplying a data set to the necessary quantity.

Generative adversarial networks, the algorithms responsible for deepfakes, have developed a bit of a bad rap of late. But their ability to synthesize highly realistic images could also have important benefits for medical diagnosis.

Deep-learning algorithms are excellent at pattern-matching in images; they can be trained to detect different types of cancer in a CT scan, differentiate diseases in MRIs, and identify abnormalities in an x-ray. But because of privacy concerns, researchers often don’t have enough training data. This is where GANs come in: they can synthesize more medical images that are indistinguishable from the real ones, effectively multiplying a data set to the necessary quantity.

There is another challenge, though. Deep-learning algorithms need to train on high-resolution images to produce the best predictions, yet synthesizing such high-res images, especially in 3D, takes a lot of computational power. That means it requires special and expensive hardware, making its large-scale use impractical in hospitals.

So researchers from the Institute of Medical Informatics at the University of Lübeck proposed a new approach to make the process much less intensive. They broke it up into stages: the GAN first generates the whole image in low-res, then generates the details at the right resolution one small section at a time. Through experiments, the researchers showed not only that their method generated realistic high-res 2D and 3D images with low computational resources, but that the expenditure also stayed constant regardless of image size.