Confirming a diagnosis of Alzheimer’s disease requires an expensive PET scan that uses a high dose of full-body radiation. With seed grant support from the Stanford Institute for Human-Centered Artificial Intelligence (HAI), a group of Stanford researchers can now diagnose Alzheimer’s Disease just as successfully by applying artificial intelligence (AI) to low-dose PET scans and simultaneously acquired MRI images. “This work has the advantage for the patients of being safer, lower dose, faster, cheaper all the things you’d want as a patient,” said Greg Zaharchuk, professor of radiology at Stanford University and 2018 HAI seed grantee. Using artificial intelligence, Zaharchuk’s team has become adept at what’s called image transformation. They can take one image or set of images and use a type of AI called a convolutional neural network (CNN) to produce a new set of images as the output. “If the information you want exists in the images you have acquired, then you can train a classifier using a CNN,” Zaharchuk said.
Machine learning approaches like CNNs typically feed a computer a labeled set of data that trains the computer to recognize something in the data. In the case of image transformation work, where the goal is to produce a better image, the image is the label, Zaharchuck says. “Every pixel is the answer I want to predict.” Think, for example, of a grainy image on a black and white TV, said Kevin Chen, a postdoctoral student in Stanford’s radiology department who worked on the low-dose PET project. If a neural net is trained on grainy and crisp images of the same object, it can learn how to output crisp images when given grainy images even without their crisp counterparts. “The human visual system is great for tracking a tiger on the Serengeti,” Zaharchuk said, “but it wasn’t built to see different contrasts like this.” A neural net, by contrast, is agnostic to the challenges of interpreting subtle contrasts. “If there is information in an image, a neural net can efficiently find it,” Zaharchuk said.