Google uses AI biomedical medical retinal fundus image to measure heart disease

For Google researchers, the eye is a sign of a person's health. Now, the tech giant is predicting people's blood pressure, age and smoking by analyzing human retinal photographs and studying them in depth. The preliminary results show that machines can predict heart disease through this information, thereby making effective preventive measures.

The study relies on convolution neural networks, a depth-learning algorithm that changes biologists ' analysis of images. Scientists are using this method to look for mutations in the genome and to predict changes in the layout of individual cells. Google's approach has been discussed in a paper published last August. It can be said that this is part of a new type of deep learning application that makes image processing easier and can even identify neglected biological phenomena.

"Previously, it was impossible to apply machine learning to some areas of biology," said Philip Nelson, director of the Google Research program at Mountain View in California. "Not only is it now, but it is even more pleasing to see what human beings have not discovered," he said. ”

Convolution neural network enables the computer to process images efficiently and comprehensively, and does not need to decompose images into multiple parts. Thanks to the progress of computing power and storage technology, this method has achieved initial success in about 2012 years. For example, Facebook uses this kind of depth learning technique to identify faces in photos.

But scientists are still trying to apply these networks to the biology sector, in part because of cultural differences between the various fields. "Let a group of intelligent biologists and a group of equally intelligent computer scientists work in the same room, they communicate in two different languages and ways of thinking," said Daphne Koller, chief computational officer at Calico, a biotech company owned by Google's parent company Alphabet. ”

Scientists must also determine what types of research can be done over the network, which must undergo a lot of image training to make predictions. When Google wanted to use in-depth learning to discover mutations in the genome, scientists had to convert the alphabet of DNA into images that the computer could recognize.

Then they trained the neural network on the DNA fragment, which was aligned with the reference genome, and its mutation was known. The end result is deepvariant (deep variation), a tool released in December that can detect small changes in DNA sequences. At least in the test, Deepvariant's performance is as good as the traditional tools.

Cell biologists from the Allen Institute in Seattle are using convolution neural networks to convert flat, gray images of cells captured under an optical microscope into 3D images, some of which are marked in color.

This eliminates the need for cell staining-a process that takes more time and more complex laboratory equipment, and can also damage cells. Last month, the team unveiled the details of an advanced technology that could use some data (such as the contours of cells) to predict more cellular morphology and location.

Anne Carpenter, director of the Harvard-MIT Institute of Broad Institute of MIT and Harvard, said:

"What you see now is that machine learning can accomplish the biological tasks associated with imaging, an unprecedented shift." ”

In 2015, her interdisciplinary team began using convolution neural networks to process cell images; now, Carpenter says, the neural network is processing about 15% of its research center's image data. In her view, this approach will become the main data-processing model for its research center in the next few years.

Some people think that using convolution neural network to analyze images may inadvertently reveal some subtle biological phenomena, which may lead to some problems that have never been considered before and will be excited. "The most interesting word in science is not ' found, '" Nelson said. ' But ' that's weird, what's going on. ’。 ”

Rick Horwitz, executive director of the Allen Institute, said the accidental discovery could drive the development of disease research. He argues that deep learning can reveal signs of cancer in individual cells, which could help researchers improve the classification of tumours. Conversely, there may be new hypotheses about the spread of cancer.

In addition, other biological machine learning experts have turned their sights on other areas, although convolution neural networks have been applied to image processing. "Images are important, but chemistry and molecular data are just as important," he said. ”

--------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------