AI Spots Dead Cells 100 Times Faster Than Humans, Which Could Help Cure Alzheimer's
Understanding when and why a cell dies is fundamental to studying human development, disease and aging. For neurodegenerative diseases such as Lou Gehrig's disease, Alzheimer's disease and Parkinson's disease, identifying dead and dying neurons is crucial for developing and testing new treatments.
But identifying dead cells can be challenging. Until now, scientists had to manually mark under a microscope which cells looked alive and which were dead. Dead cells have a characteristic shriveled appearance that is relatively easy to recognize if you know what to look for.
Researchers hired an army of hourly-paid student interns to review thousands of images and count when each neuron in a sample died. Unfortunately, doing this by hand is a slow, expensive and sometimes error-prone process.
Making matters more complicated, scientists have recently started using automated microscopes to continuously capture images of cells as they change over time. While automated microscopes make imaging easier, they also generate a huge number of images to sort by hand.
It became clear that manual curation is neither accurate nor efficient. Moreover, most imaging methods can only detect late stages of cell death, sometimes days after a cell has already begun to decay. This makes it difficult to distinguish what actually caused a cell to die from factors that were merely involved in its breakdown.
Scientists have tried for some time to automate the curation process. Early attempts could not handle the wide range of cell types and microscopes used in research, and could not rival the accuracy of trained interns.
But a new artificial intelligence technology developed by the research team can identify dead cells with superhuman accuracy and speed. This progress could potentially accelerate all kinds of biomedical research, especially in the field of neurodegenerative diseases.
Recently, artificial intelligence has taken microscopy by storm. A form of AI called convolutional neural networks, or CNNs, is of particular interest because it can analyze images just as accurately as humans can.
Convolutional neural networks can be taught to recognize and detect complex patterns in images. As with human vision, giving a CNN many example images and indicating which features to focus on can teach a computer to recognize the patterns of interest.
These patterns can include biological phenomena that are hard to see with the naked eye. For example, one research group was able to train a CNN to identify skin cancer more accurately than trained dermatologists. More recently, scientists have trained CNNs to identify complex biological signatures, such as cell type, in microscopic images.
Building on this work, researchers at the Gladstone Institutes at the University of California, San Francisco developed a new technology called biomarker-optimized CNNs (BO-CNNs) to identify dead cells. First, they needed to teach BO-CNN to distinguish clearly dead cells from clearly live ones.
So they prepared a Petri dish with mouse neurons that were engineered to produce a non-toxic protein called the genetically encoded death indicator (GEDI), which stained live cells green and dead cells yellow. BO-CNN could easily learn that green means "alive" and yellow means "dead." But it also studied other features that distinguish live cells from dead ones that are not so obvious to the human eye.
After BO-CNN learned to identify the features that distinguish green cells from yellow ones, it was shown neurons that did not differ in color. BO-CNN was able to correctly label live and dead cells significantly faster and more accurately than people trained to do the same. The model could even look at images of cell types it had not seen before, taken with different types of microscopes, and still correctly identify dead cells.
One obvious question remained: why was the researchers' model so effective at detecting dead cells?
Scientists often treat decisions made by CNNs as black boxes, with the strategy a computer uses to solve a visual task considered less important than how well it performs. However, since there must be some patterns in cell structure that the model relies on when making decisions, uncovering these patterns can help scientists better define what cell death looks like and understand why it happens.
To find out what these patterns are, the team used additional computational tools to create visual representations of BO-CNN's decisions. It turned out that their BO-CNN model partially detects cell death by focusing on changes in fluorescence patterns in the cell nucleus. This is a feature human curators previously did not know about, and it may explain why the design of earlier AI models was less accurate than BO-CNN.
The researchers believe their approach represents a significant advance in using artificial intelligence to study complex biology, and this proof of concept could be widely applied beyond detecting cell death in microscopic images. Their software is open source and publicly available.
Live-cell microscopy is extremely information-rich, which researchers find difficult to interpret. But with technologies such as BO-CNN, they can now use signals from the cells themselves to train AI to recognize and interpret signals in other cells. By removing human guesswork, BO-CNNs improve the reproducibility and speed of research and can help scientists discover new phenomena in images that they otherwise might not easily recognize.
Using artificial intelligence, the research team is currently working to extend its BO-CNN technology to predict the future, identifying damaged cells before they even begin to die. They believe this could be a game changer in neurodegenerative disease research, helping to identify new ways to prevent neuron death and, ultimately, lead to more effective treatments.
