A new machine-learning strategy can distinguish between healthy and degraded red blood cells, potentially offering a faster and better way to identify high-quality blood for transfusions.
Deep learning model assesses quality of stored blood
Over time, blood stored in banks undergoes changes that impair their health and ability to carry oxygen. These changes — collectively called “storage lesions” — render the blood unusable in treatments like life-saving transfusions. Experts currently search for signs of degradation in blood samples under a microscope, but this time-consuming task can be destructive to fragile cells and subject to human bias.
Now a new imaging and machine learning method could offer a less laborious, faster, and more accurate way to assess the viability of stored blood, thanks to a collaborative team of scientists from the Broad Institute of MIT and Harvard and Ryerson University in Toronto, Canada as well as contributors in Switzerland, Germany, and the United Kingdom.
As the team reports in the Proceedings of the National Academy of Sciences, the method rapidly identifies differences between new and old red blood cells that are being stored for transfusions, offering an automated and effective way to monitor blood for signs of storage lesion. The work lays the groundwork for a new way for blood banks to assess blood quality, as well as a more general machine learning strategy that could potentially improve the diagnosis and monitoring of hematological diseases.
“It's always a struggle to maintain the blood supply globally, especially during a pandemic,” said Anne Carpenter, a senior author on the paper and senior director of the Broad Institute’s Imaging Platform. “What this study demonstrates is the power of deep learning to not only replace the manual work of classifying red blood cells, but to provide a dramatically more accurate assessment of blood quality.”
Signs of bad blood
The telltale sign of unusable blood lies in the shape of its red blood cells. After about six to eight weeks in storage, healthy red blood cells begin to transform from smooth, concave discs to lumpy, round balls and then to swollen, dense spheres. To gauge the overall health of a bag of blood, researchers examine red blood cells under a microscope and manually rank each cell on a scale of one to six based on its shape. They then compile those rankings into a metric called a morphology index. The process can be slow and prone to human error and subjectivity.
“Researchers looking at the same cells can, and oftentimes do, disagree about which category they fall into,” said Minh Doan, first author of the paper and former postdoctoral associate in the Carpenter Lab. Doan now works as Head of Bioimaging Analytics at GlaxoSmithKline.
Doan and his collaborators at the Broad wondered if there was an automated, more objective solution. They hypothesized a method involving microfluidic imaging flow cytometry — a technique that captures snapshots of hundreds of thousands of single cells within minutes — paired with deep-learning algorithms that automatically analyze the images for cell quality. Doan found multiple teams from around the world producing suitable images of blood. He emailed them asking to collaborate, and they agreed.
Together, the international team of researchers created the world’s largest freely available set of annotated red blood cell images, comprising more than 67,400 cells, each one annotated by hand.
The researchers used such images to train a deep learning model to classify healthy and unhealthy morphologies, and tested it against annotations from experts from two different countries. The model agreed with the experts roughly 77 percent of the time, an improvement from the 14 percent accuracy of random guessing and close to the 82 percent of the time experts agree with each other.
Better blood analysis
Because experts themselves don’t always agree with each other in their blood cell assessments, the researchers set out to improve the model’s performance by eliminating the human component of annotation entirely. They developed an alternate, weakly supervised deep-learning strategy that could automatically extract biologically relevant morphological differences in red blood cells directly from images that haven’t been annotated by people.
The scientists used microfluidic imaging flow cytometry to take thousands of pictures of cells from blood bags at specific time points during storage. They then trained their deep-learning algorithms on these images, allowing the model to categorize the images, from healthy cells to degraded ones.
The team’s model revealed a continuum of morphological changes that can be used to objectively predict blood quality, eliminating the need for human annotation. The system also generated cell quality scores that better matched the results from a biochemical assay of red blood cell quality than those from human examination.
If validated and adopted in clinical practice, the approach could improve the accuracy, speed, and cost of assessing stored blood for transfusion, which is in short supply in many areas of the world.
“It’s really exciting because we demonstrated that deep learning combined with imaging flow can reconstruct a biological progression without human assessment,” said Doan. “That has profound potential for many future applications that require monitoring of diseases.”
The researchers say a study this extensive wouldn’t have been possible without their many international partners, who provided imaging data, blood samples, annotations, analysis, and most importantly, a willingness to collaborate after they received Doan’s initial email.
“The fact that some people are so willing to share data to move the science along faster is a really important aspect of modern science, which I think is just so helpful in translating science into new technologies that could potentially be useful in the clinic,” Carpenter said.
Support for this research was provided in part by the US National Science Foundation, UK Biotechnology and Biological Sciences Research Council, Natural Sciences and Engineering Research Council of Canada, Canadian Institutes of Health Research, and Carigest S.A. of Geneva, Switzerland.