OR WAIT null SECS
Because a fifth of all microbiology technologists will reach retirement by 2022, laboratories will need new tools to multiply their labor.
Bloodstream infections (BSI) can be subtle, deadly, and difficult to identify. To make matters worse, a fifth of all microbiology technologists trained to spot them will reach retirement age in the next 5 years.
Those 2 facts create a need for innovation in the space, and researchers from Beth Israel Deaconess Medical Center (BIDMC) think an artificial intelligence (AI)-enabled microscope might fit the bill. In a study that was recently accepted by the Journal of Clinical Microbiology, they tested out how well the technology worked.
Samples from patients with suspected BSIs were incubated (to increase the presence of any bacteria) and then stained to create slides, which were fed to the microscope by the 10s of thousands in order to train it. The microscope made use of convolutional neural networking (CNN) to categorize the bacteria samples: In all, it was shown over 100,000 clinically-validated images, just 146 by 146 pixels each, of the most common infections (E. coli, Staphylococcus, and Streptococcus).
“Excessive background increases the chance that a CNN will learn features during training that are unrelated to bacterial Gram stain classification,” the researchers wrote. “This…results in a model with high accuracy in classifying images on which it was trained (the training set) but poor accuracy when presented with an independent validation set.” Using a custom Python script, the team developed a system in which the trainer was able to click areas of the image that contained potential bacteria, teaching the model to ignore unnecessary information.
Because each of the 3 bacteria come in a unique shape (rod-shaped, chains or pairs of spheres, and clusters of spheres) the system was able to visually discern what it was presented with 94.9% accuracy in a test set.
After training, the system was given 189 slides that had yet to be analyzed by human microbiologists. Across the board it identified the bacteria with over 93% accuracy. It was most successful at spotting Streptococcus, identifying chains of the bacteria 98.4% of the time.
“Taken together, our data support proof-of-concept for a fully automated classification methodology for blood culture Gram stains,” the authors wrote. “The algorithm was highly adept at identifying image crops with organisms and could be used to present prescreened, classified crops to technologists to accelerate smear review,” and it could someday be used to for all Gram stain interpretive activities in a laboratory, they added.
On top of all of the microbiology technologists expected retire in coming years, 9% of such roles are already unfilled. There may not be an answer to those personnel shortages, but automated analysis like this could be a force multiplier for the experts already in the field. According to senior author James Kirby, MD, AI-enabled microscopes could “conceivably reducing technologist read time from minutes to seconds.”
Kirby, the Director of the Clinical Microbiology Laboratory at BIDMC and an associate professor of pathology at Harvard Medical School, sees potential for the technology beyond diagnosis.
“The tool becomes a living data repository as we use it,” he said. “[It] could be used to train new staff and ensure competency. It can provide an unprecedented level of detail as a research tool.”