Scientists in Boston have developed an automated artificial intelligence (AI)-guided microscopy system that can help diagnose serious bloodstream infections (BSIs) quickly and accurately. The technology, which uses a trained convolutional neural network (CNN) to recognize the different shapes and distribution of pathogenic bacteria, could help to speed diagnosis and potentially save patient lives, as well as address the current lack of trained microbiology technologists, suggest its developers at Harvard Medical School and Beth Israel Deaconess Medical Center (BIDMC).
“This marks the first demonstration of machine learning in the diagnostic area,” comments James Kirby, M.D., director of the Clinical Microbiology Laboratory at BIDMC, and associate professor of pathology at Harvard Medical School. “With further development, we believe this technology could form the basis of a future diagnostic platform that augments the capabilities of clinical laboratories, ultimately speeding the delivery of patient care.” The researchers report on the technology in the Journal of Clinical Investigation, in a paper entitled “Automated Interpretation of Blood Culture Gram Stains using a Deep Convolutional Neural Network.”
The mortality rate from BSIs can be up to 40%, and every day’s delay in starting appropriate treatment increases the risk of death, so fast identification of the causative bacteria is critical. Initial diagnosis is based on the Gram stain smear, but human interpretation of Gram-stained slide images is “labor and time intensive, and highly operator-dependent,” the authors explain. “With consolidation of hospital systems, increasing workloads, and potential unavailability of highly trained microbiologists on site, automated image collection paired with computational interpretation of Gram stains to augment and complement manual testing would provide benefit.”
Dr. Kirby’s team now reports on the development of a trained CNN-based model that is designed to overcome some of the current technical difficulties associated with automating Gram stain analysis. The approach combines an automated slide imaging platform equipped with a 40x air objective and a trained CNN-based model that can recognize bacterial morphology.
“Like a child, the system needed training,” said Kirby. “Learning to recognize bacteria required a lot of practice, making mistakes and learning from those errors.” The researchers first generated 100,000 crops from more than 25,000 images obtained from positive blood-culture Gram stains on slides prepared during routine clinical workup. They used these crops to train the CNN to recognize pathogenic bacteria, based on their shape and distribution. These included rod-shaped bacterial including Escherichia coli, round clusters of Staphylococcus species, and pairs or chains of Streptococcus species. The trained system demonstrated an accuracy of nearly 95%.
CNNs do not interpret raw images directly, the authors explain. “Rather, they consist of a number of layers, each of which convolutes regions of the image to detect specific features. During each step of the learning process, a subset of images is presented to the network, allowing function parameters to be changed such that the CNN identifies features important for classification based on optimization of output accuracy.”
After further rounds of validation, the researchers ultimately tested the model on a set of 189 whole slides that hadn’t been part of training, validation, or test sets. “Overall, we found that our trained model performed well on whole-slide image classification,” they write. “Where cells were detected, we achieved overall classification accuracy of 92.5% specificity of >93% for all classification labels with no human intervention.”
Human technologists provide highly accurate diagnosis and training, but the American Society for Clinical Pathology estimates that in past years about 9% of trained microbiologist jobs in the U.S. were not filled, and nearly 20% of existing personnel plan to retire within 5 years. The Boston team projects that their AI platform technology could help to reduce the burden on existing staff by speeding classification by microbiology technologists and reducing read time “from minutes to seconds.” And with further training and development, the platform could one day be used as a fully automated classification system.
“In the era of laboratory consolidation and limitations in the number of skilled technologists, we believe our system could provide enhanced opportunities for rapid Gram stain classification at the site of care or during understaffed shifts in conjunction with later analysis at a central laboratory or day shifts,” the authors write. “We believe that this technology could form the basis of a future diagnostic platform that provides automated smear classification results and augments capabilities of clinical laboratories.This concept could potentially be extended to all Gram stain interpretive activities in the clinical laboratory.”
The system could also be used for microbiology training and research. “The tool becomes a living data repository as we use it,” he said. “And could be used to train new staff and ensure competency. It can provide unprecedented level of detail as a research tool,” Kirby suggests.