AI used to speed up detection of cyclospora in CPS project
Scott Lenaghan, Ph.D., with the University of Tennessee, is enlisting artificial intelligence and machine learning to speed detection of Cyclospora cayetanensis’ infectious life stage.
“Right now, the only way to know whether it’s viable is a sporulation assay,” he said. “You have to determine that they have sporulated, and the whole process is labor intensive.”
The eventual high-throughput automated system also should significantly increase the number of potential Cyclospora inactivation methods for which researchers are able to screen. As part of the two-year project, Lenaghan and his colleagues plan to validate four inactivation strategies: gamma radiation, ultraviolet light, ozonation and chlorine dioxide gas. In addition, they plan to screen numerous antimicrobials and identify at least two novel inactivation methods.
“Right now, there are no inactivating strategies in the produce environment,” Lenaghan said. “The industry needs some guidance or direction. At least we can give them the data.”
Joining Lenaghan are co-principal investigators Qixin Zhong, Ph.D., and Mark Morgan, Ph.D., both with the University of Tennessee.
Zhong has developed a library of GRAS, or generally recognized as safe, compounds that will be screened for their potential to inactivate Cyclospora. Morgan brings expertise in process engineering with a focus on ozonation and chlorine dioxide.
Cyclospora has a complex life cycle and requires a human host as an intermediary to complete it. An infected human sheds unsporulated — or immature, non-infective — oocysts in their feces. It takes one to two weeks of favorable conditions outside the host for the oocysts to mature, sporulate and become infective to a human who consumes them.
In the past, Lenaghan said, scientists would incubate a sample for several days and manually look through a microscope to determine whether Cyclospora oocysts were viable and had sporulated.
Cyclospora is endemic to many parts of the world, including Central and South America. Until 2018, outbreaks in the United States were linked to produce imported from those endemic countries. More recently, a few Cyclospora outbreaks have been tied to U.S.-grown produce.
As part of their initial proof-of-concept trials, Lenaghan said they will use Eimeria oocysts as surrogates for Cyclospora oocysts. Both pathogens are protozoan and have very similar life cycles and similar appearances. But Eimeria oocysts, which are found in poultry droppings, are more readily available than Cyclospora oocysts.
The researchers have enlisted machine learning and artificial intelligence that involves a computerized system to compare microorganisms in a sample to an image library of confirmed pathogens. The more samples that are run, the more the system “learns” how to identify the target organism. The researchers also will use robotics to automate the process.
The first step involved running samples through the high-throughput system to determine how accurate it was in identifying oocysts. Lenaghan said they have reached a 95% confidence level.
Currently, the researchers are working to train the system to differentiate between sporulated and non-sporulated oocysts.
“We have a training set of thousands of images of oocysts,” Lenaghan said. “We manually identify sets that are sporulated or non-sporulated, then train the system.”
Once that is complete, the researchers will run samples of oocysts through the system while a human operator simultaneously scores the results to determine accuracy. For this, they will use Eimeria oocysts obtained from the U.S. Department of Agriculture. Their goal is also to achieve an equally high confidence level of identification.
The second part of their research will involve validating strategies that use gamma radiation, ultraviolet light, ozonation or chlorine dioxide gas to inactivate Cyclospora oocysts.
Finally, they will screen additional compounds, including some chemicals and washes currently used in agricultural systems. The high-throughput system also will allow them to look at different concentrations. Without artificial intelligence and machine learning, Lenaghan said screening this many compounds would likely be impossible.
A human operator would have a difficult time continuously looking through a microscope,” he said. “The advantage of automation and machine learning is that the instrument can run 24 hours per day.”
The research project will also include an economic analysis of the general costs of each treatment.
— Center for Produce Safety