New AI Enables Rapid Detection of Harmful Bacteria

UCLA uses deep neural networks to rapidly identify disease-causing bacteria.

Posted Jul 11, 2020

Geralt/Pixabay
Source: Geralt/Pixabay

Testing for pathogens is a critical component of maintaining public health and safety. Having a method to rapidly and reliably test for harmful germs is essential for diagnosing diseases, maintaining clean drinking water, regulating food safety, conducting scientific research, and other important functions of modern society.

In recent research, scientists from University of California, Los Angeles (UCLA), have demonstrated that artificial intelligence (AI) can detect harmful bacteria from a water sample up to 12 hours faster than the current gold-standard Environmental Protection Agency (EPA) methods.

In a new study published yesterday in Light: Science and Applications, the researchers created a time-lapse imaging platform that uses two separate deep neural networks (DNNs) for the detection and classification of bacteria. The team tested the high-throughput bacterial colony growth detection and classification system using water suspensions with added coliform bacteria of E. coli (including chlorine-stressed E. coli), K. pneumoniae, and K. aerogenes, grown on chromogenic agar as the culture medium. These samples were placed in a lens-free imaging system that can automatically capture time-lapsed holographic images every half hour during a 24-hour period.

These holographic images are digitally woven together and reconstructed to show the bacterial growth patterns on the agar medium. A DNN-based detection model was used to filter the bacterial from the nonbacterial objects. The training used 66 agar plates, 30,000 non-colony objects, and over 13,700 growing bacterial colonies of E. coli, K. pneumoniae and K. aerogenes. To finalize the network models, the team used over 2500 colonies and more than 13,000 non-colony objects from five agar plates. Next, the team used a DNN model to classify the species of the detected colonies that was trained with over 7,900 growing colonies.

Amazingly, the team reported that the entire training process was completed in about five hours using a desktop computer with dual GPUs (GTX1080Ti, Nvidia). Next, they tested the system with new 965 colonies from 15 agar plates that were not used for the previous network training nor validation.

The researchers reported that their system can automatically detect bacterial growth as early as in three hours and 90 percent of bacterial colonies within seven to ten hours with a precision of 99.2 to 100 percent accuracy. This represents an overall time savings of over 12 hours compared to the gold-standard methods such as the Colilert test and Standard Method 9222B that require 18-24 hours, according to the researchers.

The new system has other advantages as well. As a lens-free holographic microscopy system, it can autofocus onto the object plane computationally, without the need for time-consuming mechanical axial focusing. The system is flexible in handling structural changes and variations, making optical alignment easier. The researchers pointed out that for shorter time intervals than what was used for their proof-of-concept, an image processing procedure could be sped up if programmed in C/C++ instead of MATLAB and Python/PyTorch.

Through the cross-disciplinary application of AI deep learning, holographic imaging, and microbiology, scientists have developed a significantly faster way to provide early and rapid screening for disease-causing bacteria that may help improve health care, food safety, and water cleanliness to benefit humankind in the future.   

Copyright © 2020 Cami Rosso All rights reserved.