Source-linked AI summary
Early-detection and classification of live bacteria using time-lapse coherent imaging and deep learning
Hongda Wang, Hatice Ceylan Koydemir, Yunzhe Qiu, Bijie Bai, Yibo Zhang, Yiyin Jin, Sabiha Tok, Enis Cagatay Yilmaz, Esin Gumustekin, Yair Rivenson, Aydogan Ozcan
TL;DR
Rapid, high-throughput bacterial detection and classification remain needed. This paper presents a time-lapse coherent imaging platform with deep neural networks that detects 90% of colonies within 7–10 hours and classifies about 80% of tested colonies within 7.6–12 hours.
Problem
The paper addresses the need for an automated method for rapid, high-throughput bacterial detection.
Method
The system combines time-lapse coherent imaging of bacterial growth with deep neural networks for automated detection and species classification.
Results
90% of bacterial colonies were detected within 7–10 h, >95% within 12 h at 99.2–100% precision, and ~80% were classified within 7.6–12 h.
Takeaways & Limitations
The platform provides an automated approach for early detection and classification of bacterial colonies.
Takeaways & Limitations
The reported classification results represent the lower limits of the system’s classification capabilities.
Abstract
from arXiv · showhide
We present a computational live bacteria detection system that periodically captures coherent microscopy images of bacterial growth inside a 60 mm diameter agar-plate and analyzes these time-lapsed holograms using deep neural networks for rapid detection of bacterial growth and classification of the corresponding species. The performance of our system was demonstrated by rapid detection of Escherichia coli and total coliform bacteria (i.e., Klebsiella aerogenes and Klebsiella pneumoniae subsp. pneumoniae) in water samples. These results were confirmed against gold-standard culture-based results, shortening the detection time of bacterial growth by >12 h as compared to the Environmental Protection Agency (EPA)-approved analytical methods. Our experiments further confirmed that this method successfully detects 90% of bacterial colonies within 7-10 h (and >95% within 12 h) with a precision of 99.2-100%, and correctly identifies their species in 7.6-12 h with 80% accuracy. Using pre-incubation of samples in growth media, our system achieved a limit of detection (LOD) of ~1 colony forming unit (CFU)/L within 9 h of total test time. This computational bacteria detection and classification platform is highly cost-effective (~$0.6 per test) and high-throughput with a scanning speed of 24 cm2/min over the entire plate surface, making it highly suitable for integration with the existing analytical methods currently used for bacteria detection on agar plates. Powered by deep learning, this automated and cost-effective live bacteria detection platform can be transformative for a wide range of applications in microbiology by significantly reducing the detection time, also automating the identification of colonies, without labeling or the need for an expert.
Introduction
The introduction identifies the need for rapid, sensitive, automated live-bacteria detection because culture-based methods are slow and expert-dependent, while molecular methods may lack sensitivity and cannot distinguish live from dead microorganisms. It presents a time-lapse coherent imaging platform using deep neural networks to detect, classify, and count bacterial colonies, achieving rapid detection and species identification in water samples.
- Motivation: Waterborne diseases affect more than 2 billion people worldwide, and treatment costs exceed $2 billion annually in the United States.
- Limitations of existing methods: 24–48 h is often required for bacterial colonies to reach macroscopic visibility with traditional culture-based detection methods.
- Limitations of existing methods: Molecular detection can reduce assay time to a few hours but may miss bacteria at 1 CFU per 100-1000 mL and cannot distinguish live from dead microorganisms.
- Proposed platform: The proposed platform combines time-lapse coherent imaging with two deep neural networks to detect bacterial growth and classify species from spatio-temporal holographic features.
- Performance: ~1 CFU/L was detected within ≤ 9 h of total test time, saving more than 12 h versus gold-standard EPA methods requiring at least 24 h.
- Performance: 90% detection sensitivity was achieved within 7–10 h and >95% within 12 h, with ~99.2-100% precision and 80% species identification within 7.6–12 h.
Results
The system rapidly detected bacterial colonies and classified their species from time-lapse holographic images of agar plates. It achieved high detection precision, species-classification accuracy, and a ~1 CFU/L detection limit within ≤9 h using pre-incubation.
- Detection performance: 90% of true positive colonies were detected after ~1 additional hour beyond species-specific 80% detection times, and >95% of all three species within 12 h.80% detection occurred at ~6.0 h for K. pneumoniae, ~6.8 h for E. coli, and ~8.8 h for K. aerogenes.
- Detection performance: 99.2-100% precision was maintained for all tested species after 7 h of incubation.The precision quickly rose to ~100% within 6 h.
- Species classification: ~80% of colonies were correctly classified within ~7.6 h for K. pneumoniae, ~8 h for E. coli, and ~12 h for K. aerogenes.Species classification additionally distinguishes total-coliform sub-species, which traditional agar-plate counting cannot do.
- Benchmark comparison: The system surpassed Colilert®-18 sensitivity within ~8 h and detected 2 CFU at 8.5 h for the lowest tested concentration of ~1 CFU/L.For the same contaminated water sample, Colilert®-18 achieved 1.4±1.6 CFU/L after 18 hours of incubation.
- Sensitivity and detection limit: ~1 CFU/L was detected within ≤9 h of total test time when pre-incubation was used.The system detected more than 2 CFU per test in ≤9 h across tested concentrations of ~1-160 CFU/L.
Discussion
The platform enables early, automated bacterial growth detection and species classification while integrating with EPA-approved agar-plate methods. Its lens-free, computationally refocused design provides high-throughput, alignment-tolerant imaging and supports scalable microbiology applications.
- Performance: More than 12 h were saved versus gold-standard EPA-approved methods requiring 18–24 h, while remaining compatible with existing analytical workflows.The platform can be integrated with EPA-approved agar-plate methods for improved analysis.
- Performance: 90% of bacterial colonies were detected within 7–10 h, >95% within 12 h, with 99.2–100% precision, and ~80% were correctly classified.Classification required 7.6, 8.8, and 12 h for K. pneumoniae, E. coli, and K. aerogenes, respectively.
- Imaging advantages: ~24 cm2/min effective imaging throughput enabled whole-plate scanning, compared with ~128 min for a traditional 4× bright-field microscope.The prototype completes a 242-tile scan in 87 s; computational autofocusing avoids mechanical axial focusing.
- Imaging advantages: Computational refocusing tolerates structural variation and localizes colonies at different depths in ~5 mm-thick culture media from a single hologram.The system resolved colonies at depths including ~2170 µm from the agar surface.
- Scalability: The modular platform is scalable to larger samples and shorter scan intervals, while hardware and control optimization could raise throughput above 50 cm2/min.The monitoring field of view is limited by image acquisition time and stage speed.
Methods
The study used three identified bacterial strains, CHROMagar™ ECC plates, and lens-free imaging to monitor growth. Methods also included chlorine-stressed E. coli samples and comparisons with EPA-approved and culture-based measurements.
- Organisms and safety: The cultures comprised E. coli ATCC® 25922™, K. aerogenes ATCC® 49701™, and K. pneumoniae subsp. pneumoniae ATCC®13883™.Experiments were conducted in a Biosafety Level 2 laboratory under UCLA environmental, health, and safety rules.
- Agar preparation: CHROMagar™ ECC was prepared as the solid growth medium and dispensed as 10 mL into 60 mm × 15 mm Petri dishes after heating and cooling.The mixture used 8.2 g substrate in 250 mL reagent grade water, heated to 100 °C and cooled to ~50 °C before dispensing.
- Lens-free imaging preparation: 1–200 CFU per 0.1 mL suspensions were spread onto CHROMagar™ ECC plates, which were inverted and incubated at 37 °C in the optical platform.Suspensions were prepared daily from 24-hour agar cultures and quantified with a spectrophotometer before dilution.
- Comparison measurements: Performance was compared with Colilert® 18, an EPA-approved enzyme-based method, and plate counting on TSA and ECC ChromoSelect Selective Agar plates.Water samples included 1 L and 1.2 L preparations, with bacterial suspensions spiked into samples and replicate plates.
b. Design of the high-throughput time-resolved microorganism monitoring platform
The platform integrates five modules for automated time-resolved holographic monitoring: imaging, translation, incubation, control circuitry, and software. Its components support programmable acquisition, raster scanning, temperature control, and coordinated operation.
- Platform architecture: Five modules comprise the platform: holographic imaging, mechanical translation, incubation, control circuitry, and a controlling program.Each module is described as part of the integrated system.
- Holographic imaging: The imaging system transmits partially coherent laser light through the agar plate to form inline holograms on a CMOS sensor.The sensor has 1.67 μm pixels and a 6.4 mm × 4.6 mm active area.
- Holographic imaging: 4 ms to 167 ms exposure times are pre-calibrated by illumination intensity, and images are saved as 8-bit bitmap files.A C++ program remotely controls the illumination and software-triggers the sensor through USB 3.0.
- Mechanical translation: Two stepper motors drive the customized two-dimensional mechanical stage, which scans the entire Petri dish in a raster pattern.The stage uses linear rails, bearing rods, linear bearings, and 3D-printed structural parts.
- Incubation and control: The incubation unit uses a microscope-incubator heating plate, while an Arduino-based circuit controls motion, sensor connection, and power.The heating plate is maintained at 47 °C, and the control circuit includes stepper drivers and a MOSFET-based digital switch.
c. Data acquisition
The system acquired full-plate lens-free holograms every 30 minutes under 532 nm illumination, then used independent validation imaging and colony datasets to train and test classification models.
- Acquisition protocol: Images of entire agar plates were captured at 30-minute intervals using 532 nm illumination at ~400 μW intensity.Images were buffered in computer memory before being written to disk by an independent thread to maximize acquisition speed.
- Acquisition protocol: After 24 h of incubation, a benchtop scanning microscope imaged each plate in reflection mode, and images were automatically stitched into a full-FOV comparison image.The imager was an Olympus IX83 microscope.
- Dataset construction: Approximately 6,969 E. coli, 2,613 K. aerogenes, and 6,727 K. pneumoniae colonies populated the training and validation datasets.These data comprised time-lapse lens-free images of individual bacterial colonies.
- Dataset construction: Another 965 colonies from 3 different species across 15 independent agar plates were used for blind testing of the machine-learning models.The test set was separate from the colonies used to train and validate the models.
d. Image processing and analysis
The pipeline stitches holographic tile scans into full-plate frames, identifies colony candidates through time-lapse differential analysis, and uses two deep neural networks for bacterial detection and species classification. Colony counting is verified against late-incubation ground truth while restricting analysis to the central 50 mm-diameter plate area.
- Colony candidate selection: >50 connective pixels above intensity threshold 12 across four consecutive frames define colony candidates.Low-pass and high-pass filtering suppress high-frequency noise and slowly varying background signals, while non-bacterial objects are removed by two DNNs.
- DNN-enabled detection: 160 × 160-pixel candidate regions are represented as 2 × 4 × 160 × 160 phase/amplitude-time-x-y arrays for DNN detection.The custom network replaces 2D convolutional layers with P3D convolutions to incorporate the temporal dimension.
- Colony counting: 50 mm-diameter central imaging limits reduce boundary distortion, while time-lapse images verify merged colonies against ground truth to avoid over-counting.Ground-truth colony information is created after >24 h incubation.
f. Calculation of imaging throughput
The section defines space-bandwidth product (SBP) as the basis for comparing imaging throughput between the proposed system and a conventional lens-based scanning microscope. It specifies the information-content and resolution parameters used in that comparison.
- f. Calculation of imaging throughput: Imaging throughput was compared using the space-bandwidth product (SBP) between the proposed system and a conventional lens-based scanning microscope.The comparison is reported in Supplementary Table S2.
- f. Calculation of imaging throughput: The SBP calculation uses effective pixel-count, half-pitch resolution, digital sampling factor, and independent spatial information from phase and amplitude images.The digital sampling factor applies along the x and y directions.
- f. Calculation of imaging throughput: The holographic reconstruction uses α=2 for phase-and-amplitude information, whereas the lens-based bright-field microscope uses α=1 for amplitude-only information.The lens-based microscope used a color camera with 7.4 µm pixels and a 4× objective, limiting image resolution to ~3.7 µm by Nyquist sampling.