(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Spatio-temporal based deep learning for rapid detection and identification of bacterial colonies through lens-free microscopy time-lapses [1]

['Paul Paquin', 'Univ. Grenoble Alpes', 'Cea', 'Leti', 'Dtbs', 'Lsiv', 'Grenoble', 'Claire Durmort', 'Cnrs', 'Ibs']

Date: 2022-11 

Detection and identification of pathogenic bacteria isolated from biological samples (blood, urine, sputum, etc.) are crucial steps in accelerated clinical diagnosis. However, accurate and rapid identification remain difficult to achieve due to the challenge of having to analyse complex and large samples. Current solutions (mass spectrometry, automated biochemical testing, etc.) propose a trade-off between time and accuracy, achieving satisfactory results at the expense of time-consuming processes, which can also be intrusive, destructive and costly. Moreover, those techniques tend to require an overnight subculture on solid agar medium delaying bacteria identification by 12–48 hours, thus preventing rapid prescription of appropriate treatment as it hinders antibiotic susceptibility testing. In this study, lens-free imaging is presented as a possible solution to achieve a quick and accurate wide range, non-destructive, label-free pathogenic bacteria detection and identification in real-time using micro colonies (10–500 μm) kinetic growth pattern combined with a two-stage deep learning architecture. Bacterial colonies growth time-lapses were acquired thanks to a live-cell lens-free imaging system and a thin-layer agar media made of 20 μl BHI (Brain Heart Infusion) to train our deep learning networks. Our architecture proposal achieved interesting results on a dataset constituted of seven different pathogenic bacteria—Staphylococcus aureus (S. aureus), Enterococcus faecium (E. faecium), Enterococcus faecalis (E. faecalis), Staphylococcus epidermidis (S. epidermidis), Streptococcus pneumoniae R6 (S. pneumoniae), Streptococcus pyogenes (S. pyogenes), Lactococcus Lactis (L. Lactis). At T = 8h, our detection network reached an average 96.0% detection rate while our classification network precision and sensitivity averaged around 93.1% and 94.0% respectively, both were tested on 1908 colonies. Our classification network even obtained a perfect score for E. faecalis (60 colonies) and very high score for S. epidermidis at 99.7% (647 colonies). Our method achieved those results thanks to a novel technique coupling convolutional and recurrent neural networks together to extract spatio-temporal patterns from unreconstructed lens-free microscopy time-lapses.

In order to combat the spread of resistances and to meet the specific needs of infection diagnosis in rural and remote settings, there is a strong need for new diagnostic tools that are faster, space efficient and less expensive than conventional ones while still being accurate. To answer those criteria, optical identification methods such as lens-free microscopy seem best-suited. Indeed, lens-free microscopy is an imaging technique that allows for continuous image acquisition during incubation, as it is non-destructive and compact enough to be integrated into an incubator. Here, we propose the combined use of lens-free microscopy imaging and deep learning to train neural networks able to detect and identify pathogenic species thanks to bacteria growth time-lapses on thin-layer agar and a novel deep learning architecture combining recurrent and convolutional neural networks. We achieved high quality detection and identification from 8 hours onwards on a dataset composed of 7 different pathogenic strains. Our study provides new results proving the usefulness of lens-free imaging on pathogenic species for rapid detection and identification tasks in low-resource settings.

3. Introduction

The ever-increasing number of multidrug-resistant bacteria is having a growing influence on mortality worldwide. With the emergence of resistance to Colistin, a last resort antibiotic effective against Gram-negative bacteria and especially the Enterobacterales, at the end of 2015 [1,2], the situation is worsening. Unnecessary usage of wide-spectrum drugs is accelerating the global spread of multidrug resistant organisms. Antibioresistance is even more alarming when it comes to last-resort antibiotics. Infections resulting from those resistant bacteria are more and more difficult to cure. Such is the case with Escherichia coli, which is seeing Colistin resistances emerge in already multidrug-resistant strains [3]. The World Health Organization predicts that by 2050 infectious diseases will once again be the leading cause of death in the world with 10 million deaths annually [4].

Rapid identification of isolated pathogenic bacteria in biological samples (urine, blood, sputum, etc.) is a key step in the diagnosis of an infection, and therefore in its management. Indeed, speed and reliability of identification determine how quickly a specific and optimised antibiotic therapy–more effective and less likely to generate new resistance to antibiotics–will be administered by precisely targeting the identified pathogen.

Current solutions for microbial species identification providing high quality outputs still have many shortcomings resulting from many concessions done in order to achieve those very high identification performance. That is to say, they often require complex, time-consuming and costly identification protocols in exchange for accurate results.

Identification methods commonly used are either based on biochemical profiling by using identification cards such as Vitek 2 [5] or on time-of-flight (TOF) mass-spectrogram in particular with matrix assisted laser desorption ionization (MALDI-TOF) [6,7]. Those techniques usually require isolated colonies grown on overnight cultures; this requires users to wait 18 to 48 hours for at least one identifiable colony to fully incubate. There is also genotypic analyses, which allow identification of a target bacteria or virus thanks to its genomic sequence: the most popular methods using this principle are the sequencing of 16S [8] for prokaryotes and 18S ribosomal Ribonucleic acid (rRNA) [9] for eukaryotes or PCR-based methods for rapid identification [10]. Such analyses allow reliable and robust pathogenic bacteria identification in samples; nevertheless, it is at the expense of an a priori knowledge and understanding of the infection.

For this reason, the current lines of research on microbial identification are diversified. The first direction is the development of identification techniques that address the specific needs of low-resource settings (LRS) or small isolated laboratories [11]. These techniques would expand the use of diagnosis prior to antibiotic therapy, thereby reducing unnecessary use or misuse of antibiotics. The second strand of research focuses on automation: many steps in the entire process have not yet been automated and require human processing. For instance, in the case of MALDI-TOF, the choice of the colonies to be analysed, their sampling, the mixing with specific reagents, are several tricky steps to automate. A fully automated identification method would lead to a “smart incubator”, [12] i.e. an incubator able to provide an automatic identification on every incubated plate, as soon as enough phenotypic information is available. In that manner, identification results could be delivered during plate incubation, usually at night, and all human processing done the day after could be dedicated to antibiotic susceptibility testing (AST). The third and last area of research is single-cell characterization [13–16]: if a technique can provide an identification on very low biomasses, typically 100 cfu (colony-forming unit) and less, the results could be obtained before the incubation step, which would reduce time-to-result significantly (18 h to 48 h less).

Optical identification methods, particularly when they do not require labelling techniques (i.e. fluorescence microscopy), have decisive advantages: non-invasive, rapid and non-destructive, they could potentially allow automated, reliable, low-cost and high-throughput diagnosis. These techniques are based on various optical properties such as scattering, absorption and emission, reflection or even phase contrast each of which allows for the study of numerous bacterial characteristics [17,18] (S1 Table). On the one hand, spectroscopic methods highlight bacterial biochemistry by collecting spectral data over a certain range through the use of absorption and emission (Fourier transform infrared spectroscopy) [19] or inelastic scattering (Raman microspectrocopy [20,21]). On the other hand, scattering methods tend to focus on colony morphology due to their use of diffraction and reflection principles to produce spatial data. In particular, elastic scattering [22,23] and more precisely forward scattering (BARDOT) [24,25] both provide specific direct two-dimensional patterns for each bacterial colony based on their shape, which makes these techniques sensitive to intra-species variation. Furthermore, some optical setups are capable of producing both spectral and spatial data thanks to hyperspectral approaches over large field of view, some even imaging full standard petri dishes. For instance, hyper spectral diffuse reflectance microscopy [26,27] allows the entire dish to be captured and its spectral signature to be acquired for later use thanks to in line scanning. Nevertheless, these systems must wait for colonies to reach a certain size for them to be imaged and are not compact enough to be placed inside an incubator to obtain real-time images. This leads to a delay in the detection and identification processes.

Lens-free imaging allows for wide-range detection and identification through direct diffractograms analysis [28] or holographic reconstructions [29,30]. Thanks to “colony fingerprint” (i.e. discriminative parameters) and through cluster analysis, Maeda et al. [31] showed that discriminating microorganisms was possible with lens-free imaging within the visible range after 8h of incubation on a limited dataset. In the same manner, Wang et al. [32] proposed an automated lens-free imaging system to detect and classify bacteria by using neural networks on a small time-lapse (2 h) obtaining impressive results (>95% detection rate and 99.2–100% precision within 12h) on 3 different Enterobacterales (Klebsiella pneumoniae, Escherichia coli, Klebsiella aerogenes).

In this paper, micro colonies long kinetic growth patterns (>6 h) are used to train a two-stage deep learning architecture that will detect and identify pathogenic bacteria over a wide-field, through growth time-lapse, in less than 12 hours.

This method differs from other lens-free techniques by its approach based on thin-layer un-reconstructed video microscopy for detection and classification on some of the most prevalent anaerobic pathogenic bacteria in clinical settings. Among other things, it differentiates itself through its use of a thin-layer culture medium that minimises the sample-sensor distance. This technique works well in a low-density context (at maximum ~ 200 colonies/mm2) as it is possible to image the sample over a large field-of-view while avoiding a magnification-induced superimposition of holograms which occurs when the sample is moved away from the sensor. This gives access to a better resolution and higher frequency information than on a Petri dish during a single-shot acquisition. Nevertheless, this technique’s sample preparation protocol includes a pour-plating step, thus colonies do not necessarily grow on the same plane as there is a height variability.

This is where our un-reconstructed holographic approach shines by avoiding either making an a priori estimation of the sample-sensor distance or having to produce multiple reconstructions at different heights. It also removes the need for auto-focus, which is not the case for Petri dishes monitored over time due to inter-plate (Agar pouring height) and intra-plate height variability (Agar drying out in the incubator) during an experiment. This method thus avoids all the problems inherent with holographic reconstruction (phase wrapping, twin images, etc.) by learning directly from a series of holograms within a height range. Finally, it is the use of neural networks throughout the processing pipeline that allows this method to obtain rapid and robust results, on par with other work. This pipeline consists of a region-based convolutional network for detection and a novel dual association of a spatial encoder (3D convolutional network) with a temporal encoder (2D recurrent network).

[END]
---
[1] Url: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000122

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/