Multispectral Infrared Colony Phenotyping for High‐Throughput Microbiological Control of Waters

TL;DR

A new infrared imaging tool can quickly and accurately identify microbes in water samples, improving water quality monitoring.

Contribution

A prototype using DFIR multispectral imaging for rapid microbial identification on filtration membranes is developed and tested.

Findings

The prototype achieved 96.5% accuracy in identifying microbial colonies from a database of 3,230 colonies.

The system captures both spectral and morphological fingerprints of colonies without the need for labels or destructive methods.

Abstract

Microbiological water quality assessment relies on culture‐based methods that are time‐consuming, resource‐intensive, and often lack specificity. To address these limitations, we developed a prototype for automated, label‐free, and nondestructive microbial identification based on discrete frequency infrared (DFIR) multispectral imaging. By combining monochromatic quantum cascade lasers (QCLs) with an uncooled bolometer array, this prototype captures spectral and morphological fingerprints of colonies directly on filtration membranes. A demonstration database of 3230 colonies from 11 strains across 7 genera was acquired. In average, deep‐learning based classification achieved a 96.5% ± 1.3% correct identification rate. Overall, this prototype brings DFIR imaging one step closer to an industry‐ready microbial identification tool.

Microbiological water quality monitoring still relies on slow, culture‐based methods with limited specificity. We present an automated, label‐free DFIR multispectral imaging prototype that captures both spectral and morphological fingerprints of microbial colonies directly on filtration membranes. Applied to a 3,230‐colony database, it achieved 96.5% accuracy, highlighting its potential as an industry‐ready tool for rapid microbial identification.

Figures

Optical setup description. (A) Laser head schematic. (B) Dichroic filter spectral response (identical for all four filters). (C) Fiber‐guided optical path from laser head to sample and sensor. BM: bolometer matrix; DF: dichroic filter; FH: fiber holder; HCF: hollow‐core fiber; HR: holding rod; OAP: off‐axis parabolic mirror; PM: prism mirror; QCL: quantum cascade laser; RM: rotating motor; SH: sample holder; XY‐MS: XY microscopy stage.

Method overview.

Acquisitions examples. (A) Region of interest localization on a complete membrane scan, with colonies of E. coli highlighted in white and references positions in red. Since analysis is purely phenotypical, only isolated colonies are selected. Image was acquired at 1655 cm−1. Fringes observed around colonies 8, 9 and 10 are due to interferences between the membrane and the bolometer. (B) Colonies representative of the different strains represented in the database.

Architecture of the deep learning model used for classification. It takes as input a stacked infrared image of dimension [8, 80, 80], and outputs a vector which presents a probability distribution over the 11 possible classes. An ArgMax operation (not represented) selects the class with the highest probability as the final output. Convolutional operations (in dark blue) are accompanied by their respective parameters: in_channels and out_channels (kernel_size, stride, padding).

Cross‐validation results on validation dataset.

Identification rates for the four polymicrobial membranes. Each bar represents the percentage of correctly (green) or wrongly (light purple) classified colonies. Correct identification counts are labeled on the right side (correctly labeled colonies/total number of colonies). Strains are grouped in four pairs, corresponding to the four polymicrobial membranes.

Tables

TABLE 1. Wavelengths of interest.

Wavenumber (cm−1)	Wavelength (μm)	Biochemical significance
1236	8.1	Asymmetric stretching of P=O bond in phosphodiester groups
1274	7.8	“Amide III” peak. Mixture of several coordinate displacements
1353	7.4	Reference
1397	7.1	Symmetric stretching of C=O in ionized carboxylic groups
1553	6.4	“Amide II” peak. Stretching of C—N in peptide bond
1656	6.0	“Amide I” peak. Stretching of C=O in peptide bond
1713	5.8	Stretching of C=O in carboxylic acid groups
1736	5.8	Stretching of C=O in ester groups

TABLE 2. Database composition.

Genera	Species	ATCC reference	Colonies acquired
Escherichia	coli	25922	221
Escherichia	coli	8739	219
Escherichia	coli	43888	266
Enterobacter	cloacae	13047	158
Klebsiella	aerogenes	13048	271
Serratia	liquefaciens	27592	254
Enterococcus	faecium	700221	347
Enterococcus	durans	6056	505
Streptococcus	bovis	33317	384
Staphylococcus	epidermidis	14990	329
Staphylococcus	saprophyticus	35661	276

Full text

Introduction

1

In 2024, microbial quality of water was brought under the spotlight by the Olympics free water swimming competition, when concentrations of Escherichia coli and Enterococcus of the Seine sparked international interest. Though far less public, these microbiological quality controls are in fact common to the more than 22 000 bathing waters of the European Union [1]. To these must be added the countless daily controls performed on drinking or industrial waters intended for human consumption. This large‐scale monitoring tracks waterborne pathogens (Legionella, Salmonella, etc.), as well as bacterial species relevant to fecal contamination, mainly E. coli , coliforms, and fecal enterococci [2].

Standard protocols for identification and enumeration of microorganisms rely on a capture of microbial cells by filtration of a water sample (ISO 8199, ISO 9308, ISO 7899). Culturing these cells on selective and/or chromogenic media then allows isolation or coloration of specific bacterial genera, enabling a certain level of identification [3]. Despite the availability of methods such as qPCR [4] or chemical labeling [5, 6, 7], culture‐based identification remains largely in use. This is due to the wide range of media available for different targets, as well as their tolerance to mutations that may otherwise imper identification by PCR. Still, the low specificity of these media limit the precision of identification, which raises sanitary and quality monitoring questions [8].

To mitigate this issue, culture is sometimes coupled with modern identification techniques such as matrix‐assisted laser desorption ionization–time of flight mass spectrometry (MALDI‐TOF MS). This technology, currently considered a gold standard in clinical microbiology, is known for its speed and identification performances, sometimes up to the strain level [9, 10, 11]. However, instrumentation is expensive, especially for small laboratories [12, 13, 14].

Here, we suggest a novel technique that could lead to the development of a bench‐top instrument dedicated to identification of bacterial colonies. This technique, called discrete frequency infrared multispectral imaging (DFIR imaging), is reagent‐free and provides optical fingerprints that blends chemical and morphological data [15]. It relies on the same physical principle as infrared spectroscopy, a technology that recently entered the market as a tool for microbial typing [16, 17, 18]. Infrared spectroscopy uses the absorption of infrared light by the chemical bounds of biomolecules to gain insights on the composition of a sample. Applied to microbial identification, it provides a powerful tool for fingerprinting a given biochemical phenotype [19]. Fourier transformed infrared spectroscopy (FTIR) specifically gained growing interest due to its concordance with whole genome sequencing [20] and performances in strain‐typing rivaling MALDI‐TOF [21, 22]. However, its sensitivity to culture variance is source of errors, even with strong standardization processes [11, 23].

Compared to FTIR spectroscopy, DFIR imaging introduces three major evolutions. First, it replaces the broadband, low‐power thermal source used in FTIR interferometer with a series of monochromatic lasers with illumination power of up to 100 mW. These quantum cascade lasers (QCLs) are powerful enough to allow live analysis of biological samples of up to 100 μm in thickness, which was previously impossible due to the strong absorbance of water in the mid‐infrared range (3–10 μm) [24]. Next, the continuous spectral acquisition of FTIR spectroscopy is replaced by a multispectral acquisition. Here, only a few relevant wavelengths are considered, which limits the amount of spectral data available but considerably speeds up acquisitions. Due to the redundancy of information in bacterial infrared spectra, this restricted set of wavelengths should still explain most variance between spectra of different species [21, 25]. Finally, replacing the single‐point sensor by an infrared imager allows us to perform spectral imaging, which combines both morphological and spectral information. Spectral imaging is well established in the literature as a way to gain a great amount of data for colony identification [26, 27, 28, 29]. However, attempts to transpose this technique in the mid‐infrared have been limited by the long acquisition times of FTIR hyperspectral imaging and the need to report colonies from culture medium to an IR‐transparent imaging support [30]. Whereas in DFIR, the illumination power of QCLs is high enough to image colonies directly through the filtration membrane that acts as both culture and imaging support [31]. Moreover, DFIR is orders of magnitude faster than FTIR for tasks such as imaging (minutes to image a few cm^2^ instead of hours), especially with a restricted number of wavelengths of interest [32].

We previously argued that DFIR imaging could overcome the limitations of FTIR and open new possibilities for microbial analysis [15, 31, 33, 34]. Here, we report the creation of a prototype for a future bench‐top instrument. To assess performance, a simple demonstration database, dedicated to the specific case of water quality assessment, was acquired; 3230 colonies were imaged using our prototype. Colonies were grown on monoculture plates and belonged to 11 strains of genera commonly encountered in water (Escherichia, Klebsiella, Enterobacter, Serratia, Enterococcus, Streptococcus, and Staphylococcus). A deep learning model was trained to classify colonies and evaluated on a few cases of polymicrobial membranes. Based on these early results, we argue that such a tool could make infrared‐based identification available for water quality assessment laboratories.

Material and Methods

2

DFIR Method

2.1

Optical Principle

2.1.1

Chemical bonds are excited by absorbing infrared light at specific wavelengths, leading to transitions to rovibrational energy levels [19]. Each chemical bond and molecule absorbs light at specific wavelengths. Infrared spectroscopy uses this principle to identify chemical compounds, based on their spectral fingerprint. For multispectral acquisitions, fingerprinting is based on absorbance at only a handful of wavelengths. Here, eight wavelengths, listed in Table 1, were selected based on our previous studies [31] and literature review [19, 30, 35]. They all correspond to absorption peaks of major biomolecules used for identification or analysis of biological samples.

Prototype Architecture

2.1.2

Our prototype, described in Figure 1, is composed of three main components:

1. A laser head containing eight mid‐infrared QCLs.
2. A sample holder mounted on a motorized XY microscopy stage.
3. A detection head based on a 2D uncooled bolometer array.

The laser head (Figure 1A) includes eight QCLs, each emitting at one wavelength of interest. The Gaussian beams from each laser (∼3 mm in diameter) are first combined two‐by‐two using a custom dichroic filter (Materion, see spectral response on Figure 1B). The resulting four beams are directed through three successive silver‐coated knife‐edge right‐angle prism mirrors (MRAK25‐P01, Thorlabs) into an optical reducer composed of two silver‐coated off‐axis parabolic mirrors (focal lengths: 76.2 mm and 15 mm, MPD039‐P01 and MPD00M9‐P01, Thorlabs), reducing the beam size to ~0.6 mm.

Optical setup description. (A) Laser head schematic. (B) Dichroic filter spectral response (identical for all four filters). (C) Fiber‐guided optical path from laser head to sample and sensor. BM: bolometer matrix; DF: dichroic filter; FH: fiber holder; HCF: hollow‐core fiber; HR: holding rod; OAP: off‐axis parabolic mirror; PM: prism mirror; QCL: quantum cascade laser; RM: rotating motor; SH: sample holder; XY‐MS: XY microscopy stage.

Due to imperfect spatial overlap, the beams are injected into a 1 m‐long hollow‐core fiber, as described in Figure 1C (1.5 mm core, metallic coating, HF500MWLW‐SMA‐1m, Guiding Photonics). The fiber fulfills two objectives. It serves as a sequential multiplexer of the four optical beams thanks to the high numerical aperture (divergence angle at ±30 mrad), and as a homogenization technique to ensure similar output illumination regardless of the wavelength. To do so, this fiber, annealed for minimal bending loss, is set into large‐amplitude (cm), low‐frequency (of the order of magnitude of hundreds of rpm) motion using a rotating motor (Maxon) and holding rod to promote optical mode mixing and increase the number of fiber‐induced random phase shifts between coherent points of the beam section on the sensor. The resulting speckle contrast defined as the ratio of the standard deviation over the mean value of the beam intensity on the sensor plane can reach values as low as 15%–10%, depending on the laser wavelength.

The output beam is then expanded using two silver‐coated parabolic mirrors (focal lengths: 50.8 mm and 152.4 mm, MPD129‐P01 and MPD169‐P01, Thorlabs) to produce a uniform illumination spot of 4–5 mm in diameter. The laser head is controlled by a custom electronic board that regulates temperature and sequentially enables each laser to image the sample at a specific wavelength, delivering around 0.5 mW mm^−2^ of surface power density per channel.

The sample holder secures 47 mm‐diameter membranes (see Section 2.2.2) using magnets. It is mounted on a Thorlabs MLS203‐1 motorized XY microscopy stage with a 110 mm × 75 mm travel range.

The detection head features an 80 × 80 pixels uncooled bolometer array (Micro80, Lynred), developed with pixel‐level packaging [36]. This design minimizes the distance between the final sensor window and the focal plane array (FPA) to just a few tenths of micrometers—compared to several millimeters in conventional packaging—allowing close proximity to the sample. The sensor window has been removed to allow broadband operation across the full 5–10 μm range, unlike typical imagers limited to 3–5 μm or 8–12 μm. Pixels are 25 μm wide with a 34 μm pitch, covering a total area of 2.72 mm × 2.72 mm.

Database Constitution and Sample Preparation

2.2

Bacterial Strains

2.2.1

Eleven references strains were selected for this study (Table 2). They include two Staphylococci, one Streptococci, and two Enterococci, as well as six strains representative of the coliform group. Among these, three different strains of E. coli were selected, including American Type Culture Collection (ATCC) 4388, a nonpathogenic model representative of the toxic O157 phenotype. All of these strains were chosen for their relevance for water quality assessment [37, 38, 39]. Biological samples were obtained from Microbiologics (Kwik‐Stik lyophilized strains from ATCC).

Filtration Membranes

2.2.2

Due to the absorbance of water, plastics, and glass in the mid‐infrared, culture plates are opaque at all wavelengths of interest. Infrared studies traditionally transfer samples onto an IR‐transparent imaging medium [21, 30], but this process damages colonies' morphology. Here, we take advantage of the filtration membranes commonly used in water quality analysis [38, 40]. After filtration of a sample of 100 mL, these membranes are deposited on top of a culture medium to allow colony growth. Some materials, such as aluminum oxide (alumina), are transparent in the mid‐infrared (up to 8.1 μm), meaning nanoporous alumina membranes can be used both as a growth support and, after retrieval from culture medium, imaging support [31, 41].

In this study, we used Anodiscs membranes (Whatman, 514‐0518; thickness: 60 μm, pore diameter: 200 nm), with a diameter of 47 mm. They indeed provide a wide area for microbial growth and are rigid enough to be transferred from plate to plate or to the analysis setup. Colony growth is largely unaltered, with cells doubling times remaining in the same order of magnitude as for standard agar‐plate culture [42]. Infrared image acquisitions are feasible between 1235 and 1900 cm^−1^. Typical spectrum of an anodisc membrane can be found in a previous study [33].

Sample Preparation and Culture

2.2.3

Alumina membranes were deposited on Trypticase soy agar plates (TSA, VWR, 111114ZI). Biological samples were inoculated by streaking the membrane surface with an inoculation loop. Incubation was performed at 37°C for 24 h. Membranes are then retrieved from the culture medium and air‐dried at room temperature for 15 min before being placed in the prototype for acquisitions. The process is summarized in Figure 2.

Method overview.

Image Acquisitions and Preprocessing

2.3

Acquisitions begin with a scan of the whole membrane at a single wavenumber, corresponding to the maximal absorption peak of colonies (1656 cm^−1^). The resulting image allows localization of colonies positions using a custom computer vision algorithm (Figure 3). This algorithm also detects reference regions, free of artifact, border, or biological material, that will be used for image preprocessing. After localization, another algorithm controls the motorized stage to successively place regions of interest under the imager. For each of these regions, lasers are enabled one after another to acquire a complete eight‐images multispectral stack. This process takes around 30 s. The microscopy stage, lasers, and camera are all controlled by a homemade python interface.

Acquisitions examples. (A) Region of interest localization on a complete membrane scan, with colonies of E. coli highlighted in white and references positions in red. Since analysis is purely phenotypical, only isolated colonies are selected. Image was acquired at 1655 cm−1. Fringes observed around colonies 8, 9 and 10 are due to interferences between the membrane and the bolometer. (B) Colonies representative of the different strains represented in the database.

After acquisition, raw images are preprocessed and converted to absorbance as described in our previous study [15]. Resulting images are of shape 80 × 80 × 8 in 14 bits. Each of the eight channels corresponds to one wavelength of interest. Each image presents a single isolated colony that belongs to 1 of the 11 possible learning classes. In this study, 3230 colonies were acquired over 157 membranes.

Classification by Deep Learning

2.4

Image classification was performed using deep learning approaches. To this end, we used a custom ResNet described in Figure 4. This architecture was selected based on literature review and tests carried out on images from a previous study [15, 43, 44]. Given the relatively small amount of data available, it was decided not to optimize the model and hyperparameters to avoid overfitting.

Architecture of the deep learning model used for classification. It takes as input a stacked infrared image of dimension [8, 80, 80], and outputs a vector which presents a probability distribution over the 11 possible classes. An ArgMax operation (not represented) selects the class with the highest probability as the final output. Convolutional operations (in dark blue) are accompanied by their respective parameters: in_channels and out_channels (kernel_size, stride, padding).

After its last layer (LogSoftMax), our classification model outputs a single vector. This vector represents the n probabilities for a given image to belong to any of the n possible classes. During training, weights of the model are slowly optimized, so that the maximal probability for each image matches its true class. To do so, we used the negative log‐likelihood loss. This loss measures the distance between a predicted probability and its ground truth counterpart (i.e., class). In PyTorch, negative log‐likelihood loss (nn.NLLLoss) operates on log‐probabilities as input, requiring the model's output to be log‐softmax transformed beforehand. This contrasts with the more usual cross‐entropy loss (nn.CrossEntropyLoss), which combines a log‐softmax transformation and the negative log‐likelihood computation into a single operation. Here, we choose the NLLLoss for interpretability, as we prefer our models to output softmax transformed probabilities.

Model training was performed over 200 epochs, with a batch size of 64 and an initial learning rate of 0.001. The Adam algorithm was used as optimizer. Computations were ran using an Intel i9 and a Nvidia GeForce RTX 4090 for GPU acceleration. Database was split into training and validation datasets by stratified k‐fold, with k = 5 and random splitting. Thus, validation dataset was comprised of 646 images.

Regarding the training dataset, class imbalance (see Table 2) was compensated by oversampling of the lesser‐represented classes. This resulted in a training dataset of 4444 images (444 images for each of the 11 classes). To compensate for the overrepresentation of certain images due to oversampling, a data augmentation strategy was also implemented. Random rotations, translations of one pixel, as well as horizontal and vertical flips were applied on all images of the training dataset [45].

For each training, the model was initialized and trained five times, each time with a different training and validation partition, following the strategy of cross‐validation. Training and validation datasets were created so that each image got into the validation dataset exactly once. Training and validation dataset lengths were the same over the five folds of cross‐validation. Precision, accuracy, recall, F1‐score, balanced accuracy, and Matthews correlation coefficient (MCC) were computed for each fold using scikit‐learn metrics module. Scores and confusion matrix presented in the Section 3 are averaged over the five folds of validation.

Results

3

Classification Results

3.1

Averaged cross‐validation results are presented in Figure 5. The model achieved high and consistent performances across all metrics: accuracy, precision, and recall were each 96.5% ± 1.3%, with an F1‐score of 95.7% ± 1.6%, a balanced accuracy of 95.7% ± 1.6%, and a MCC of 96.1% ± 1.4%. Misclassifications were rare and mainly occurred between closely related strains. As could be expected, the three E. coli strains yield the highest errors, as well as the largest variations from one cross‐validation fold to another (up to 8.74% for ATCC 8739). This high variability could be explained by the very close similarity of colonies, thus making the model extremely sensitive to variations in training data. Still, accuracy for all three E. coli strains remained above 89%.

Cross‐validation results on validation dataset.

Other confusions also reflect phylogenetic similarities ( E. coli and coliforms, S. epidermidis and S. saprophyticus , Enterococcus and Streptococcus). Overall, these discrimination capabilities, as well as the biological consistency of the confusions, make these results extremely encouraging.

Inferences on Polymicrobial Membranes

3.2

During database acquisition, a few bi‐microbial membranes were also prepared. For these experiments, two different strains were streaked on a single membrane—each on a different half, so that they could be visually distinguished. Others steps of sample preparation and images processing were unchanged. After acquisition, colonies were manually labeled based on their position on the membrane. Due to time constraints, only four membranes were prepared in this manner. Each presents a couple of strains, respectively: K. aerogenes / E. cloacae , S. liquefaciens / E. cloacae , S. epidermidis / S. bovis , and S. saprophyticus / E. faecium . These couples were chosen for sample preparation convenience, not to represent a real biological case. Colonies acquired on these membranes were not used for model training.

Here, we wanted to evaluate classification performances on these bi‐microbial membranes. To do so, our model was retrained from scratch on the entire training database, without train/validation split. It was then used to classify colonies from our four bi‐microbial membranes.

As summarized in Figure 6, classification performances are especially high, with between 96% and 100% of correct identification. These results are undeniably helped by the low number of samples, which prevents any firm conclusion regarding technique performances. Still, our model was able to correctly identify all seven strains present on these membranes, without preliminary knowledge. This early result tends to show that our technique and classification model could generalize identification on completely unseen data—even in the vicinity of another strain. Further experiments are obviously needed to confirm these observations, including more strains, more membranes and real polymicrobial samples without physical separation between strains. Still, this quick test remains extremely encouraging.

Identification rates for the four polymicrobial membranes. Each bar represents the percentage of correctly (green) or wrongly (light purple) classified colonies. Correct identification counts are labeled on the right side (correctly labeled colonies/total number of colonies). Strains are grouped in four pairs, corresponding to the four polymicrobial membranes.

Discussion and Conclusion

4

In this study, we introduced a prototype for DFIR multispectral imaging. It builds upon a previous research setup and is specifically designed for microbial identification in water quality assessment. The prototype was fully automated, enabling rapid identification of all microbial colonies grown on filtration membranes—an approach fully compatible with existing water quality control workflows. After a 15 min drying, a whole membrane was analyzed in around 10 min, further improving time‐to‐results of current infrared analysis systems (≈2 h) [46]. Additionally, its nondestructive analysis preserves samples for secondary testing, such as identification confirmation by sequencing or antibiotic susceptibility testing for pharmaceutical quality control [34].

As a proof‐of‐concept, a database of 3230 colonies from 11 strains across 7 genera was acquired using this prototype. A custom ResNet model correctly classified 96.5% ± 1.3% of colonies in validation. Misclassifications were rare and mostly occurred between phylogenetically related species, confirming the robustness of this method. A few tests on polymicrobial membranes hinted at a possible capability to generalize to more complex identification scenarios, even though more experiments are needed.

Future improvements will focus on enhancing sensor spatial resolution and refining the system's design based on the insights gained from this prototype. Additionally, upcoming work will focus on a new database dedicated to a single genus, to evaluate the prototype's capability to discriminate closely related species. These developments will bring DFIR imaging closer to a fully operational, industry‐ready solution for microbial identification in water quality monitoring and beyond.

Funding

This study was partially funded by the European Commission through the HORIZON‐KDT‐JU‐2023‐2‐RIA project (ATHENA No. 101139941).

Conflicts of Interest

The authors declare no conflicts of interest.