Device for detecting bacteria in a sample

Abstract

A device for detecting bacteria in a sample comprises an optical block (4) arranged to produce multispectral image sets of portions of a sample, obtained by illumination eliciting an autofluorescence response and by illumination eliciting a reflectance response. These multispectral image sets are provided to an analyzer (6) comprising a classifier that produces an image in which each pixel is derived from an identical pixel in the images of the multispectral image set and is associated with an element identifier indicating whether that pixel is associated with a bacterium or another element. This image is then used by a computer to determine bacteria detection data. Fig. 1

Description

Title of the invention: Device for detecting bacteria in a sample

[0001] The invention relates to the field of microbiology and in particular the detection of bacteria in samples.

[0002] In the agri-food sector, bacterial detection is a major issue. All stakeholders are subject to extremely rigorous hygiene standards to prevent the sale of food containing bacteria, regardless of its degree of processing. Similar problems exist in the pharmaceutical and cosmetic sectors.

[0003] Indeed, when the presence of bacteria is detected too late, many consumers often suffer the consequences, which can range from minor digestive discomfort to poisoning and / or infections that can be fatal. A. State of the art

[0004] The detection of the presence of bacteria is based in the state of the art on taking samples at regular intervals on the production lines, and on counting the bacteria present in the samples after cultivation.

[0005] This conventional approach involves cultivating bacteria for several days on a nutrient medium, followed by manual counting of bacterial colonies to determine the number of bacteria, generally measured in colony-forming units per milliliter (CFU / mL). However, although considered the gold standard in microbiology, this method has several drawbacks, notably the significant time required to obtain results, which can range from two to five days depending on the specific microorganisms being quantified. This duration is even greater in the case of sterility tests, which can take up to 15 days in the pharmaceutical and cosmetic industries.

[0006] In recent years, alternative methods have been explored to try to replace the Petri dish culture technique: the polymerase chain reaction (PCR), the most probable number approach, ATP measurement, flow or solid-phase cytometry, microscopy-based methods, methods based on the analysis of hyperspectral measurements, or the use of excitation and emission matrices, etc. PCR Methods

[0007] Among these exploratory methods, PCR offers the fastest results, generally in less than an hour. However, PCR has limitations because it can only detect specific genetic sequences (such as the 16S rRNA gene), as with the QuickScan quantitative test (registered trademark). This gene is not present in all bacterial species, which can introduce bias. Generally speaking, each PCR test focuses on a specific genetic sequence, making these tests insufficiently "generalist." Furthermore, the use of reagents in these methods makes them expensive and less accessible for routine use compared to other methods. A.2 "Most probable number" method

[0008] The most probable number approach is a culture-based approach. Biomérieux has optimized this approach by using microfluidics and coupling it with a fluorescent dye to develop a solution called Tempo (registered trademark). To test the suspension, it is loaded and incubated in a microchamber with a culture medium coupled to a fluorescent dye. The limit of detection of this method ("Limit of Detection" or "LOD" in English) is 1 CFU / mL (see the article by Cayer et al. "Evaluation of the Tempo® System: Improving the Microbiological Quality Monitoring of Human Milk", Frontiers in Pediatrics, Volume 8, 2020, https: / / doi.org / 10.3389 / fped.2020.00494) and a dynamic range of 3.7 log, allowing the detection of the total number of viable bacteria ("Total Viable Count" or "TVC" in English) between 24 and 48 hours.

[0009] This detection time is too long and does not constitute a sufficient improvement compared to culture in Petri dishes. A.3 Methods based on ATP measurement

[0010] ATP measurement involves amplifying the signal generated by the presence of adenosine triphosphate (ATP). This allows for rapid results in a few seconds, but this method is very sensitive to organic contamination caused by chemicals or detergents, which can lead to many false positives. Because ATP measurement is easy to deploy in the field and relatively inexpensive, it is frequently used, but remains unsuitable as a reliable, single measure of the presence of bacteria. A.4 Flow cytometry methods

[0011] This method is based on illuminating a capillary containing the sample with a laser. It can use dyes to label the bacteria, and an optical sensor is used to collect the photons emitted by the labeled bacteria.

[0012] Flow cytometry still has significant limitations. First, it counts both live and dead cells and requires the analysis of large sample volumes, which can take several hours and is too time-consuming. Furthermore, it is only compatible with liquid matrices and the equipment is expensive (several hundred thousand euros per machine). A.5 Solid-phase cytometry methods

[0013] The solution to be analyzed is filtered through a membrane selected to retain elements at least the size of a bacterium (e.g., 0.4 µm). This is combined with labeling the bacteria using a dye. The membrane is then scanned with a laser to excite the fluorochrome, and the quantity of bacteria is counted. As an alternative to laser detection, epifluorescence imaging can be performed.

[0014] Various implementations based on this principle have been developed by companies such as Red Berries (see for example patent application EP3861125A1), Microbs, Innosieve Diagnostics (see for example patent application EP2440331A1) or Mibic.

[0015] These methods require a filterable solution and reagents that make them expensive. Furthermore, after the detection and counting of microorganisms, the reagents used can have toxic effects on cellular metabolism, which can hinder the subsequent culture and identification of these microorganisms.

[0016] Finally, these methods require significant pretreatment steps and incubation periods of at least 20 minutes, and they exhibit a high level of false negatives due to dust or impurities affecting fluorescence, or to dye absorption. A.6 Optical Microscopy Methods

[0017] Optical microscopy methods have attracted interest for the identification and enumeration of bacteria because of their potential for rapid and non-destructive detection which requires minimal sample preparation.

[0018] Characterization and enumeration of bacteria are required for a wide range of concentrations and different types of samples, which significantly complicates these methods.

[0019] Current optical methods applied to the detection of bacteria only work with a very concentrated suspension, which makes them useless in practical applications. A.7 Epifluorescence Microscopy Methods

[0020] Fluorescence microscopy is an imaging technique that occurs when a fluorophore interacts with light of a chosen energy, which leads to the excitation of a fluorophore from the ground state to a higher energy state, before returning to a lower energy level by emitting a photon.

[0021] Epifluorescence microscopy generally uses a single band of wavelengths and does not offer the ability to differentiate bacteria from other objects without the use of optimized bandpass filters. Indeed, the multiband approach requires switching from one band to another, using a dichroic wheel, which is slow.

[0022] These methods also require sample preparation with optical dyes, such as DAPI, for counting bacteria (see the article by T. Muthukrishnan et al. "Evaluating the Reliability of Counting Bacteria Using Epifluorescence Microscopy", Journal of Marine Science and Engineering. 2017; 5(1):4, https: / / doi.org / 10.3390 / jmse5010004).

[0023] The main cases where fluorescent dyes are not necessary are known in ophthalmology, where the autofluorescence of ocular cells can provide valuable information on patient-specific pathologies (see the article by A. Gakamsky et al. "Tryptophan and Non-Tryptophan Fluorescence of the Eye Enzymes Proteins Provides Diagnostics of Cataract at the Molecular Level", Scientific Reports 7, 40375 (2017), https: / / doi.org / 10.1038 / srep40375). In such cases, fluorescence microscopy can provide a rapid and reliable detection system using intrinsic fluorescence.

[0024] However, there is currently no optimized system that can distinguish objects from bacteria at this level of precision, with a proven application in microbiology. A.8 Hyperspectral Imaging Methods

[0025] Hyperspectral imaging solutions have been studied for the quantification of bacteria. These solutions generally use white light illumination in the VNIR spectrum ("Visible and Near-infrared" in English or "Visible and Near-infrared Spectrum", between 400nm and 1100nm), associated with machine learning models ("machine learning" or "ME" in English) or deep learning models ("Deep Learning" or "DE" in English).

[0026] These models are monolithic, meaning that feature extraction is performed to form the input for the ML, or that a DL is used to process hyperspectral images that may have undergone preprocessing. This makes it impossible to determine which part of the images is associated with a bacterium and therefore limits the validation possibilities during training, or requires imaging times incompatible with industrial implementation.

[0027] Two main approaches are competing: One approach is based on measuring adulteration. For example, in the article by Zheng et al., "A Nondestructive Real-Time Detection Method of Total Viable Count in Pork by Hyperspectral Imaging Technique," Applied Sciences 7, no. 3: 213, 2017, https: / / doi.org / 10.3390 / app7030213, a model that correlates the spectral signature of pork with the total viable cell count (TVC) is trained to rapidly quantify the TVC in a piece of pork. However, this measurement primarily detects product adulteration rather than the actual quantity of bacteria. It is effective for matrices damaged by bacteria but is unsuitable for low bacterial loads (e.g., less than 100 CFU / mL) or matrices such as environmental samples and milk. One approach is based on sorting bacteria according to their spectral signature on a microscope slide, as described in the article by Michael et al., "Hyperspectral imaging of common foodborne pathogens for rapid identification and differentiation," Food Science & Nutrition, 2019, 7: 2716-2725, https: / / doi.org / 10.1002 / fsn3.1131. This approach uses white light and transmittance microscopy to differentiate Salmonella spp., E. coli, and Cronobacter sakazakii. While these hyperspectral solutions allow for differentiation between bacterial species, they require significant sample preparation and high-resolution optics. Furthermore, they do not allow for the direct calculation of colony-forming units per milliliter (CFU / mL).

[0028] Beyond the problems mentioned above, methods based on hyperspectral imaging suffer from acquisition speed problems with existing hyperspectral cameras, which make them unusable in contexts of detecting low levels of bacteria. A.9 Multispectral Imaging Methods

[0029] Multispectral imaging solutions have been studied for the investigation of sample adulteration. In general, adulteration investigation methods are unsuitable for the problem addressed by the invention.

[0030] By way of example, the article by Spyrelli et al., “Implementation of Multispectral Imaging (MSI) for Microbiological Quality Assessment of Poultry Products”, Microorganisms 2020, 8, 552. https: / / doi.org / 10.3390 / microorganisms8040552, describes the capture of reflectance at several wavelengths between 405 nm and 970 nm. These measurements are processed to form input vectors for a machine learning process, which returns a TVC.

[0031] A similar approach is presented in the article by Fengou et al. "Detection of Meat Adulteration Using Spectroscopy-Based Sensors", Foods 2021, 10, 861. https: / / doi.org / 10.3390 / foods10040861. In this article, the performance of adulteration detection is validated against an ultraviolet fluorescence approach.

[0032] All these approaches have the drawback of not being able to detect low levels of bacteria. Furthermore, they are based solely on reflectance. By their very nature, they are unsuitable for detecting single bacteria, both due to the resolution of the processed image and the signal used.

[0033] A.10 Excitation and emission matrix methods

[0034] Excitation and emission matrices (EEMs) are used for multi-component analysis and provide a molecular fingerprint for different types of samples. EEM spectroscopy has been used in various applications, particularly in the wine and oil industry, to predict taste by determining specific compounds of interest (see the article by B. Quintanilla-Casas et al., "Using fluorescence excitation-emission matrices to predict bitterness and pungency of virgin olive oil: A feasibility study", Food Chemistry, Volume 395, 30 November 2022, 1336022022, https: / / doi.org / 10.1016 / j.foodchem.2022.133602).

[0035] Compared to conventional fluorometry methods, EEM spectroscopy has the advantage of not depending on concentration and of accurately identifying certain samples.

[0036] Certain applications in microbiology have, for example, enabled the detection of bacteria in water at levels as low as 10 CFU / mL (see, for example, the article by Nakar et al., "Quantification of bacteria inwater using PLS analysis of emission spectra of fluorescence and excitation-emission matrices", Water Research, Volume 169, 2020, 115197, https: / / doi.org / 10.1016 / j.watres.2019.115197). However, in the case of food, these methods are not capable of detecting anything other than adulteration, and in particular, TVCs below 100 CFU / mL.

[0037] Finally, they require calibration of tryptophan to obtain accurate results and the use of expensive equipment such as a fluorometer. B. Summary of the invention

[0038] In general, the Petri dish method and the exploratory methods described above are not satisfactory for one or more of the following reasons: implementation time, complexity, cost, quality of detection, detection threshold, ability to determine the TVC.

[0039] The invention improves the situation. To this end, it proposes, according to a first aspect, a device for detecting bacteria in a sample, comprising an optical bench including an illumination unit comprising an optical source and a distribution optic for illuminating a sample, and a measurement unit comprising a collection optic and a measurement optic, the collection optic being arranged to redirect measurement radiation from the sample in response to illumination towards the measurement optic. The optical source is arranged to emit a first radiation with a wavelength substantially equal to 365 nm, and a second beam of wavelengths within a range between 375nm and 1000nm directed to the distribution optics to illuminate the sample, the distribution optics includes a magnification objective, and the collection optics further includes one or more splitters arranged to separate the measurement beam into at least a first beam and a second beam, the first beam presenting a portion of the measurement beam having one or more wavelengths within a first range between 400nm and 414nm, the second beam presenting a portion of the measurement beam having one or more wavelengths within a second range between 414nm and 490nm.The measurement optics includes at least a first optical sensor for the first beam, and a second optical sensor for the second beam arranged to return an image in which each pixel is associated with an area of ​​the sample illuminated by the magnification objective and with a measured light amplitude value respectively.This device is arranged to control a relative displacement of the sample with respect to the illumination block in order to illuminate successive portions of the sample with the first radiation and with the second radiation and to obtain a set of multispectral images for each of the successive portions of the sample, and further includes an analyzer comprising a classifier and a computer, the classifier comprising a deep learning neural network arranged to receive the set of multispectral images associated with a portion of the sample and to return an image in which each pixel is associated with an element type identifier chosen from a list comprising at least two elements, one of which designates a bacterium, and the computer being arranged to determine bacteria detection data from the images returned by the deep learning neural network.

[0040] This device is particularly advantageous because it makes it possible to offer a device that allows for the detection of bacteria at very low detection levels, and in times compatible with an industrial production line.

[0041] According to various embodiments, this device may have one or more of the following characteristics: - said one or more beam splitters of the collection optics are further arranged to separate the measurement radiation into a third beam distinct from the first and second beams, this third beam presenting a portion of the measurement radiation having one or more wavelengths in a third range between 470 nm and 1000 nm, the measurement optics comprising a third optical sensor (44) for measuring the third beam, the analyzer being further arranged to receive the measurements from the third optical sensor along with the measurements data from the first optical sensor and the second optical sensor to derive the bacteria detection data, - the optical bench also includes an autofocus system positioned between the optical source and the magnifying lens, - The magnification lens and autofocus are included in the distribution optics and the collection optics. - The classifier's deep learning neural network is a convolutional neural network, - The classifier's deep learning neural network is a transformer, - The classifier includes at least one convolutional neural network coupled with a transformer. - The calculator is designed to determine a ratio between the area occupied by pixels associated with a bacteria identifier and a reference area, and to calculate bacteria detection data from this ratio. - the classifier is arranged to associate a pixel with an identifier of a type of element chosen from a list including a bacterium, a matrix element, an air bubble, a membrane element or a foreign element, and - the calculator is arranged to determine the reference surface from the pixels whose identifiers are associated with a membrane element.

[0042] According to a second aspect, the invention also relates to a method for detecting bacteria in a sample comprising the following operations: a) determine one or more sets of multispectral images, each set of multispectral images of a given sample portion comprising images obtained by measuring, in a first wavelength range between 400nm and 414nm and in a second range between 414nm and 490nm, an autofluorescence beam emitted by the given sample portion when illuminated by radiation with a wavelength substantially equal to 365nm, and on the other hand a reflectance beam of the given sample portion when illuminated by radiation with wavelengths in a range between 375nm and 1000nm, b) provide each set of multispectral images to a deep learning neural network to produce an image in which each pixel is associated with an element type identifier chosen from a list comprising at least two elements, one of which designates a bacterium, and c) calculate bacteria detection data from at least some of the images from operation b).

[0043] According to various embodiments, this process may have one or more of the following characteristics: - Operation a) comprises determining a set of multispectral images for a set of non-overlapping sample portions whose union covers the entire surface of the sample, - operation a) is performed for a portion of the sample, then operation b) is performed for that portion of the sample, and operations a) and b) are repeated until a measurement end condition including exceeding a threshold or determining a set of multispectral images for the entire sample is met, - Operation a) is performed for a chosen number of sample portions, then the required number of sample portions is determined by applying operation b) to the resulting multispectral image sets, then operations a) and b) are performed with a stopping condition taking into account the required number of sample portions, - the number of sample portions required is re-evaluated each time operations a) and b) are performed, and in which the stopping condition takes into account the re-evaluated number of sample portions required, and - operation c) includes determining a ratio between the area occupied by pixels associated with a bacteria identifier and a reference area, and calculating bacteria detection data from this ratio.

[0044] Other features and advantages of the invention will become more apparent from the following description, taken from illustrative and non-limiting examples shown in the drawings: - Figure [1] represents a schematic view of a device according to the invention, - [Fig.2] represents one embodiment of the optical bench of [Fig.1], - [Fig.3] represents an example of an image obtained at the output of the optical block of [Fig.2], - [Fig.4] represents a generic view of an analyzer implemented in the embodiment of [Fig.1], - Figure [5] represents a block diagram of an implementation of the analyzer from Figure [4], - Figure [6] represents an example of an image obtained as output from the classifier of Figure [5], - Figure [7] represents a diagram illustrating an operating method of the analyzer in Figure [4] according to an operational embodiment, - Figure 8 represents the results of bacterial detection obtained with the matrix that originally appeared in Figures 3 and 6. - Figure 9 represents images similar to those in Figures 3 and 6 obtained with another matrix, - [Fig. 10] represents bacterial detection results obtained with the matrix that originated [Fig. 9], and

[0045] - Fig. 11 represents bacteria detection results obtained with a analyzer using a transformer.

[0046] The drawings and the description below contain, essentially, elements of a definite nature. They may therefore not only serve to better understand the present invention, but also contribute to its definition, if necessary. C. Examples of implementation of the invention

[0047] Fig. 1 represents a schematic view of an embodiment of a device 2 according to the invention.

[0048] The device 2 comprises an optical bench 4 and an analyzer 6. The optical bench 4 comprises an illumination block 8 and a measurement block 10.

[0049] The optical bench 4 is arranged to perform multispectral imaging operations on a sample 12 obtained by grinding and depositing a sample 14 onto a membrane, for which a bacterial detection value 16 is sought. In the embodiment described here, the sample can be wheat, flour, or even a liquid reflux (for example, to test for the presence of bacteria in water). Hereafter, the sample 14 may also be referred to as the "matrix".

[0050] Upon exiting the optical bench 4, a plurality of multispectral images 18 (hereafter referred to as the multispectral image set) are transmitted to the analyzer 6 to determine bacterial detection data 16, which may be a bacterial quantity value. In some embodiments, the bacterial quantity value may be a number of colony-forming units per milliliter (CFU / mL). In other embodiments, this bacterial quantity value may be a Boolean value indicating whether a threshold of colony-forming units per milliliter (CFU / mL) has been exceeded. Although reference is made here to colony-forming units per milliliter, any other relevant CVT measurement may be used, such as colony-forming units per gram, colony-forming units per square centimeter, etc.As will be seen below, in some embodiments, the bacteria detection data 16 may be data enabling the determination of a quantity value of bacteria by a treatment or normalization.

[0051] In the embodiment of [Fig. 1], the analyzer 6 implements a machine learning model (ML). For this reason, dashed arrows are used to show the elements for training the analyzer 6, via a Petri dish 20 in which the bacteria present in the sample 14 are conventionally cultured and then counted to provide data for training the analyzer 6. Once training is complete, only the solid arrows are implemented by device 2 to calculate a detection value of bacteria 16. The operation of the analyzer 6, its drive and its embodiment variants will be described further with the embodiments of figures 4 to 7. Optical bench

[0052] Fig. 2 illustrates one embodiment of optical bench 4.

[0053] The optical bench 4 includes an illumination block 22 and a measuring block 24.

[0054] The illumination block comprises an optical source 26 and a distribution optic 28 to illuminate the sample 12. In the example described here, the distribution optics 28 include a magnifying objective 30 and an autofocus 32. As will be seen below, the magnifying objective 30 and the autofocus 32 also serve as collection optics for performing measurements. In some embodiments, the autofocus 32 may be omitted.

[0055] In the example described here, the optical source 26 is a white light source. Generally, the optical source 26 can exhibit an emission spectrum comprising wavelengths between 360 nm and 1000 nm. The reason for this choice is that the Applicant has discovered that it is possible, using a multispectral approach, to study the spectrum in autofluorescence (also called "intrinsic fluorescence") and reflectance in order to distinguish the presence of bacteria in matrices such as water, flour, wheat, or other products to be tested.

[0056] This is highly unusual. Indeed, while reflectance is an interesting signal for detecting adulteration at the macroscopic level, it is not relevant a priori when detection levels are much lower, for example, at the level of each individual bacterium. Traditionally, reflectance is ignored in microscopy in favor of sample transmittance measured by phase-contrast imaging. However, studying transmittance using phase-contrast imaging is impossible in this case.

[0057] Against all expectations, the Applicant discovered that not only do bacteria exhibit a useful signal in reflectance, but that, moreover, unlike all existing approaches, the combination of autofluorescence and reflectance makes it possible to detect bacteria individually and to exclude foreign bodies that could distort the measurements.

[0058] The invention therefore combines the non-classical use of reflectance with autofluorescence, where the only known methods - which do not address the problem targeted by the invention - at best use only one type of signal.

[0059] In the example described here, the optical source 26 is a light-emitting diode (LED) source. Alternatively, any optical source capable of emitting radiation that induces autofluorescence in the range mentioned above will be suitable, provided that the distribution optics 28 allow for concentrating The radiation on sample 12 is used to obtain a suitable size for imaging. Thus, the optical source 26 can be a laser source or any other suitable light source.

[0060] Since autofluorescence is a low-intensity phenomenon, the optical source 26 is arranged to emit autofluorescence radiation with a wavelength approximately equal to 365 nm, that is, centered on the 365 nm wavelength, which corresponds to the autofluorescence excitation of NADH (which is the reduced form of nicotinamide adenine dinucleotide). Thus, illumination radiation does not interfere with the autofluorescence measurement. With regard to reflectance, the situation is different. Indeed, reflectance can be measured over a broad spectrum, from 300 nm to more than 1000 nm. However, in order to limit the risk of photobleaching, the optical source is arranged to emit reflectance radiation with a wavelength between 375 nm and 1000 nm. Preferably, the reflectance radiation can have a wavelength between 400nm and 1000nm.

[0061] Furthermore, the multispectral approach is advantageous compared to a hyperspectral approach because it allows for a faster imaging rate for a given resolution. More specifically, the multispectral approach, with a faster acquisition rate, makes it possible, within a constant time budget, to obtain higher resolution images, and therefore to detect higher levels of colony-forming units per milliliter (CFU / mL). Thus, the 30x magnification objective in the example described here is a 20x magnification objective, which allows for images of approximately 0.5 mm x 0.5 mm.

[0062] Alternatively, other lenses could be used. Generally, the combination of the lens and cameras used results in a resolution of approximately 84 pixels for a bacterium. In general, any combination that achieves a resolution between 1 pixel and 1500 pixels (representing a magnification of approximately 100X) allows for detection with a quality and imaging time that remain compatible with the desired performance, namely the ability to perform a reliable measurement in less than one hour.

[0063] The sample 12 is received on a movable platform 33 of the optical block 4. The movable platform 33 allows the sample 12 to be moved in an XY plane relative to the measuring block 22 and the measuring block 24, thus illuminating and measuring a different portion of the sample 12 each time. The autofocus 32 can also be controlled according to the movement of the platform 33 in order to perform the Z-axis focus adjustment. As will be seen below, the movement by the movable platform 33 can be controlled by the analyzer 6. Alternatively, the movable platform 33 can be controlled by a control module of optical bench 4. As an alternative, it is objective 30 which could be mobile relative to sample 12.

[0064] The measuring block 6 includes a collection optic 34 and a measuring optic 36. The collecting optic 34 has the role of recovering the autofluorescence and reflectance radiation from the sample 12 and redirecting it to the measuring optic 36.

[0065] As mentioned above, the optical bench 4 is designed in a compact manner, and the magnifying lens 30 and the autofocus 32 serve both the distribution optics 28 and the collection optics 34. Thus, the light beam from the light source 26 is focused and its size is reduced on the sample 12 by the lens 30 and the autofocus 32, and the radiation which comes from the sample 12 passes back through the lens 30 and the autofocus 32 which restores its initial size for the optical measurement.

[0066] In the example described here, the collection optics 34 further include three separators 38 and three lenses 39 which each redirect a portion of the radiation from the sample 12 towards the measuring optics 36.

[0067] Thus, a first separator 38 comprises a dichroic mirror which isolates a portion of the radiation from the sample 12 with wavelengths between 400 nm and 425 nm, and preferably between 400 nm and 414 nm, to form a first beam. This dichroic mirror directs the remainder of the radiation from the sample 12 towards a second separator 38.

[0068] Similarly, the second beam splitter 38 includes a dichroic mirror 38 which isolates a portion of the radiation from the sample 12 with wavelengths between 414 nm and 490 nm, and preferably between 414 nm and 470 nm, to form a second beam. This dichroic mirror directs the remainder of the radiation from the sample 12 to a third beam splitter 38.

[0069] Finally, the third separator 38 recovers the remaining radiation, which therefore has wavelengths beyond 470 nm, to form a third beam. Alternatively, the third separator 38 may include a dichroic mirror that selects a portion of the radiation from the sample 12 with wavelengths preferably between 470 nm and 520 nm to form the third beam.

[0070] The first beam, the second beam and the third beam are each received by a respective lens 39 of the collection optics 34.

[0071] The measuring optics 36 in the example described here comprises a first camera 40, a second camera 42 and a third camera 44. Each lens 39 redirects respectively the first beam to the first camera 40, the second beam to the second camera 42 and the third beam to the third camera 44.

[0072] The first camera 40, the second camera 42 and the third camera 44 are arranged to measure spectral bands which correspond to the wavelength range of respectively the first beam, the second beam and the third beam.

[0073] In certain embodiments, the third separator 38 may be omitted, and the analyzer 6 will operate on the measurements of the first and second beams only. In this case, as will be seen below, instead of two image triplets, the output of the cameras 40 to 44 will then be two pairs of images (one pair for the autofluorescence measurement and another pair for the reflectance measurement).

[0074] In the example described here, the first camera 40, the second camera 42, and the third camera 44 are Phoenix 8.1 MP Model (IMX566), and the resulting image from each shot has a resolution of 1420 x 1420, as shown in [Fig. 3]. Other camera models or optical sensors can, of course, be used, as well as other resolutions for the corresponding images. However, it must be taken into account that the diffraction limit sets a maximum useful resolution.

[0075] In each image, each pixel corresponds to an area of ​​the sample 12 that has been illuminated. The value associated with each pixel corresponds to the sum of the intensities measured for the wavelengths of the camera considered at that location. To represent the value associated with each pixel, a color can be used that depends on the sum of the measured intensities.

[0076] Thus, in operation, the optical source 26 will successively illuminate portions of the sample 12. The illumination of each portion of the sample 12 is carried out in two stages.

[0077] A first illumination of approximately 1 second at a wavelength of 365 nm is used to elicit the autofluorescence radiation from sample 12, which is measured in three beams by the three cameras 40 to 44, resulting in three images of the sample portion similar to that in [Fig. 3]. Then, a second illumination of approximately 50 ms is performed at approximately 405 nm, and the reflectance radiation from sample 12 is measured in three beams by the three cameras 40 to 44, resulting in three further images of the sample portion similar to that in [Fig. 3].

[0078] These sextuplets of images (two sets of three images respectively in autofluorescence and reflectance) are processed by the analyzer 6 in order to determine bacteria detection data 16, thus returning a bacteria detection value. As mentioned above, in the case where only two bands are measured, the analyzer 6 will be configured to process quadruplets of images.

[0079] In the example described here, autofluorescence and reflectance are measured by the same cameras because the optical source 26 includes a filter to produce the autofluorescence radiation, which is centered at 365 nm and is slightly permeable to radiation around 405 nm. Thus, the reflectance measurement can be performed by cameras 40 to 44 by significantly increasing the amplitude of the optical source 26 for the reflectance measurement.

[0080] Alternatively, the optical source 26 could include a mechanism alternating a 365nm filter with a 405nm filter between the autofluorescence and reflectance measurements. Also alternatively, off-axis illumination with white light of a wavelength greater than 400nm could be used to perform the reflectance measurement. In all cases, two image triplets (or two pairs of images) are acquired, one for autofluorescence and the other for reflectance. These two image triplets (or two pairs of images) will hereafter be referred to as "image sextuplets" ("image quadruplets" in the case of two bands, and therefore two pairs), or even "multispectral image set". C.2 Analyzer

[0081] The analyzer 6 and its processing of measurements from cameras 40 to 44 will now be described with reference to [Fig.4].

[0082] The analyzer 6 includes a data storage 100 and a calculator 102.

[0083] Data storage 100 receives sample image data, data from deep neural network, and bacteria detection data.

[0084] The sample image data is associated with coordinates indicating which part of sample 12 it originates from, as well as with camera data indicating which of the first, second, or third beams it is associated with. As explained above, the sample image data comprises pixels, each of which is associated with an intensity value representing the sum of the intensities measured by the camera for the location of sample 12 designated by the coordinates and the pixel in question. Additionally, a supplementary camera could be used to detect the absolute position of sample 12. Alternatively, the sample holder 12 can be adapted for the same purpose.

[0085] Data storage 100 can be any type of data storage suitable for receiving digital data: hard drive, hard drive with flash memory, flash memory in any form, RAM, optical disc, locally distributed storage or in the cloud, etc.

[0086] The computer (PC) 102 may include one or more processors (P) 104, a network interface 106 comprising a transmitter (Tx) 108 and a receiver (Rx) 110 to enable the device to transmit and / or receive data to other computing devices connected to a network 112 (for example, an IP network) to which network interface 106 is connected.

[0087] In certain embodiments in which the computer 102 includes a programmable processor, a computer program product (CPP) 116 may be provided to implement the phase correction processing. The CPP 116 stores a computer program (CP) 118 that includes computer readable instructions (CRI) 120. The CPP 116 may be stored on a computer readable medium (CRM) 122, which may be a non-transient computer readable medium, such as magnetic media (hard drive, SSD, magnetic tape, etc.), optical media (CD, DVD, Blu-ray disc, etc.), memory (RAM, flash memory, etc.), distributed or cloud storage, etc.Computer-readable media (CRM) 122 can also be stored in data storage 100.

[0088] The P 104 processor(s) can be any processor suitable for the calculations described below. Such a processor can be implemented in any known form, such as a microprocessor for a personal computer, laptop, tablet, or smartphone; a dedicated digital signal processor (DSP); a dedicated FPGA or SoC chip; a computing resource on a grid or in the cloud; a graphics processing unit (GPU); a microcontroller; or any other form suitable for providing the computing power necessary for the implementation described below. One or more of these elements can also be implemented as specialized electronic circuits such as an ASIC. A combination of processor(s) and electronic circuits can also be considered. Processors dedicated to machine learning may also be considered.

[0089] In the exploratory methods described in the introduction, particularly in some of the multispectral or hyperspectral methods, a deep learning neural network is used to directly extract the value of colony-forming units per millilitre (CFU / mL), from a single shot of the sample 12.

[0090] The Applicant explored several ways of processing the sample image data and determined that the use of a deep-learning neural network on a single image is not desirable.

[0091] On the contrary, the Applicant has discovered that, in order to be able to calculate a bacterial detection value usable on an industrial scale, i.e. on the order of 1 CFU / mL, it is desirable to have a spatial resolution of High-level data collection requires more than 100 exposures for a sample of approximately 500 mm². Advantageously, concentrating the light from the optical source also increases the surface intensity of illumination and thus the corresponding autofluorescence response.

[0092] It should first be noted that this number of images makes the use of hyperspectral cameras impossible because the acquisition time for each image is 2.5 minutes. But beyond that, the Applicant realized that the lower the desired level of colony-forming units per milliliter (CFU / mL), the less reliable direct detection by a deep learning neural network becomes. Indeed, with increasing resolution, the information about the presence of bacteria becomes diluted, or even spread across several images, making it difficult for the neural network to learn reliably.

[0093] The Applicant thus discovered that it is preferable to carry out the detection in two operations.

[0094] For this reason, as shown in [Fig.5], the analyzer 6 includes a classifier 50 which includes a deep learning neural network which receives a set of multispectral images as input and returns an image 52 as output, and a computer 54 which receives the image 52 as input and returns the detection data 16.

[0095] It goes without saying that the distinction between the classifier 50 and the calculator 54 is essentially functional, but that the analyzer 6 can be seen as made of a single block.

[0096] Thus, a first detection operation implemented by the classifier 50 consists of associating, from all the images concerning a given area of ​​sample 12, a label with each pixel of that area. Indeed, sample 12 is typically composed of a membrane and the atomized sample 14. Consequently, the optical measurement can therefore, for each pixel, correspond to the following natures: membrane material, matrix element, a bacterium, an air bubble, or a foreign element such as dust. In its simplest form, the label can be reduced to "bacteria" or "other element".

[0097] In other words, each pixel receives a value which indicates whether it is considered that this pixel belongs to the membrane of the sample 12, to a bacterium, to a matrix element which is the subject of the measurement of detection of bacteria, to an air bubble, or to a foreign element.

[0098] Alternatively, the deep learning neural network can return a vector indicating a probability for each pixel of belonging to a given category (for example, [0,4; 0,6] would indicate that there is a 40% chance that a given pixel belongs to a bacterium and a 60% chance that this is not the case in the case of a binary label described above). In this case, the first modulo 50 can be arranged to To determine a unique label for each pixel, either by choosing the label with the highest value, or by partially using the values ​​of neighboring pixels, etc. Alternatively, the probability of belonging to the "bacteria" category could be taken into account in the subsequent calculation of bacteria detection data.

[0099] Fig. 6 represents an example of image 52 obtained after application of the labels from the measurements shown on Fig. 3.

[0100] The Applicant tested several deep learning neural networks for this task, and found that convolutional neural networks ("CNN"), recurrent neural networks ("RNN") and transformers or self-attentive models ("Transformers") are useful for performing this first operation.

[0101] To train the deep neural network, multispectral image sets were measured for numerous samples of each matrix, and each pixel of each sextuplet of images was annotated to indicate its corresponding label. By each pixel of each sextuplet, it should be understood that a label is associated with a pixel coordinate for six images (the two triplets) corresponding to the same shot.

[0102] The Applicant has tested several of these CNN models, including Unet, Unet++, MAnet, Linknet, FPN, PSPNet, PAN, DeepLabV3, and DeepLabV3+. Access to these CNNs is available at https: / / web.archive.org / web / 20231021175139 / https: / / smp.readthedocs.io / en / latest / . Among the transformers, SegViTv2 (see https: / / arxiv.org / pdf / 2306.06289.pdf), Segmenter (see [link missing]), and SegFormer (see https: / / huggingface.co / docs / transformers / model_doc / segformer) have also been identified. The Applicant has also determined that a combination of CNN, RNN, and / or transformer can be useful.

[0103] A second part of the detection is carried out by the computer 54 and consists of quantifying the presence of bacteria on the basis of the images 52 on which the pixels have been labeled. This can be done in various ways.

[0104] According to one embodiment of the invention, the pixel area associated with a bacterium is calculated and accumulated image by image to obtain a total area occupied by the bacteria on the sample 12, which makes it possible to detect a TVC value by relating it to the detected membrane area. To determine the TVC from the ratio between the area occupied by the bacteria and the area occupied by the membrane, cultures in Petri dishes were carried out to measure the actual TVC for these samples, from standardized sampling quantities 14.

[0105] The determined TVC value obviously depends on the quantity of sample 14 that was used to prepare the sample 12. Therefore, if the quantity of sample 14 is not normalized, then the analyzer 6 may receive as input The value in grams, liters, or square centimeters allows the TVC value to be converted to CFU / g, CFU / mL, or CFU / cm². Alternatively, this standardization can be carried out later.

[0106] By "frame by frame," we mean the images obtained after processing by the first operation above, each corresponding to a given portion of the sample that has been illuminated and whose autofluorescence and reflectance response has been measured. As suggested above, the cumulative TVC value can also be compared to a threshold, and the operation can stop as soon as the threshold has been crossed or the entire sample 12 has been illuminated.

[0107] This type of treatment is illustrated in [Fig.7], which shows an operating loop of device 2.

[0108] This loop begins with an operation 60 in which a set of multispectral images of a portion of the sample 12 is obtained by imaging with the optical bench 4.

[0109] Next, in an operation 62, the analyzer 6 calculates a ratio of bacteria surface area to membrane surface area, and a bacteria detection value is calculated in an operation 64.

[0110] In operation 64, a bacteria detection value can be obtained by determining a sample portion CTV for the multispectral image set of operation 60; this sample portion CTV can then be summed over all sample portions. Alternatively, the bacteria detection value can be re-evaluated at each sample portion; that is, the bacterial surface area to membrane surface area ratio for the current measurement is updated at each iteration of the loop, and the CTV is evaluated from this total ratio (up to the current loop).

[0111] Next, in an operation 66, it is determined whether the loop should continue. If a threshold is detected, this involves comparing the result of operation 64 with that threshold. Otherwise, this operation consists of determining whether there is another portion of the sample to be illuminated. Finally, the loop ends in an operation 68 with the sending of the bacteria detection data thus determined.

[0112] Fig. 8 shows the performance of the measurement carried out by the loop of the Fig. 7 method.

[0113] Fig. 9 shows respectively the images corresponding to Fig. 3 (left) and Fig. 6 (right) obtained with wheat as the matrix, and Fig. 10 shows the performance of the measurement carried out by this method with wheat as the matrix.

[0114] The Applicant has further discovered that it is possible to speed up the measurement by reducing the number of shots taken.

[0115] More specifically, in a particular embodiment, 5 non-overlapping images (for example) are obtained to initialize the measurement loop. Based on these initial values, a number of shots allowing the TVC value to be calculated with an accuracy of ±0.5 LOG.

[0116] The table below represents numbers of shots as a function of TVC values. TVC Value Shots 0 1960 1 1846 2 212 3 3 4 2 5 2 6 2 7 2

[0117] In a first variant, a loop similar to that of [Fig.7] may be implemented, but with a fixed number of shots taken from the table above as a stopping condition.

[0118] In a second variant, the loop may be slightly modified by evaluating the number of shots at each loop iteration based on the evolution of the TVC value of the previous loop, and by defining a loop termination condition based on the comparison between the updated number of shots needed and the number of loops already performed and / or exceeding a TVC value threshold where applicable.

[0119] The Applicant's work suggests that there is no preferred path for taking photographs. Nevertheless, in order to limit any risk, it is preferable to traverse the sample by selecting photographs that are substantially evenly distributed among them. For example, it is possible to define non-overlapping zones, take a measurement in each zone, and then repeat the process.

[0120] In other variants, some of the 52 images could be ignored.

[0121] Thus, the device described above makes it possible to carry out a measurement of TVC with high precision in very low detection ranges, with a measurement time which is less than 2 hours, and is regularly less than 10 minutes.

[0122] As mentioned above, the Applicant has also experimented with the use of transformers as a deep learning neural network. Indeed, despite their qualities demonstrated above, CNNs exhibit narrow effective receptive fields.

[0123] The transformers introduce a self-attention mechanism that is not local and that can perform operations combining distant pixels in the image. This significantly increases the effective receptive field and improves the model's performance for a given computing power.

[0124] To validate this theory, different machine learning models were tested on a dataset of 56 images (10 for validation and 46 for training). In this dataset, 11 samples contained only wheat and 45 samples contained only bacteria. These samples were annotated at the pixel level.

[0125] The performance of transformer models of different sizes and that of a convolutional neural network model. The transformer used is of the Segformer type (b0, bl, b2, b3 - b0 being the one requiring the least computing power and b3 the most), while the convolutional neural networks used are DeeplabV3+ coupled with Resnet 101. In the latter case, the DeeplabV3+ neural network is used as the "backbone" to extract a set of features which is processed by the Resnet neural network used as the "head" to determine the TVC value. The backbone transforms the image into features which are then used by the head which, in classic applications, will either perform classification, regression, object detection, or, in the example described here, semantic segmentation (one class per pixel).

[0126] It is also possible to combine a convolutional or recurrent neural network with a transformer. In this case, the transformer will always be used as the "head", while the other neural network will be used as the "backbone".

[0127] The models are trained on an Nvidia T4 GPU with 4 Intel Cascade Lake virtual CPUs and 16 GB of RAM.

[0128] This comparison comprises two parts:

[0129] 1. the intersection over the union ("loU" or "Intersection over Union" in English) on the validation set (the 10 images mentioned above)

[0130] 2. the end-to-end prediction of the total number of viable bacteria (TVC) on Five unused training samples, each composed of five images mixing wheat and bacteria (remembering that the model was trained only on pure samples), were used. The total number of viable bacteria obtained was then compared with that measured in the laboratory using Petri dishes in terms of absolute error and precision. Precision was defined as the number of samples for which the absolute error was less than 0.5.

[0131] The table below summarizes the results of the evaluations of the first part: Model 1 Best I oU Iterations per minute Accuracy of T VC Average error of TVC SegFormer bO 69.88 207 60% 0.30 SegFormer b 1 70.42 184 80% 0.33 SegFormer b2 70.49 157 80% 0.23 SegFormer b3 70.26 88 100% 0.12 DeeplabV3+ Resnet 101 69.72 22 100% 0.24

[0132] The results show that the use of transformation models has a definite advantage.

[0133] - All transformer models are more efficient than the CNN model in intersection terms on the union (loU) on the validation set.

[0134] - All transformers are faster than the CNN in terms of number of iterations per minute.

[0135] - SegFormer b2 and b3 are more efficient than the CNN in terms of TVC error average.

[0136] - SegFormer b3 equals CNN in terms of TVC accuracy.

[0137] Figure 11 shows the results of the evaluations of the second part. This figure shows in particular the results obtained using the SegFormer b3 transformer, as well as their comparison with the Petri dish measurements.

Claims

1. Demands A device for detecting bacteria in a sample, comprising an optical bench (4) including an illumination block (22) comprising an optical source (26) and a distribution optic (28) for illuminating a sample, and a measurement block (24) including a collection optic (34) and a measurement optic (36), the collection optic (34) being arranged to redirect measurement radiation from the sample (12) in response to illumination towards the measurement optic (36), characterized in that the optical source (26) is arranged to emit a first radiation with a wavelength substantially equal to 365nm, and a second radiation with wavelengths in a range between 375nm and 1000nm towards the distribution optic (28) in order to illuminate the sample (12), the distribution optic (28) including a magnification objective (30),the collection optics (34) further includes one or more splitters (38) arranged to separate the measurement radiation into at least a first beam and a second beam, the first beam having a portion of the measurement radiation having one or more wavelengths in a first range between 400nm and 414nm, the second beam having a portion of the measurement radiation having one or more wavelengths in a second range between 414nm and 490nm, the measurement optics (36) including at least a first optical sensor (40) for the first beam, and a second optical sensor (42) for the second beam arranged to return an image in which each pixel is associated with an area of ​​the sample (12) illuminated by the magnifying lens (30) and with a measured light amplitude value respectively,the device (2) being arranged to control a relative displacement of the sample (12) with respect to the illumination block (22) in order to illuminate successive portions of the sample (12) with the first radiation and with the second radiation and to obtain a set of multispectral images for each of the successive portions of the sample (12), the device (2) further comprising an analyzer (6) comprising a classifier (50) and a computer (54), the classifier (50) comprising a deep learning neural network arranged to receive the set of multispectral images associated with a portion of the sample (12) and to return an image (52) in which each pixel is associated with an element type identifier chosen from a list comprising at least two elements, one of which designates a bacterium, and the computer (54) being arranged to determine bacterium detection data (16) from the images returned by the deep learning neural network.

2. Device according to claim 1, wherein said one or more separators (38) of the collection optics (34) are further arranged to separate the measurement radiation into a third beam distinct from the first beam and the second beam, this third beam having a portion of the measurement radiation having one or more wavelengths in a third range between 470nm and 1000nm, the measurement optics (36) comprising a third optical sensor (44) for measuring the third beam, the analyzer (6) being further arranged to receive the measurements from the third optical sensor (44) with the measurements from the first optical sensor (40) and the second optical sensor (42) to derive the data for detecting bacteria (16).

3. Device according to claim 1 or 2, wherein the optical bench (4) further comprises an autofocus (32) disposed between the optical source (26) and the magnifying lens (30).

4. device according to claim 3, wherein the magnifying lens (30) and the autofocus (32) are included in the dispensing optics (28) and in the collecting optics (34).

5. Device according to any one of the preceding claims, wherein the deep learning neural network of the classifier (50) is a convolutional neural network.

6. Device according to any one of claims 1 to 4, wherein the deep learning neural network of the classifier (50) is a transformer.

7. Device according to any one of claims 1 to 4, wherein the classifier (50) comprises at least one convolutional neural network coupled to a transformer.

8. Device according to any one of the preceding claims, wherein the calculator (54) is arranged to determine a ratio between the area occupied by pixels associated with a bacteria identifier and a reference area, and to calculate bacteria detection data (16) from this ratio.

9. Device according to any one of the preceding claims, wherein the classifier (50) is arranged to associate a pixel with an identifier of element type selected from a list comprising a bacterium, a matrix element, an air bubble, a membrane element or a foreign element.

10. Device according to claim 8 in combination with claim 9, wherein the computer (54) is arranged to determine the reference surface from pixels whose identifiers are associated with a membrane element.

11. A method for detecting bacteria in a sample comprising the following operations: a) determining one or more sets of multispectral images, each set of multispectral images of a given sample portion comprising images obtained by measuring, in a first wavelength range between 400nm and 414nm and in a second range between 414nm and 490nm, an autofluorescence beam emitted by the given sample portion when illuminated by radiation of wavelength substantially equal to 365nm, and on the other hand a reflectance beam of the given sample portion when illuminated by radiation of wavelengths in a range between 375nm and 1000nm,b) provide each set of multispectral images to a deep learning neural network to produce an image (52) in which each pixel is associated with an element type identifier chosen from a list comprising at least two elements, one of which designates a bacterium, and c) compute bacterium detection data from at least some of the images from operation b).

12. A method according to claim 11, wherein operation a) comprises determining a set of multispectral images for a set of non-overlapping sample portions whose union covers the entire surface of the sample.

13. A method according to claim 11, wherein operation a) is performed for a portion of the sample, then operation b) is performed for that portion of the sample, and operations a) and b) are repeated until a measurement end condition including exceeding a threshold or determining a set of multispectral images for the whole sample is met.

14. A method according to claim 11, wherein operation a) is performed for a chosen number of sample portions, then a required number of sample portions is determined by applying operation b) to the resulting multispectral image sets, and then operations a) and b) are performed with a stopping condition taking into account the required number of sample portions.

15. A method according to claim 14, wherein the number of sample portions required is re-evaluated each time operations a) and b) are carried out, and wherein the stopping condition takes into account the re-evaluated number of sample portions required.

16. A method according to any one of claims 11 to 15, wherein operation c) comprises determining a ratio between the area occupied by pixels associated with a bacteria identifier and a reference area, and calculating bacteria detection data (16) from this ratio.