Method for determining the susceptibility of a microorganism to an antimicrobial agent from a spectral image

Abstract

The invention relates to a computer-implemented method for characterizing a microbial strain, comprising the following steps: - obtaining a spectral image of a colony, representing a colony of a microbial strain, the spectral image of the colony comprising several pixels, each pixel comprising several spectral channels; - filtering the spectral image of the colony to obtain a "test image", each pixel of the test image being a function of an average of a packet of neighboring pixels of the colony image, each packet comprising M

Description

Title of the invention: Method for determining the susceptibility of a microorganism to an antimicrobial agent from a spectral image. Technical field

[0001] The invention relates to the field of microbiological analysis, and in particular to the characterization of microorganisms, notably the prediction of the sensitivity or resistance of yeasts, molds, and bacteria to an antimicrobial agent. More specifically, the invention applies to the analysis of a spectral image of one or more colonies of bacteria, molds, or yeasts grown in or on an observable culture medium. STATE OF THE ART

[0002] In the field of in vitro diagnostics of microorganisms, particularly pathogens, the characterization of a microorganism primarily consists of identifying its species and its susceptibility to an antimicrobial agent (or "antibiogram") in order to determine a treatment for the patient infected by that microorganism. To do this, a complex microbiological process is typically implemented in the laboratory, a process that most often requires prior knowledge of other properties of the microorganism, notably its kingdom (e.g., yeast or bacteria), and, in the case of bacteria, its Gram stain type or whether it is fermentative or not. Such a process involves a significant number of manual steps (for example, for Gram stain determination, fixation, staining, mordanting, washing, overstaining, etc.), and is therefore time-consuming to implement.

[0003] To reduce human intervention, spectral imaging in the visible and near-infrared has been used, reducing the number of samples and mass spectrometry or biomolecular analyses downstream of the culture stage by classifying colonies into distinct groups, particularly at the species level, or at lower levels of classification in the phylogenetic tree, notably for typing or detecting these clonal groups. Documents WO2019 / 122732 and WO2023 / 094775 describe such processes. It is known that hyperspectral and multispectral imaging in the visible and near-infrared allows microbial characterization on Petri dishes with correct classification rates exceeding 90%. Since microbial characterization is a highly sensitive application, particularly in terms of human, animal, and environmental health, ever-increasing performance is essential. Description of the invention

[0004] The invention makes it possible to predict the susceptibility of a microorganism to an antimicrobial agent using hyperspectral or multispectral imaging of a microbial colony grown on a culture medium without the presence of said antimicrobial.

[0005] To this end, the invention proposes a computer-implemented method for characterizing a microbial strain, comprising the following steps:

[0006] - obtaining a spectral image of a colony, representing a colony of a microbial strain, the spectral image of the colony comprising several pixels, each pixel comprising several spectral channels;

[0007] - filtering processing of the spectral image of the colony to obtain an "image of the test image, each pixel of the test image being a function of an average of a packet of neighboring pixels from the colony image, each packet comprising M <N pixels avec N le nombre de pixels de l’image de la colonie ;

[0008] - classification of each pixel of the test image predicting whether the spectrum of pixel belongs to a class of a prediction model with at least two classes of microbial strain characterization;

[0009] - determination of the class characterizing the microbial strain by a vote majority across all predicted classes for each pixel;

[0010] - determination of the susceptibility of the microbial strain to a microbial agent as being that associated with the class characterizing the microbial strain.

[0011] According to one embodiment, the filtering consists of grouping the N pixels of the colony image into packets of M <N pixels voisins, un pixel de l’image de test étant la moyenne des M pixels voisins de l’image de la colonie, l’image de test présentant un nombre de pixels inférieurs à celui de l’image de la colonie.

[0012] According to one embodiment, the filtering process consists of applying a convolutional filter to the colony image, a pixel of the test image being obtained by processing several neighboring pixels of the colony image using a convolutional kernel, the test image comprising N pixels, each pixel corresponding to an average of M <N pixels voisins de l’image de la colonie, l’image de test présentant un nombre de pixels identique à celui de l’image de la colonie.

[0013] According to one embodiment, M is between 2 and 10, preferably equal to 5.

[0014] According to one embodiment, the threshold for the majority vote is between 60% and 80%, in particular 60% or 80%.

[0015] According to one embodiment, a prediction model is a model with at least three classes, each class corresponding to a microbial strain of a type, each strain being previously characterized as susceptible or not to a microbial agent, the susceptibility or resistance of the microbial strain to a microbial agent being that associated with the microbial class.

[0016] According to one embodiment the prediction model is a 2-class model, each class indicating susceptibility or not to a microbial agent, the susceptibility of the microbial strain to a microbial agent being the result of the step of determining the class characterizing the microbial strain.

[0017] The method of the invention may include training the prediction model, the training comprising the following steps:

[0018] - obtaining spectral images of colonies for several microbial strains different;

[0019] - training of classification on spectral images of different colonies.

[0020] According to one embodiment, the spectral image comprises for each pixel 240 channels spaced 2.1nm apart over a band between 390 nm and 900 nm, the spectral image being hyperspectral.

[0021] According to one embodiment, the spectral image comprises for each pixel 23 channels spaced 10 nm apart over a spectral band between 390 nm and 900 nm, the image being multispectral.

[0022] According to one embodiment, the microbial strain is a strain of Staphylococcus aureus and the antimicrobial agent is methicillin.

[0023] According to one embodiment, the microbial strain is obtained from a sample taken from a patient and in which the process includes a step of selecting an antimicrobial therapy based on the determination of the susceptibility of the microbial strain and administering said therapy to the patient.

[0024] The invention also relates to a computer program product comprising code instructions for the execution of a process according to the invention.

[0025] The invention also relates to a method for learning a prediction model implemented in the characterization method according to the invention, comprising the following steps:

[0026] - creation of a training database comprising for each strain of a species, its susceptibility to the microbial agent as well as the corresponding colony spectra;

[0027] - supervised training of the prediction model so as to obtain parameters of the prediction model on spectra of different colonies.

[0028] Filtering the spectral image improves performance on classification, which is therefore more robust.

[0029] Furthermore, spectral imaging, preferably between 390 nm and 900 nm, contains sufficient information to predict that two microbial strains are clonal or derived from the same lineage and thus share the same susceptibility to the antimicrobial agent. By knowing the susceptibility of a class, by predicting that a strain belongs to the class, the strain to be characterized is predicted to have the susceptibility of the class. PRESENTATION OF THE FIGURES

[0030] Other features, purposes and advantages of the invention will become apparent from the following description, which is purely illustrative and not limiting, and which should be read in conjunction with the accompanying drawings on which:

[0031] - Fig. 1 illustrates a system for characterizing a microbial strain conforming to the invention;

[0032] - Figure [Fig. 2] illustrates a spectral imaging system according to a first mode of realization ;

[0033] - Figure [Fig. 3] illustrates a spectral imaging system according to a second mode of realization ;

[0034] - Figure 4 illustrates steps in a process for characterizing a strain according to a method of implementing the invention;

[0035] - Figure 5 illustrates obtaining a test image from an image of a colony according to a first variant of a filtering implemented in the characterization process according to the invention;

[0036] - Figure 6 illustrates obtaining a test image from an image of a colony according to a second variant of a filtering implemented in the characterization process according to the invention;

[0037] - [Fig.7] shows strains used by prediction models.

[0038] In all figures, similar elements bear identical references. DETAILED DESCRIPTION

[0039] System

[0040] Fig. 1 illustrates a system S for characterizing a microbial strain comprising one or more colony(ies) of yeast or bacteria grown on an agar poured into a Petri dish.

[0041] Such a system S comprises a device 1, 1' for acquiring spectral images of a microbial strain 2 preferably contained in a Petri dish 21 and a processing unit 3.

[0042] The processing unit 3 is connected to the spectral image acquisition device 1,1' by wire or wireless means. It comprises one or more processors configured to optionally control the acquisition device, and one or more data storage units storing instructions to implement steps of a process for characterizing a microbial strain based on acquired spectral image(s) described below.

[0043] The processing unit 3 is also connected to one or more servers 4 which store one or more database(s). A database can store spectral images acquired for processing, training data, classification data, etc.

[0044] The processing unit 3 can be associated with a user interface and a screen.

[0045] Imaging device

[0046] In relation to [Fig. 2], the spectral imaging device 1 is, in a first embodiment, of the hyperspectral type. It advantageously comprises:

[0047] - a so-called hyperspectral camera 18, consisting of a digital sensor comprising an array of elementary sensors, for example a digital sensor of type CCD or CMOS, sensitive in a range of wavelengths for example [4in; 4rax] = [400 nm ; 900 nm]' and a light dispersive element or a spectrograph to select a wavelength to be acquired by the sensor;

[0048] - a lens 20 for focusing the image onto the digital sensor of the camera 18 optics of strain 2 from which we seek to acquire a hyperspectral image;

[0049] - a front lighting 24, for example consisting of one or more halogen lamps, for example two or four lamps, capable of emitting light in the range [2^; 2max] and for achieving uniform front lighting of strain 2. For example, the lights are white light lamps;

[0050] - a rear light 26, for example consisting of a matrix of LEDs with light white, to achieve, depending on the case, a uniform rear lighting of strain 2 in the range;

[0051] - a trolley 28 on which rests the Petri dish 21 containing the strain 2 and allowing the latter to pass in front of lens 20 in order to obtain a full image by scanning.

[0052] It is noted that the rear lighting is used with or instead of the front lighting to enable transmission acquisition from camera 18.

[0053] The device 1 is, for example, configured to acquire the image of a 90 mm by 90 mm region with a sampling step of 160 micrometers (estimated spatial resolution of 300 micrometers) and with a spectral resolution of a few nanometers over the range [2 min; 2 max]. More than 200 channels can be acquired over a range of approximately 500 nm. In particular, the field of view and depth of field of the objective 20 are chosen to obtain images that can include complete colonies with a radius of up to 1 cm, preferably up to 0.9 cm, and even more preferably 0.5 cm.

[0054] Device 1 thus produces a digital HSI (Hyper Spectral Imaging) image of the light reflected by sample 2, improperly called a "hypercube" because it is in fact three-dimensional: two spatial dimensions and one spectral dimension, each pixel (or rather voxel due to the three-dimensional nature of the HSI image) representing the radiance measured at a point in the sample 22 for a spectral channel. A pixel thus contains a spectrum.

[0055] The radiance of a pixel, commonly called "light intensity", corresponds here to the amount of light incident on the surface of the corresponding elementary sensitive site of the camera sensor 18 during the exposure time, as is known in itself from the field of digital photography for example.

[0056] In relation to Figure 3, the spectral imaging device 1' is in a second embodiment of the multispectral type and comprises a camera 34, advantageously a high spatial resolution CMOS or CCD camera, coupled to a set of spectral filters 36, for example arranged in front of the lens 20 between the lens 20 and the sensor of the camera 32. The set of filters 36 consists of a number NF of separate bandpass filters, each configured to transmit only light in a part of the range [2min; 2max], with a spectral width at half maximum (FWHM for "full width half maximum") less than or equal to 10 nm, 25 nm, 40 nm or 50 nm, and preferably less than or equal to 20 nm.Assembly 36, for example, is a filter wheel that can accommodate several different filters, a wheel controlled by the data processing unit which operates it to scroll the filters past the camera and trigger an image capture for each of them.

[0057] The device 1' thus produces a digital MSI (for Multi Spectral Imaging) image of the light reflected by the sample 2.

[0058] Characterization method

[0059] A method for characterizing a microbial strain is described in relation to [Fig.4],

[0060] In a step of strain preparation (step 11), a Petri dish 21 is inoculated with a biological sample, for example, taken from a patient (urine, blood, or more generally any biological fluid), so as to grow bacterial colonies on the surface of a nutrient, or "culture," medium deposited in the Petri dish. The nutrient medium's main purpose is to promote colony growth and, optionally, to enhance the accuracy of the characterization by limiting light disturbances. This type of growth is well known to those skilled in the art and will not be described in further detail.

[0061] Advantageously, the preparation is carried out so that there is at least one area of ​​the Petri dish on which the colonies are spaced apart from each other so as to be easily separable on the image which will be acquired and therefore their segmentation.

[0062] Once colony growth is complete, for example after 16h, 24h, 36h or 48h, the Petri dish is opened for analysis. This is referred to as the microbial strain to be characterized.

[0063] The box 2 is then positioned on the carriage 28 of the imaging device (step E2) and at least one spectral image of the box is acquired 2 and stored in the processing unit 3 (step E3).

[0064] Depending on the case, we have a hyperspectral image I_HSI acquired with the imaging device 1 of [Fig.2] or a multispectral image I_MSI acquired with the imaging device 1' of [Fig.3].

[0065] The acquired image is advantageously pre-processed to filter fluctuations in the emission spectrum of the lighting sources 24, 26. For example, a "fiat field correction" (FFC) type correction is implemented to obtain the reflectance, which also has the advantage of correcting pixel-to-pixel sensor response dispersions (step E4).

[0066] In the case of hyperspectral imaging, each pixel of the image, and therefore each spectrum, comprises 240 channels over the wavelength range between 400 nm and 900 nm with a step size of 2.1 nm. Such a resolution is considered high, which then allows, by grouping, the simulation of the spectral width of commercially available color filters. For example, the 10 nm width for the multispectral image is simulated by averaging three contiguous channels of a hyperspectral image.

[0067] In the case of multispectral imaging, each pixel comprises several channels with widths of 10, 25, 40, or 50 nm and a step size of 1 nm. Preferably, the width is 10 nm. For a FWHM width of 10 nm, each pixel comprises 48 channels. For FWHM widths of 25 nm and 50 nm, each pixel comprises 18 channels.

[0068] Identification and segmentation of bacterial colonies are then performed (step E5). For example, a bacterial colony identification algorithm is implemented on I_HSI or I_MSI images. Any conventional pattern and object recognition algorithm can be implemented to extract an area of ​​the I_HSI or I_MSI image, denoted Col_HSI or I_MSI, corresponding to a colony. Alternatively, this selection can be made manually by an operator who selects this area in the image using the display screen and a user interface. In general, the segmentation makes it possible to detect all colonies of interest, removing artifacts such as filaments or dust. The segmentation can be implemented in any known way. Alternatively, an operator selects, for example using a graphical interface, one or more colonies to be characterized.

[0069] It should be noted that each pixel of the spectral image comprises a spectrum which is a curve representing the light intensity measured at the colony scale as a function of the frequency. This is a vector of size the number of channels of the spectral image. As an example with the acquisition device described above, a colony that can be in the shape of a circle with a diameter of a few tenths of a millimeter to a few millimeters can extend over an area of ​​the spectral image of a dozen to a few hundred pixels, xels. In the case where the spectral image represents several colonies of the same strain, the procedure described here applies to each colony or to a set of colonies chosen according to criteria of size or position in the culture medium for example.

[0070] Advantageously, smoothing of the segmented image is implemented using conventional techniques (step E6). Techniques known by their English names, such as Raw, Smoothing, SNV, First derivative or Second derivative, can be applied.

[0071] Once a colony is identified, a processing (step E7) of the spectral image of the colony is implemented.

[0072] According to a first embodiment, such processing consists of applying convolutional filtering to the spectral image of the colony. The filtered image is the same size as the colony image. Such filtering reduces the noise in the spectral image. For example, each pixel in the filtered image is the average of a group of neighboring pixels in the spectral image of the colony. Thus, such filtering groups the pixels of the colony image in a sliding manner. According to this first embodiment, between 1 and 9 neighboring pixels, preferably 4 neighboring pixels, are considered. Figure 5 illustrates such a first embodiment. On the left is the original image of the colony, and on the right is the filtered image (test image). A pixel in the test image corresponds to the average of 4 neighboring pixels in the colony image. The 4 neighboring pixels are grouped in a sliding manner.

[0073] According to a second embodiment, such processing also consists of grouping neighboring pixels of the colony image into pixel packets. Each pixel of the filtered image then corresponds to the average of the pixels in each packet. In this case, the filtered image is smaller than the colony image. According to this second embodiment, neighboring pixels are grouped into packets of 2, 3, 4, or 5, preferably 5, and then the spectra of the selected pixels are averaged. The filtered image then comprises pixels that are the average of several pixels from the spectral or hyperspectral image. By grouping the pixels, the noise is also reduced. Figure 6 illustrates such a second embodiment. On the left is the colony image (i) and on the right is the filtered image (t) (test image). One pixel of the test image corresponds to the average of 4 neighboring pixels in the colony image.Here there is no sliding grouping; the pixel packets are neighbors and independent.

[0074] With the processing according to the first or second embodiment, a "test image" is obtained, each pixel being a function of the pixels of the colony image and preferably of the average of several neighboring pixels.

[0075] It should be noted that, according to the prior art of microbial characterization by hyperspectral or multispectral imaging, for example described in applications WO2019 / 122732 and WO2023 / 094775, the spectra of the pixels of a colony are averaged, thus producing a single spectrum per colony, a single spectrum which is then processed to obtain the desired information. All pixels are therefore averaged.

[0076] In another variant, each pixel of the colony, and therefore each spectrum associated with a colony, is processed to obtain this information and a final consensus is obtained by a majority vote. To improve the performance of these prior art processes based on colony averaging or pixel-by-pixel processing followed by majority voting, the aim was to improve training sets (particularly in terms of diversity, quantity of data, cross-validation strategy, etc.) and the application of other algorithms, in particular those based on deep learning (especially convolutional neural networks).

[0077] In the invention described herein, the inventors do not seek to improve training databases or prediction models but rather to characterize the signal-to-noise ratio of spectra. Indeed, the inventors have observed that averaging spectra over a colony, and therefore averaging several dozen, or even a hundred, spectra, results in the loss of important information for characterizing microorganisms. On the other hand, the spectrum of a pixel is so noisy that this "fine" information is difficult to exploit. The inventors have thus found that averaging 2 to 10 spectra reduces noise while preserving this information, with the best performance observed being associated with averaging 3 to 7 spectra, the peak performance occurring at 5 spectra.

[0078] This shows the advantage of grouping the pixels of the colony image into several packets so that the pixels of the test image are a function of an average of these packets.

[0079] The process then continues with a classification of each pixel of the test image, predicting whether the spectrum of the pixel in the test image belongs to a class of a prediction model with at least two classes for characterizing the microbial strain (step E8). Following classification, each pixel is either associated with a class or is unassigned.

[0080] Next, the class characterizing the microbial strain is determined by a majority vote on all the predicted classes for each pixel (step E9). This involves taking all the classes and calculating which class of distinguishes. The strain class is the one with the most votes. Class determination is effective for a vote percentage between 60% and 80%, specifically a vote of 60% or 80%.

[0081] Based on the strain class, the susceptibility of the microbial strain to a microbial agent is determined as being that associated with the class characterizing the microbial strain (step E10). Indeed, each class is associated with susceptibility or not to a microbial agent.

[0082] Multi-class model

[0083] According to one embodiment, the prediction model makes it possible to predict which strain, among several strains, the spectrum of the pixel corresponds to, each strain being susceptible or not to a microbial agent. This is then referred to as a multi-class model (more than 2). In the case of a Staphylococcus aureus microbial strain, the prediction model includes 73 strains or classes that are susceptible or not to methicillin. After classification, each pixel is associated with a strain or is left unassigned. After voting, a Staphylococcus aureus strain is obtained, and its susceptibility to methicillin is associated with this strain in the database. For example, MRSA stands for "Methicillin-resistant Staphylococcus aureus" and MSSA for "Methicillin-sensitive Staphylococcus aureus," that is, Staphylococcus aureus strains that are respectively resistant or not to the methicillin antibiotic.

[0084] In this respect, the model uses a database comprising 73 strains of Staphylococcus aureus, to which are associated spectral images of colonies obtained under the same conditions as those of the characterization process. Each strain is associated with an MSA or MRSA susceptibility. The model provides the strain here.

[0085] According to one embodiment, the prediction model makes it possible to predict the microbial species to which the colony belongs.

[0086] 2-class model

[0087] In one embodiment, the prediction model makes it possible to predict whether a pixel belongs to a strain susceptible or not to a microbial agent. This is referred to as a two-class model. In the case of a Staphylococcus aureus microbial strain, the prediction model comprises two classes, namely MRSA or MSA. Following classification, each pixel is associated with a susceptibility or is unassigned. Majority voting determines whether the strain is MRSA or MSA.

[0088] In this respect, the model uses a database comprising 73 strains of Staphylococcus aureus, to which are associated spectral images of colonies obtained under the same conditions as those of the characterization process. Each strain is associated with an MSA or MRSA susceptibility. The model directly provides the methicillin susceptibility.

[0089] The 2-class or multi-class model is a support vector classification (SVM) model or a convolutional neural network (CNN). In the case of an SVM, for example, an RBF kernel SVM (Radial Basis Function) is chosen.

[0090] Learning

[0091] The method advantageously includes a learning step (step E0) by means of a training database stored in the processing unit 4.

[0092] The training database for the microbial species considered is constituted as follows (step E01).

[0093] For each strain of a species, the strain is prepared in a Petri dish as in step El described previously.

[0094] In particular, for each strain, several samples (e.g., three, preferably four) are prepared, and spectral images are acquired for each colony of each sample using device 1, 1' as described in Figures 2 and 3, preferably from several colonies per Petri dish. The spectrum is processed as in step E4. The spectra of the first two samples are used for training the microbial strain classes, while the spectra of the other samples (e.g., the third sample in the case of triplet production) are used to test the performance of the training. Optionally, the acquisition is also performed on several samples using different capture devices in order to capture the variability of the spectra caused by differences in the characteristics of the devices (e.g., the variability of the light sources between devices, etc.).

[0095] The phenotypic measurement of the strain's susceptibility to the antimicrobial agent, and its memorization, is carried out.

[0096] Finally, a genomic characterization of the strain is carried out using a complete sequencing of its genome and the establishment of a wgMLST profile as described in the document "MLST revisited: the gene-by-gene approach to bacterial genomics" by Martin CJ Maiden, Nature Reviews Microbiology, 2013, and the storage of this characterization.

[0097] The learning database then includes, for each strain of a species, its susceptibility to the microbial agent as well as corresponding colony spectra.

[0098] In the case of Staphylococcus aureus, 73 different strains are used for training, some of which are associated with susceptibility to methicillin. This is shown in [Fig. 5].

[0099] Once the database has been established, the training of the prediction model itself (step E02) is implemented in a conventional manner. Examples

[0100] A prediction model typically comprises two parts: a first part that outputs a score for each class (for example, between 0 and 1 for a multi-class prediction), possibly associated with a confidence index, and a second part that applies a predetermined rule to these scores to produce a final classification result, for example, selecting the class associated with the highest score or selecting the classes whose scores exceed a predetermined threshold. In this second case, for the processing of each pixel followed by a majority vote, a pixel is unassigned when none of the scores exceeds the single threshold, or, when a confidence index is generated, when the latter does not exceed a predetermined threshold. Similarly, when the majority of pixels do not exceed the voting threshold, the colony is assigned the unassigned status.

[0101] For treatments based on the average spectrum of a colony, the latter is assigned the status of unassigned when none of the scores exceeds a threshold or, where applicable, the confidence index is too low.

[0102] Regardless of the algorithm used, how it predicts classes and / or calculates a confidence index, it should be noted that for sensitive applications, such as the in vitro diagnosis of microorganisms infecting a patient, it is important to have the highest possible specificity for positive results, even at the expense of sensitivity (and therefore the "unassigned" prediction). Indeed, false positives or true negatives can have adverse consequences for the patient's health, such as choosing an ineffective antibiotic based on a prediction of a bacterium's susceptibility to that antibiotic. When a prediction yields an "unassigned" result, microbiologists can then initiate further tests to obtain the desired information.

[0103] Thus, improving a predictive model applied to in vitro diagnosis consists in particular of being demanding on the certainty of the prediction while limiting as much as possible the rate of non-assignment which requires carrying out other tests.

[0104] The invention has been applied to the characterization of methicillin-resistant Staphylococcus aureus microbial strains in order to identify MRSA and MSA strains. Figure 7 shows MRSA and MSA strains used.

[0105] Several levels were tested:

[0106] Colony: classification according to the whole colony by averaging the spectra of the pixels and direct classification by a classification model.

[0107] Pixel: classification of each pixel by the classification model.

[0108] Packets: grouping into packets of 5 pixels and using spectra for each packet and classification of each packet by the classification model.

[0109] In what follows, the classification rate corresponds to a correct classification, while not assigned means that the classification did not return a result, as explained above.

[0110] Scenario 1: 73-Class Model - Hyperspectral Image Level Classification Rate Unassigned Colony Colony 79.0% # Pixel Pixel 42.5% # Colony 60% vote 77.8% 60.4% Colony 80% vote 86.1% 81.7% Bundles Grouping 65.8% # Grouping (60% vote) 78.8% 21.1% Grouping (80% vote) 84.1% 43.7%

[0111] The best performance was achieved with the pixel level and a majority vote of 80%. However, 81.7% were unassigned.

[0112] The packet level made it possible to reduce the percentage of unassigned to 43.7% with a very satisfactory classification rate.

[0113] Scenario 2: MRSA / MSA 2-Class Model - Hyperspectral Image Level Classification Rate Unassigned Colony 94.3% # Pixel Pixel 81.2% # Colony 60% vote 94.0% 8.3% Colony 80% vote 97.2% 31.6% Packets Grouping 90.5% # Grouping (60% vote) 95.2% 3.7% Grouping (80% vote) 97.9% 13.1%

[0114] The best performance is clearly achieved for the packet level with a low percentage of unassigned.

[0115] Scenario 3 - 73-class model - comparison of hyperspectral HSI / multispectral MSI image (23 channels, 10 nm wide) Level Image type Classification rate Not assigned HSI Colony 79.9% # MSI 76.6% # HSI Packages 68.2% # MSI 65.6% # HSI Colony votes at 60% 85.5% 21.1% MSI Colony votes at 60% 85.2% 23.0% HSI Colony votes at 80% 92.9% 43.7% MSI Colony votes at 80% 92.1% 47.5%

[0116] Scenario 4 - 2-class model - comparison of HS1 hyperspectral / MSI multispectral image (23 channels, 10 nm wide) Level Image Type Classification Rate Unassigned Colony HSI 94.3% # MSI 93.1% # Packets HSI 90.5% # MSI 89.4% # HSI Colony votes 60% 95.2% 3.7% MSI Colony votes 60% 94.4% 3.0% HSI Colony votes 80% 97.9% 13.1% MSI Colonie votes at 80% 97.1% 15.3%

[0117] Scenarios 3 and 4 show that good performance is obtained using the packet level, particularly at the 2-class model level, by reducing the percentage of unassigned values ​​compared to the 73-class model. Using an MSI image allows for good performance while reducing the complexity of image acquisition.

[0118] In summary, the best performance for phenotype identification is obtained at the package level and 80% majority vote with a classification rate of 97.1% and 15.3% of colonies not assigned.

[0119] Compared with the entire spectrum at the same level of analysis, all performances are significantly lower before the majority vote. The vote improves performance.

[0120] The use of an MSI image with an FWHM width of 10 nm leads to good performance.

Claims

Demands

1. A computer-implemented method for characterizing a microbial strain, comprising the following steps: - obtaining (E3) a spectral image of a colony, representing a colony of a microbial strain, the spectral image of the colony comprising several pixels, each pixel comprising several spectral channels; - filtering (E7) the spectral image of the colony to obtain a "test image", each pixel of the test image being a function of an average of a packet of neighboring pixels of the colony image, each packet comprising M <N pixels avec N le nombre de pixels de l’image de la colonie ; - classification (E8) de chaque pixel de l’image de test prédisant si le spectre du pixel appartient à une classe d’un modèle de prédiction à moins deux classes de caractérisation de la souche microbienne ;- determination (E9) of the class characterizing the microbial strain by a majority vote on all the classes predicted for each pixel; - determination (E10) of a susceptibility of the microbial strain to a microbial agent as being that associated with the class characterizing the microbial strain.

2. A method according to claim 1, wherein the filtering consists of grouping the N pixels of the colony image into packets of M <N pixels voisins, un pixel de l’image de test étant la moyenne des M pixels voisins de l’image de la colonie, l’image de test présentant un nombre de pixels inférieurs à celui de l’image de la colonie.

3. A method according to claim 1, wherein the filtering process consists of applying a convolutional filter to the colony image, a pixel of the test image being obtained by processing several neighboring pixels of the colony image using a convolutional kernel, the test image comprising N pixels, each pixel corresponding to an average of M <N pixels voisins de l’image de la colonie, l’image de test présentant un nombre de pixels identique à celui de l’image de la colonie.

4. A method according to any one of claims 1 to 3 wherein M is between 2 and 10, preferably equal to 5.

5. A method according to any one of the preceding claims, wherein the threshold for the majority vote is between 60% and 80%, in particular 60% or 80%.

6. A method according to any one of claims 1 to 5, wherein the prediction model is a model with at least three classes, each class corresponding to a microbial strain of a type, each strain being previously characterized as susceptible or not to a microbial agent, the susceptibility or resistance of the microbial strain to a microbial agent being that associated with the microbial class.

7. A method according to any one of claims 1 to 5, wherein the prediction model is a 2-class model, each class indicating susceptibility or non-susceptibility to a microbial agent, the susceptibility of the microbial strain to a microbial agent being the result of the step of determining the class characterizing the microbial strain.

8. A method according to any one of claims 6 to 7, comprising a learning (E0) of the prediction model, the training comprising the following steps: - obtaining spectral images of different colonies for several microbial strains; - training the classification on the spectral images of different colonies.

9. A method according to any one of the preceding claims, wherein the spectral image comprises for each pixel 240 channels spaced 2.1nm apart over a band between 390 nm and 900 nm, the spectral image being hyperspectral.

10. A method according to any one of claims 1 to 8, wherein the spectral image comprises for each pixel 23 channels spaced 10 nm apart over a spectral band between 390 nm and 900 nm, the image being multispectral.

11. A method according to any one of the preceding claims, wherein the microbial strain is a strain of Staphylococcus aureus and the antimicrobial agent is methicillin.

12. Product computer program comprising code instructions for carrying out the steps of the process according to any one of the preceding claims when said program is executed on a computer.

13. A computer-implemented method for training a prediction model implemented in the characterization method according to any one of claims 1 to 11, comprising the following steps: - creation (E01) of a training database including, for each strain of a species, its susceptibility to the microbial agent and corresponding colony spectra; - supervised training (E02) of the prediction model so as to obtain parameters of the prediction model on different colony spectra.

Images