Method and system for identifying micro-organisms taking colony form

Abstract

The invention relates to a method for identifying a microbial strain taking the form of a colony, comprising acquiring a set of at least three spectral channels of the colony contained in the visible and infrared spectrum; and making, by means of a computer, a prediction of the species depending on the acquired spectral channels. The three channels of said set that contribute the most to the performance of the prediction are selected from a first spectral band [730-780] nm and from a second spectral band [550-650] nm.

Description

[0001] METHOD AND SYSTEM FOR IDENTIFYING MICROORGANISMS IN COLONY FORM

[0002] FIELD OF THE INVENTION

[0003] The invention relates to the field of microbiological analysis, and in particular to the identification of bacteria, yeasts and molds having grown in the form of a colony on a nutrient medium, for example poured into a Petri dish, and preferably a non-chromogenic, non-fluorogenic and dye-free nutrient medium.

[0004] Advantageously, the invention can be implemented using a portable device such as a smartphone or tablet.

[0005] STATE OF THE ART

[0006] It is known that the visible and near infrared spectrum contains the information necessary for identifying the species of a bacterial colony grown on a culture medium, for example contained in a Petri dish, the medium and the dish being hereinafter referred to together as "Petri dish".

[0007] For example, the Applicant achieves an accuracy of identification of the species forming a colony by analysis by means of automated learning of the image of this colony of around ten pixels taken by a hyperspectral camera with 240 spectral channels in the range [390-900] nm greater than 95%, very often greater than 98%. Reference may be made to the documents Feng, Y. Z & al., "Invasive weed optimization for optimizing one-agar-for-all classification of bacterial colonies based on hyperspectral imaging." Sensors and Actuators B-Chemical (2018) 269: 264-270 and the document "Hyperspectral image analysis for rapid and accurate discrimination of bacterial infections: A benchmark study" by Arrigoni et al., Computers in Biology and Medicine 88 (2017)

[0008] While hyperspectral technology is now mature and inexpensive, with a camera of this type costing around tens of thousands of euros, it remains inaccessible in many situations, particularly in developing countries. In addition to the cost remaining too high, hyperspectral technology is also not very mobile. The bulky cameras remain permanently located in laboratories, while microbiologists travel to reach patients in remote areas of the country.

[0009] In this context, there is a need for identification that can be carried out using even less expensive, mobile, space-saving devices that perform the entire identification process on their own, from image capture to analysis rendering, with performances equal to or very close to those obtained by using hyperspectral imaging with hundreds of spectral channels.

[0010] STATEMENT OF THE INVENTION

[0011] The aim of the present invention is to propose an efficient method for identifying microorganisms in the form of colonies by analyzing a limited number of channels in the visible and near infrared range.

[0012] To this end, the invention relates to a method for identifying a microbial strain in the form of a colony, comprising:

[0013] A. the acquisition of a set of at least three spectral channels of the colony included in the visible and infrared spectrum; and

[0014] B. computer implementation of a species prediction based on the acquired spectral channels.

[0015] According to the invention, the three channels of said set which contribute most to the prediction performance are chosen from:

[0016] - a first spectral band [730-780]nm;

[0017] - a second spectral band [550-650]nm.

[0018] The invention also relates to a method for identifying a microorganism likely to be contained in a biological sample of a predefined type, comprising:

[0019] A. the acquisition of at least one multispectral digital image of at most ten channels of at least one microbial colony having grown on a culture medium following the spreading of the sample on said medium;

[0020] B. the prediction, implemented by computer, of the microbial species forming said colony among a predetermined panel of microbial species based on the digital multispectral image(s) acquired;

[0021] According to the invention, the panel of microbial species, the optical filters and the identification algorithm are chosen so that the specificity of the identification is greater than or equal to 90%, and preferably greater than or equal to 94%;

[0022] - if the prediction model delivers the identity of a species o then said identity is confirmed without implementing another method of identifying the microorganisms; o otherwise another method of identifying the microorganism is implemented. The invention also relates to a Device for identifying a microbial colony from a predetermined panel of microbial species responsible for pathogenic infections of a predetermined type, in particular urinary, comprising:

[0023] A. a digital camera configured to acquire a digital image of said colony;

[0024] B. and a computer processing unit connected to said camera to receive the digital image of the colony;

[0025] According to the invention, the digital camera comprises:

[0026] A.1. a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm;

[0027] A.2.a lens capable of forming an image on said sensor in said range;

[0028] A.3.a set of at most ten optical bandpass filters centered respectively on different wavelengths of the wavelength range [390-900] nm, each having a bandwidth in the range [8-50] nm, each of said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; and the three channels of said set that contribute most to the prediction performance are chosen from:

[0029] - a first spectral band [730-780]nm;

[0030] - a second spectral band [550-660]nm

[0031] The invention also relates to a device for identifying a microbial colony from a predetermined panel of microbial species responsible for pathogenic infections of a predetermined type, in particular urinary infections, comprising:

[0032] A. a digital camera configured to acquire a digital image of said colony;

[0033] B. and a computer processing unit connected to said camera to receive the digital image of the colony;

[0034] According to the invention, the digital camera comprises:

[0035] A.1. a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm;

[0036] A.2. a lens capable of forming an image on said sensor in said range;

[0037] A.3. a set of at most ten optical bandpass filters respectively centered on different wavelengths of the wavelength range [390-900] nm, each having a bandwidth in the range [8-50] nm, each of said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; and the computer processing unit is configured to implement an algorithm for identifying the species forming the colony based on the digital images, and the panel of microbial species corresponds to a prevalence of at least 70% of pathological infections, and preferably greater than or equal to 80%, and the panel of microbial species, the optical filters and the identification algorithm are chosen so that the specificity of the identification is greater than or equal to 90%, and preferably greater than or equal to 94%.

[0038] The invention also relates to a device for characterizing a microbial colony, comprising:

[0039] A. a digital camera configured to acquire a digital image of said colony;

[0040] B. and a computer processing unit connected to said camera for receiving the digital image of the colony; according to the invention, the digital camera comprises:

[0041] Al a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm;

[0042] A.2.a lens capable of forming an image on said sensor in said range;

[0043] A.3.a set of up to ten optical bandpass filters, each having a bandwidth in the range [8-50]nm, each of said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; and the set of up to 10 filters comprises:

[0044] A.3.1.a first filter centered on a wavelength in the range [415-440]nm;

[0045] A.3.2.second filters centered respectively on different wavelengths of the wavelength range [390-900] nm and the computer processing unit is configured to identify:

[0046] B.1. the gram of the species forming the colony based on the digital image of said colony acquired using the first filter;

[0047] B.2. the identity of the species forming the colony based on the digital images acquired using the second filters.

[0048] BRIEF DESCRIPTION OF THE FIGURES

[0049] The invention will be better understood on reading the following description, given solely by way of example, and read in conjunction with the appended drawings, in which identical references designate identical or similar elements, and in which

[0050] - Figure 1 is a schematic of a system for multispectral acquisition and identification of colony-forming microbial species according to the invention;

[0051] - figure 2 is an exploded view of this same system; - figure 3 is a simplified top view of a filter wheel forming part of this system;

[0052] - figure 4 is a schematic view of the frequency response of a bandpass filter forming part of the filter wheel;

[0053] - Figure 5 illustrates the reflectance at different spectral channels for different species, reflectance here illustrated by its box-plot at 95 e percentile;

[0054] - Figure 6 is the curve of the classification error rate as a function of the channels selected using a "feed-forward" approach;

[0055] - Figure 7 illustrates a decision tree for the classification of 8 uropathogens based on reflectance at different spectral channels

[0056] - Figure 8 illustrates another decision tree for the classification of 8 uropathogens based on reflectance at different spectral channels

[0057] - Figure 9 is a flowchart for processing urine samples based on the invention.

[0058] DETAILED DESCRIPTION OF THE INVENTION

[0059] A. Multispectral identification system according to the invention

[0060] Referring to Figures 1 to 4, a system 10 for identifying microorganisms having grown in the form of a colony comprises:

[0061] - an image acquisition device 12 comprising a matrix image sensor 14, a lens 15, and a processing circuit receiving the signals from the sensor 14 and configuring them for transmission, for example via a USB-C or Bluetooth connection, to a processing unit. The sensor 14 is, for example, a 5 million pixel black and white CMOS sensor marketed by the company E-Con Systems, India under the reference See3CAM_CU55M, controllable from a smartphone application connected by USB-C;

[0062] - a filter wheel 16, comprising a set of bandpass filters 18, of a number less than 10. The wheel is configured relative to the device 12 to be rotated by the user so as to sequentially present each of the filters 18 in front of the objective 15. Referring to FIG. 4, the width of a filter, defined as the width at half maximum of the frequency response, is between 8nm and 50nm. For example, the filters are dichroic bandpass filters with a width of 10nm;

[0063] - optionally a lighting device 20, advantageously in the form of a dome under which a Petri dish 22 comprising the microbial colonies to be identified; an image processing unit 26, preferably in the form of a smartphone or a tablet, connectable to the image acquisition device 12, for example by means of the USB-C cable

[0064] The image sensor 14 of the device 12, of CMOS or CCD technology, is a "bare" sensor in that it is not coated with an infrared filter or a Bayer matrix as is the case with sensors equipping smartphones and tablets intended for the general public. In the absence of the filters 18, this monochrome sensor is therefore configured to integrate the light flux incident on its photosites over the entire range [390-900]nm. The optical assembly, comprising the sensor, the objective(s) with their aperture(s), as well as the distance from the image plane of the sensor to the Petri dish 22, is configured so as to obtain a sharp image of at least 5 pixels, and preferably at least 10 pixels, of a microbial colony having a diameter of between 3 and 5mm, regardless of the filter 18 present in front of the objectives.

[0065] The lighting device 20 comprises in its internal lower part a ring of LEDs 24 selected to emit lighting in the visible and near infrared, in particular in the entire range [390-900]nm, for example white LEDs marketed by the company Yujiled, China. The device 20 is designed to be a light shaper and comprises for this purpose on its internal wall 26 a white or silver reflective surface, for example a 17% gray, so as to obtain uniform lighting avoiding specularity phenomena on the colonies having grown on the Petri dish 22. This device also comprises, in the field of vision of the optical assembly, a white surface in order to carry out a correction of the color temperature, or correction of the white point.

[0066] The acquisition device 12 is provided in addition to a processing unit 12 in the form of a smartphone or a consumer tablet. These are often the only recent IT resource available to healthcare personnel and have multiple uses, both professional and personal. They are therefore more subject to breakages, breakdowns, or frequent changes. By thus clearly separating the image acquisition device from the smartphone responsible for the computer processing of the images, a double advantage is thus obtained: a) the most "fragile" part is replaceable without the acquisition device having to be changed; b) since smartphones are mass-produced, their cost is therefore lower than a device that would be produced specifically to implement the invention.However, the invention covers a smartphone or tablet whose sensor and optical system have been specifically designed and produced to implement the invention, the device 12 then being omitted and the lens of the smartphone directly positioned in front of the filter wheel 16. The lighting device 20, although optional, is provided to greatly facilitate the use of the invention. Indeed, it makes it possible to obtain adequate lighting inside the dome, and therefore on the surface of the Petri dish, regardless of the ambient lighting. This device thus avoids the user having to worry about this aspect, and therefore to obtain a microbial diagnosis regardless of the lighting conditions or the user's skills in terms of photographic settings.Furthermore, it allows the device 12 to be positioned always at the same distance from the Petri dish, thus avoiding the user having to worry about adjusting the focus on the Petri dish. The combination of the image acquisition device 12 and the illuminating dome 20, with a combined height of about ten centimeters and a maximum diameter of the same order, therefore allows anyone to take images of the Petri dish 22, without having to undergo lengthy training, including the doctors' assistants, thus freeing up time for the latter. Advantageously, the filter wheel 16 is an integral part of the upper part of the illuminating dome 20 so as to protect this element as much as possible, which is the most fragile and the most expensive.If the dome is presented here as rigid, it can be designed to be foldable to increase its transportability, for example like the transportable light boxes known in the field of advertising or culinary photography.

[0067] The processing of the acquired images is advantageously implemented in the form of an application downloaded onto the smartphone 26. While the configuration of this processing may require significant computing resources, particularly in the case of an embodiment based on automated learning (or "machine learning"), the use of the processing is, however, not very demanding in computing resources, so that the computing power of entry-level smartphones at the date of filing of the present application is largely sufficient. It will thus be noted that the system according to the invention, in addition to its ease of use, has a very moderate cost and can therefore be adopted on a large scale, particularly in developing countries. The processing of the images will be detailed below.

[0068] B. Identification by a multispectral device using a limited number of spectral channels

[0069] There will now be a method for identifying the colony-forming microbial species grown on a COS (Columbia Blood Agar) culture medium among the species Enterococcus faecalis (1), Escherichia coli (2), Klebsiella pneumoniae (3), Proteus mirabilis (4), Proteus vulgaris (5), Pseudomonas aeruginosa (6), Staphylococcus aureus (7), and Streptococcus agalactiae (8). The species listed here are chosen for diagnosis of urinary tract infection (UTI) and constitute more than 90% of these infections in the majority of hospitals. The number in parentheses represents their label in the figures. For the design of this embodiment, the inventors jointly pursued three objectives: i. minimize the number of bandpass filters which constitute to date the most expensive part of the system; ii.obtain the lowest possible classification error, at most equal to 6%, this threshold representing the value beyond which microbiologists and clinicians consider the diagnosis as scientifically too unreliable to be used routinely; iii. and design an identification algorithm that is understandable by users, unlike, for example, deep learning-based algorithms which are "black boxes" that can form a barrier to adoption or which can be subject to overfitting. The inventors have, however, successfully tested other prediction algorithms, such as automated learning based on SVM ("Support Vector Machine") with linear and radial kernels, or PLS-DA.

[0070] To do this, the inventors relied on hyperspectral imaging. It is known that the 390-900nm wavelength range contains the information necessary to identify colony-forming species with an accuracy close to 100%. The inventors therefore worked to discover specific spectral channels in this range that are most informative for identification, while ensuring that statistical correlation phenomena between channels that can occur in this type of study are eliminated.

[0071] To do this, a database of hyperspectral spectra of colonies of the species is first acquired. In particular, several strains were collected for each species, then each strain was spread on three Petri dishes, cultured at 37°C for 24 hours, then hyperspectral images of the dishes are acquired. A total of 1745 colonies grown on COS medium were analyzed, each consisting of about ten to a hundred pixels. The spectra from two dishes are used to learn an algorithm described below and the spectra from the third dish are used to test the learned algorithm. For example, the acquisition of images of Petri dishes in the range [390-900]nm, and therefore of the colonies, is carried out using a 240-channel CMOS hyperspectral camera with a spectral resolution of 2.1 nm. This acquisition system is, for example, that described in patent application WO 2019 / 122732.The preprocessing of the spectra acquired for each pixel of the bacterial colonies is then implemented in order to obtain reflectance spectra corrected for the white and black points, for example the preprocessing described in this document. The preprocessed reflectance spectra of a colony are then averaged over the pixels of this colony to obtain a single average spectrum per colony. Then the average reflectance spectrum obtained, with an initial resolution of 2.1nm, is subsampled in the wavelength domain in order to obtain a final spectral resolution compatible with commercial bandpass filters whose width is generally between 8nm and 50nm. For example, a sliding window averaging the intensities of the spectrum over the width of the window is applied. As an illustration, to obtain a final resolution of approximately 10nm, a window of width equal to 4 channels is applied, leading to a resolution of 4*2.1nm.The final reflectance spectrum is then stored in the database.

[0072] A total of 1745 colonies grown on COS medium were analyzed. Each bacterial strain acquired for the study was spread under three Petri dishes, two dishes being used for training the identification algorithms, the last one being used to test the algorithms' performance. Several predictive algorithms were tested, including Support Vector Machine predictions, with linear or non-linear kernels, for example radial, the learning being carried out using the "10-folds" technique.

[0073] An exploratory work of the descriptors of these algorithms was undertaken to determine a minimum number of channels allowing a classification error rate of less than 10%, and preferably less than or equal to 6%. In particular, a "feed forward" approach was implemented. In a first step, the number of descriptors is set equal to 1, and all the channels are scanned to determine the most predictive one. This channel is retained and the number of descriptors is set equal to 2, and the remaining channels are scanned to determine the second most predictive channel, and so on. As illustrated in Figure 5, the error rate drops rapidly as a function of the number of channels selected, this phenomenon being explained by the fact that the classes (labels of the species listed above) are separable by simple visual inspection of the reflectance of said channels (Figure 4).

[0074] The inventors have thus discovered that only 4 channels, with a width between 8nm and 50nm and distributed in the bands [580-660]nm and [730-780]nm, make it possible to obtain a classification error rate of less than 6%. In particular, the channels comprising the wavelength close to 750 nm (here 747nm), 620nm (here 619nm) and 600nm (here 595nm) make it possible to obtain the targeted performances.

[0075] More specifically, the inventors have discovered that the three channels with the highest predictive values ​​for the identification (in particular in terms of misclassification rate or in terms of specificity) of colony-forming species are included in the bands [580-660]nm and [730-780]nm. The predictive value of a channel is for example determined by the feed-forward approach as described above. It is for example also determined by the use of the PPS score (for "Power Predictive Score", described for example in the document "RIP correlation. Introducing the Predictive Power Score", Florian Wetschoreckde, Toward Data Science, 2020).In particular, by excluding these bands, the inventors have discovered that it is not possible to obtain a prediction error less than or equal to 6%, at least using classification algorithms successfully employed with all channels, for example SVMs with different kernels (linear and radial), PLS-DA or logistic regression algorithms (e.g. LASSO type). Advantageously, a fourth channel contributing most to the prediction performance is chosen in the spectral band [580-660]nm or [490-540]nm.

[0076] In a preferred embodiment, the prediction of species, for example implemented, without being limited thereto, by the device described in relation to FIG. 1, comprises the acquisition of images in different spectral channels, of a width for example between 8 and 50 nm, these channels being included only in the bands [580-660] nm and [730-780] nm. In a variant, the channels are included only in the bands [580-660] nm, [730-780] nm and [490-540] nm.

[0077] In a next step, the inventors focused on designing a simple interpretation prediction algorithm for easy adoption of this technology. Advantageously, as illustrated in Figure 7, a decision tree based solely on the four wavelengths mentioned above makes it possible to predict the identity of the colony-forming bacterial species. In particular, a channel comprising the 750nm wavelength makes it possible to directly identify classes 1, 4 and 5 whose botploxes are distinct from each other, which makes it possible to define intensity ranges separating them. The remaining classes (2, 3, 6, 7 and 8) are then separated according to a channel comprising the wavelength 620nm, making it possible to identify class 8, and a first group comprising classes 2 and 3, and a second group comprising classes 6 and 7. Finally, a channel comprising the wavelength 595nm makes it possible to separate classes 2 and 3 on the one hand and classes 6 and 7 on the other hand.The "feed forward" approach, which uses machine learning tools to identify channels with high prediction weight, therefore also makes it possible to identify channels for which the light intensities of the colony-forming species can be separated by simple thresholding. It should be noted that in this embodiment based on a decision tree, no machine learning technique is used to identify the species, which in particular makes it possible to overcome learning biases and minimize the risk of overfitting. Obviously, the invention also covers automated learning models, in particular Support Vector Machine models, for example with a linear kernel. The study carried out by the inventors on the basis of a hyperspectral camera therefore confirms the possibility of implementing a multispectral approach (i.e. based on a limited number of channels) based on simple prediction algorithms for understanding (i.e.a decision tree) for the identification of bacterial species on Petri dishes with acceptable performance in terms of in vitro diagnosis. More specifically, it is demonstrated that the use of a multispectral technology whose signal acquisition performance in the selected channels having the same quality as the hyperspectral technology leads to an appropriate identification.

[0078] C. Industrialization of multispectral identification

[0079] As is sometimes the case, the transition from a laboratory prototype based on a technology specifically designed for a purpose (a hyperspectral technology designed for acquisition in the visible and near infrared and used in a controlled environment) to a commercial device based on another technology (here a monochromatic camera based on CMOS or CDD whose response beyond 700nm is noisy, associated with commercial optical filters and used in a less controlled environment) induces a degradation of the expected performances. Thus, the choice of filters corresponding to the channels determined on the basis of the hyperspectral technology leads to a significant drop in performances, the classification error rate being able to be higher than 10%, or even 15%.

[0080] According to the invention, it is therefore a question of restoring in vitro diagnostic performance, while respecting the constraints of use, robustness, and cost described above. To do this, the inventors have therefore parameterized the multispectral device according to the invention by seeking a synergy between technical characteristics and clinical microbiological IVD workflow characteristics: i. Multispectral technical characteristics a. restriction of the wavelength range, in particular by reducing the value of the largest channel to move it away from the near-infrared wavelength range where the sensor has a degraded response b. addition of one or more additional channels, and therefore of optical bandpass filters, beyond the 4 channels determined via hyperspectral technology. At most ten filters are used.Beyond this, the system as a whole, which preferentially encases the filters in the dome 20 for their protection, becomes too bulky and / or too expensive; c. optionally, the choice of bandpass filters to set a different channel width, in particular a more restricted one if this allows for increased performance. ii. Clinical IVD workflow characteristics d. choice of a more restricted panel of species to be identified if this helps to improve the specificity of the prediction (i.e. the ability to correctly predict the correct species among the samples actually comprising at least one of the species in the panel), the restriction being however carried out to minimally degrade the prevalence of this panel in infections, here urinary tract infections. e. adjustment of the prediction algorithm so as to prioritize specificity over the rejection rate (i.e. percentage of samples for which no identification result is returned).

[0081] To achieve the synergy of features a and b, the "feed forward" approach is relaunched based on hyperspectral technology, with features d and e being hyperparameters of the prediction algorithm. Regarding feature e, as is known per se, classification-based algorithms comprise two stages: a first stage returning one or more scores quantifying one or more distances of the sample under test to the different classes of the algorithm and a second stage that determines a class identity based on the calculated distances by comparing them to a predefined threshold. Threshold selection allows for a compromise between sensitivity or specificity, for example. If a decision tree is used, the value of the intensity ranges to separate the different classes and class groups also leads to a compromise between specificity and rejection rate. Here again, specificity is favored.

[0082] Concerning characteristic d, it is for example a question of parameterizing characteristics a, b and e according to different panels chosen from a pre-established list of species, in particular the eight species described above.

[0083] By proceeding in this manner, the inventors designed a first embodiment of the device of Figure 1: i. based on 6 optical filters with a width of 10nm, respectively comprising the wavelengths 650nm 730nm, 670nm, 630nm, 471nm and 580nm; ii. for the panel of species Enterococcus faecalis (1), Escherichia coli (2), Klebsiella pneumoniae (3), Proteus mirabilis (4) and Streptococcus agalactiae (5) which account for more than 80% of urinary tract infections iii. with a specificity equal to 96% at the expense of a sensitivity equal to 79%.

[0084] By continuing the synergistic exploration of the ad characteristics, the inventors designed a second embodiment of the device of Figure 1: i. based on 5 optical filters with a width of 10nm, respectively comprising the wavelengths 640nm, 660nm, 580nm, and 750nm, ii. for the panel of species Enterococcus faecalis (1), Escherichia coli (2), Klebsiella pneumoniae (3), Proteus mirabilis (4), Pseudomonas aeruginosa (5), Staphylococcus aureus (6), and Streptococcus agalactiae (7) which account for more than 80% of urinary tract infections iii. again with a specificity equal to greater than 94%, at the expense of a sensitivity close to 80%.

[0085] Figure 8 illustrates a decision tree designed by the inventors for the 7-species embodiment, again with a view to adoption, the classes being directly separable by thresholding between different channels by a particular choice of the sequence of channels to be analyzed.

[0086] Therefore, the doctor can have confidence in the system when the latter returns an identification result, the specificity being in accordance with the obligations that a certified "in vitro diagnostic" device must respect (i.e. scientifically rendering a result that the doctor can take as microbiological "truth", without therefore having to carry out additional tests or to interpret the latter on the basis of his know-how and / or detailed knowledge of the bacterial ecology with which the patient is confronted).

[0087] The rejection rate of positive samples (i.e. samples for which colony growth is observed and a prediction that returns no identity) is however approximately 20%, meaning that for one positive sample in 5, the physician has no clinical information. Conversely, 80% of positive samples are therefore characterized with certainty, which should be considered in light of certain cases where no sample is characterized at all.

[0088] In a preferred embodiment, the rejection rate of positive samples is reduced by at least one of the following additional features, alone or in combination: e. one vote per dish: the invention makes it possible to analyze several colonies, and not just one as is the case with conventional clinical workflows. Indeed, it is customary in the clinical laboratory to sample a single colony per dish, a choice made according to the experience of the laboratory technician, to carry out subsequent tests, such as for example an identification test using the applicant's Vitek MS spectrometer or an antibiogram test using the applicant's Vitek 2. Thanks to the invention, the entire Petri dish is imaged, and the operator can select, for example by means of a screen and a pointer controlled by the application downloaded on the smartphone or using an automatic segmentation algorithm, several colonies.Each of the colonies is then analyzed, thus reducing the risk of not making any identification for the dish. A majority vote on the identified species is then returned to the user, for example a majority vote strictly greater than 50%, in particular greater than or equal to 60%. f. the addition of a bandpass filter in the 415-440nm range and the determination of the gram of the colony-forming bacteria, for example, as described in patent application WO 2017 / 216190. g. enumeration of the colonies. By having at least one complete image of the Petri dish, it is possible to characterize the bacterial load of the sample, for example by measuring the surface area of ​​the colonies on the image (i.e., the light surfaces on a generally dark culture medium) and by calculating the ratio of the surface area of ​​the colonies to the surface area of ​​the dish.In the case of a urine sample, for example, sterile when the patient is not infected, this makes it possible to confirm the presence of an infection in the patient.

[0089] Regarding the characteristic f, the Gram prediction is set to maximize its specificity. In this way, if the device does not make any identification, it still returns a result on the Gram of at least one colony-forming species on which the physician can rely to determine empirical antibiotic therapy. Thus, the percentage of samples with a positive culture, having no clinical characterization, namely neither the identity nor the Gram of a pathogen included in a sample, is very low, less than or equal to 5%.

[0090] One embodiment of a method for antimicrobial diagnosis and therapy is illustrated in Figure 9 in relation to urine samples.

[0091] The process begins in 40 by taking a urine sample from a patient suspected of having a urinary tract infection, then spreading the sample in 42 on a culture medium included in a Petri dish (COS, TSS, etc.) and then culturing the dish, in 44, at 37°C for at least 12 hours. Once the incubation is complete, at 46 an operator places the box under the illuminating dome 20 of the multispectral device 10, and rotates the filter wheel while taking an image for each filter using the application downloaded on the smartphone 26. Once all the images have been acquired, one or more colonies are selected at 48, either by the operator using the smartphone screen which displays the image of the box by one of the filters or an aggregated image, or selected using a segmentation algorithm selecting circular 2D objects in a manner known per se (for example by using a Hough-type image filter).For each selected colony, an averaging of the colony pixel intensity is performed for each channel at 50, thus leading to a single intensity value per channel for said colony. This average multispectral spectrum is then optionally preprocessed at 52, for example to adjust its white and black points and / or transform its intensity into reflectance as described for example in document WO 2019 / 122732. An identification of the identity of the microbial species(s) forming the selected colony(ies) is then implemented at 54. If an identity is returned by the prediction, the physician can then, 56, select the antimicrobial therapy based on this identity and administer it to the patient.If no identification is returned, the Petri dish, or an aliquot of the initial sample, is then, for example, transported, at 58, to a laboratory equipped with instruments capable of a more complete characterization at 60, the results of which are returned, for example, to the doctor on his smartphone 26 who chooses and administers the therapy accordingly.

[0092] In a variant, when the identification result is negative, the Gram is calculated if the device 10 has the necessary filter, in a manner described above, or any other medical value useful to the doctor (enumeration, box voting, etc.).

[0093] In a variant of the invention, it is described that it is the application downloaded on the smartphone that carries out the storage of the images and the prediction of the species present in the sample. In a variant, when the smartphone is connectable to a data transmission network (ADSL, fiber, satellite, etc.), this storage and this prediction are carried out on a remote server, the smartphone being used to transmit the images to said server. In this way, the maintenance of the smartphone and the updating of the prediction software are simpler and more robust, the monitoring of patient files can be centralized in real time, or a qualified technician receiving the images centrally can put his expertise at the service of several traveling doctors, provide live support, etc.

[0094] A computer processing unit means in particular any hardware comprising one or more processors or microprocessors associated with computer memories (cache, RAM, ROM, etc.) capable of storing instructions for the implementation of calculations and data processing when they are executed.

Claims

CLAIMS 1. Method for identifying a microbial strain in the form of a colony, comprising: A. acquisition of a set of at least three spectral channels of the colony included in the visible and infrared spectrum; B. the computer implementation of a prediction of the species based on the acquired spectral channels, characterized in that the three channels of said set which contribute most to the performance of the prediction are chosen from: - a first spectral band [730-780]nm; - a second spectral band [550-660]nm.

2. Method for identifying a microorganism likely to be contained in a biological sample of a predefined type, comprising: A. the acquisition of at least one multispectral digital image of at most ten channels of at least one microbial colony having grown on a culture medium following the spreading of the sample on said medium; B. the prediction, implemented by computer, of the microbial species forming said colony from a predetermined panel of microbial species based on the digital multispectral image(s) acquired; characterized in that: - the panel of microbial species, the optical filters and the identification algorithm are chosen so that the specificity of the identification is greater than or equal to 90%, and preferably greater than or equal to 94%; - if the prediction model delivers the identity of a species o then said identity is confirmed without implementing another method of identifying the microorganism; o otherwise another method of identifying the microorganism is implemented.

3. Method according to claim 1 or 2, characterized in that each spectral channel has a width of between 8nm and 50nm.

4. Method according to claim 1, 2 or 3, wherein said set of channels comprises a fourth channel contributing most to the performance of the prediction included in the second spectral band.

5. Method according to one of the preceding claims, in which said set of channels comprises a fourth channel contributing most to the performance of the prediction included in the spectral band [490-540]nm.

6. Method according to one of the preceding claims, characterized in that - the acquisition is carried out by means of a monochromatic image sensor in front of which bandpass filters corresponding to the spectral channels are positioned in a removable manner; and - the spectral width at mid-height of said filters is between 8nm and 50nm.

7. Method according to any one of the preceding claims, characterized in that the set of spectral channels comprises at most 10 channels, and preferably at most 8 spectral channels.

8. Method according to any one of the claims, characterized in that the prediction of the species of the bacterial strain implements a decision tree in which each node comprises a prediction made as a function of a single spectral channel of the set of spectral channels.

9. Method according to any one of the preceding claims, characterized in that the bacterial species is a uropathogen among the species Escherichia coli, Enterococcus faecalis. Staphylococcus aureus, Proteus mirabilis, Proteus vulgaris, Klebsiella pneumoniae, Pseudomonas aeruginosa, and Streptococcus agalactiae.

10. Device for identifying a microbial colony from a predetermined panel of microbial species responsible for pathogenic infections of a predetermined type, in particular urinary infections, comprising: A. a digital camera configured to acquire a digital image of said colony; B. and a computer processing unit connected to said camera for receiving the digital image of the colony; characterized in that the digital camera comprises: A.

1. a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm; A.2.a lens capable of forming an image on said sensor in said range; A.3.a set of not more than ten optical bandpass filters centered respectively on different wavelengths in the wavelength range [390-900] nm, each having a bandwidth in the range [8-50] nm, each said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; and in that the three channels of said set which contribute most to the performance of the prediction are chosen from: - a first spectral band [730-780]nm; - a second spectral band [550-660]nm 11. Device for identifying a microbial colony from a predetermined panel of microbial species responsible for pathogenic infections of a predetermined type, in particular urinary infections, comprising: A. a digital camera configured to acquire a digital image of said colony; B. and a computer processing unit connected to said camera for receiving the digital image of the colony; characterized in that the digital camera comprises: A.

1. a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm; A.2.a lens capable of forming an image on said sensor in said range; A.3.a set of at most ten optical bandpass filters centered respectively on different wavelengths of the wavelength range [390-900] nm, each having a bandwidth in the range [8-50] nm, each of said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; in that the computer processing unit is configured to implement an algorithm for identifying the species forming the colony based on the digital images, in that the panel of microbial species corresponds to a prevalence of at least 70% of pathological infections, and preferably greater than or equal to 80%, and in that the panel of microbial species, the optical filters and the identification algorithm are chosen so that the specificity of the identification is greater than or equal to 90%, and preferably greater than or equal to 94%.

12. Device for characterizing a microbial colony, comprising: A. a digital camera configured to acquire a digital image of said colony; B. and a computer processing unit connected to said camera for receiving the digital image of the colony; characterized in that the digital camera comprises: A.

1. a monochromatic image sensor capable of capturing incident light in the wavelength range [390-900] nm; A.2.a lens capable of forming an image on said sensor in said range; A.3.a set of at most ten optical bandpass filters, each having a bandwidth in the range [8-50]nm, each of said filters being configured to be removably placed in front of the objective so as to acquire a multispectral digital image of the microbial colony; in that the set of at most 10 filters comprises: A.3.1.a first filter centered on a wavelength in the range [415-440]nm; A.3.2.second filters centered respectively on different wavelengths of the wavelength range [390-900] nm and in that the computer processing unit is configured to identify: B.1.1e gram of the species forming the colony based on the digital image of said colony acquired using the first filter; B.

2. the identity of the species forming the colony based on the digital images acquired using the second filters.

13. Device according to one of claims 10 to 12, characterized in that it is configured to implement a method according to one of claims 1 to 10.

Images