Method for detecting microbial colonies in sample placed in culture medium

Abstract

The invention relates to a method for detecting microbial colonies in a biological sample placed in a solid medium by means of a detection system comprising at least one incubator, an analysis unit and at least one automatic image capture system, comprising a first stage, the present invention relates to a method for detecting and counting objects exhibiting growth in the field of view of an image capture system, and to a second stage applied only to those samples deemed negative in the first stage of the method, the second stage being based on the detection of anomalies on at least the last image acquired in a culture sequence.

Description

Technical Field

[0001] This invention is particularly applicable to the field of microbial control of sterile products and their production environments. The invention relates to the field of contamination detection systems for petri dishes or other culture media (preferably solid media). More specifically, the invention relates to the field of colony-forming unit (CFU) detection systems on petri dishes, based on continuous automated readings of the petri dishes throughout the incubation period. Technical Background In the pharmaceutical industry, microbial environmental control of production areas is conducted regularly, especially in Grade A and Grade B controlled aseptic production areas. Traditionally, settling dishes are used for air analysis, or "contact" dishes are used for surface analysis, with incubation periods of 5 to 7 days. Results (whether contamination occurs) are read by the operator after incubation. This method of controlling aseptic areas has some specific limitations because: i. In accordance with established control plans and regulatory requirements, hundreds of petri dishes containing solid culture media may be collected daily; ii. Most samples collected in these areas were negative (between 95% and 99.9%); iii. Not just the exact number of colonies grown in solid culture media, but also the contamination recovery rate as defined in regulatory texts (“USP”). <1116> The qualitative status of the samples is monitored (CFU=0 is negative, CFU>0 is positive).

[0003] If contamination occurs, an investigation is conducted to identify the contaminating microorganisms, thereby understanding their type and / or strain, and tracing the cause of the contamination.

[0004] This traditional method is time-consuming, cumbersome, difficult to track, and relies heavily on the operator's skills (training, visual ability, etc.). Even as a reference, the method's performance is not perfect, with an estimated false negative rate of around 2%.

[0005] To make the method more efficient and traceable without compromising its performance, an automated reading system based on digital imaging and related processing has been developed.

[0006] On the one hand, there are traditional colony reading systems, known as "endpoint" systems, which replace manual reading by automatically recording and processing images taken on the culture dish at the end of the culture. These systems are suitable for counting applications when the microbial load (the number of microorganisms in a sample) is typically large (tens or even hundreds of colony-forming units) and the purpose is not to perform a qualitative test for presence / absence. However, in qualitative contamination detection applications, their performance is generally inferior to traditional methods in terms of sensitivity (the ability to accurately detect colonies when they are present) and specificity (the ability to avoid triggering false alarms).

[0007] Regardless of the quality of the optical acquisition system, image processing, or algorithm used, many artifacts (dust, stains, defects on the agar surface, etc.) do exist on the petri dish and are often mistaken for bacterial colonies. Figure 1 As shown, the basic information obtained contains an inherent trade-off between false positives and false negatives, which is inconsistent with the requirements of qualitative pollution detection applications.

[0008] To overcome the limitations of automated "endpoint" reading systems and achieve performance equivalence with traditional methods, particularly for environmental microbial control in aseptic production areas, a novel system called "continuous reading" has been developed. This system integrates an incubator and is based on automated reading, which involves capturing images throughout the incubation process and processing the resulting sequences.

[0009] These systems read data from sealed petri dishes to avoid cross-contamination in incubators containing multiple dishes and typically integrate multiple contrast sources (transmitted, diffuse, grazing illumination, etc.) to capture maximum information and compensate for reading biases that may occur due to the presence of the lid (which can be removed at the end of the culture in traditional reading methods). In these systems, continuous images of the petri dishes are typically recorded at fixed intervals (e.g., one image per hour) during the culture process, and the images are analyzed using image processing operations and common protocols during acquisition, such as... Figure 2 As shown.

[0010] These systems integrate the image processing chain with key steps such as intensity normalization (to eliminate variability caused by different lighting or changes in the growth medium) or image registration relative to the previous image, thereby allowing matching of detected objects and calculation of growth criteria.

[0011] Key parameters of the algorithm that allow for adjusting the balance between sensitivity and specificity include: - Threshold criteria for segmenting and extracting objects above the background. - The detected size change of the object, which allows for the declaration of growth between two images, - The number of times the object grows consecutively, to declare it as a colony-forming unit.

[0012] Compared to endpoint reading systems, these continuous reading systems can significantly improve performance because growth criteria are universal for colonies but not for artifacts, since the size of artifacts usually does not change, or changes rather randomly.

[0013] However, despite the significant improvements brought by continuous reading systems, some false positives and false negatives still exist, albeit in small numbers. For example, if a colony begins to grow in a shaded area (near the edge of the petri dish, in shadows, or near stationary particles), the algorithm will fail to detect it because it will quickly be labeled as an artifact (appearing in the first image after the minimum number of consecutive growths has been reached). Similarly, shadows that gradually increase in size across several images will also be labeled as colonies. Clearly, colonies with atypical growth behavior can be detected by adjusting algorithm parameters (segmentation threshold, size change threshold, and number of consecutive growths), but this comes at the cost of increasing the false positive rate. Therefore, a certain false positive and false negative rate still exists in growth detection methods.

[0014] These false negatives and false positives remain a major problem in the application of sterile area environmental monitoring: currently, in such continuous reading systems, the false positive rate must be kept at around 5% to 10% to ensure that the false negative rate is comparable to that of traditional methods (about 2%), which still raises questions about justifying the economic benefits of automating traditional methods. Summary of the Invention

[0015] The present invention aims to overcome all or part of the above-mentioned defects, and in particular to further reduce the number of false negatives and / or false positives in continuous reading systems for culture dishes.

[0016] Therefore, the present invention relates to a method for detecting microbial colonies in a sample of a solid culture medium, the method being implemented by a detection system comprising at least one incubator, an analysis unit, and at least one automated image capture system, the method comprising a first stage, namely, detecting and counting objects exhibiting growth within the field of view of the image capture system, the first stage comprising at least the following steps: • During the incubation of the sample in an incubator, multiple images of the culture medium within the field of view of the image capture system are acquired using an image capture system. • Objects in the detected image are segmented using a thresholding method; • Extract parameters showing the size increase of the detected objects across consecutive images. • The presence or absence of microorganisms is determined by detecting the growth of at least one object using the analysis unit: (i) if the growth of the object is detected, the analysis unit considers at least one microbial colony to be present in the sample, and the sample is designated as positive; (ii) if the growth of the object is not detected, the analysis unit considers no microbial colony to be present in the sample, and the sample is designated as negative. • The analysis unit counts the objects exhibiting growth. The method is characterized by further including a second stage applied only to samples deemed negative in the first stage of the method, the second stage being based on anomaly detection applied to at least the last image acquired during the culture sequence, and comprising the following steps: • Objects in the acquired image are detected by segmenting using a thresholding method; • Extract the location and / or size and / or shape and / or intensity parameters of the detected object. • Apply a discriminative model estimated from the learning database contained in the analysis unit, for example, a score calculated based on all parameters extracted from the extraction step, to distinguish anomalies from artifacts.

[0017] This method has several advantages, detailed below.

[0018] According to the method of the present invention, even if this means a reduction in sensitivity (the ability to detect colonies), specificity (the ability not to trigger false alarms) can be obtained in the first stage, and sensitivity is guaranteed in the second stage, which ultimately enables a reduction in the false positive rate in addition to the conventional goal of reducing the false negative rate.

[0019] The method of this invention is intended only to identify anomalies associated with the presence of one or more colonies: the classification criteria for these anomalies differ from those conventionally used for automated reading of petri dishes at the endpoint, particularly including the location of the colonies. Specifically, if an object detected in the second stage is far from the edge of the petri dish (the center of the dish) and was not detected in the first stage of the algorithm, then that object is more likely to be an artifact.

[0020] Since the second stage is primarily applicable to the last image obtained from the sequence at the end of the culture, the colonies present at this stage are usually clearly visible and larger than 500µm in size, making them exceptionally clear and distinguishable. In contrast, in the first stage, detection is performed when the object appears, and the object to be detected is much smaller, only on the order of tens of micrometers.

[0021] Furthermore, the second phase also significantly impacted the development of the first-phase model. Specifically, in state-of-the-art continuous readout detection / counting systems, the initial constraint for model optimization is false negatives (the most significant risk being allowing contamination to pass through). Based on error estimates using conventional methods, the target rate is set at around 2%. Based on this target rate, the algorithm is optimized using image sequences encompassing all possible scenarios, particularly for so-called "hard-to-detect" cases—colonies growing at the edge of the petri dish, with low contrast, and / or slow growth. The algorithm criteria (the threshold for segmenting the object and extracting it from the background, the rate of size change between two images indicating growth, and the number of consecutive growths required to annotate the object as a colony-forming unit) are applicable to fully achieving this target rate across a complete database of all possible scenarios. The model may be "optimal" across the entire database, or "suboptimal" for a database consisting only of simple cases. Since, in the principles of this invention, the second phase is applied only to samples judged negative at the end of the first phase, the optimization criteria for the first phase can be modified, particularly by modifying the computational basis used to estimate the false negative rate: cases that are, in principle, difficult to detect, will be removed based on the growth criteria, as they will be recovered in the second phase specifically for these cases. These cases must meet the following two conditions to be removed from the Phase 1 model database: 1. Removing these cases can improve performance (reduce false positive and false negative rates), but only in "simple" cases; 2. Even if the colonies cannot be fully identified, their presence will produce a clearly visible anomaly in the last image.

[0022] In this application, the term "object" refers to an element detected on an image by a thresholding / segmentation algorithm of the analysis unit. Both stages of the detection method are based on the first step of object detection. However, the detected objects are not necessarily the same because (i) on the one hand, the algorithms are not necessarily the same; (ii) on the other hand, the contrast mode (or contrast fusion) used is not necessarily the same; (iii) lastly, but equally importantly, the timing of object detection is not the same: when it appears during incubation (small objects), and on the final point image of the second stage of the algorithm (larger objects in principle).

[0023] In this invention, the term "abnormality" refers to an object detected in the second stage of the algorithm and classified as a colony-forming unit. The detected object may not resemble a colony; it may actually correspond to a "fusion" between a colony and the edge of the petri dish, or even multiple fused colonies, hence the term "abnormality."

[0024] In this invention, the term "artifact" refers to an object that is detected by the algorithm but not recognized as a colony-forming unit by the algorithm (whether correctly or incorrectly). Similarly, artifacts may exist in the first stage and possibly in the second stage; these objects will be detected on the final image but will ultimately not be classified as colony-forming units by the discrimination model.

[0025] According to the features of the invention, the second stage may include one or more additional preprocessing steps, such as fusion of different contrast modalities obtained at the endpoint, conversion to grayscale (8 bits or more), normalization of image intensity to compensate for possible illumination variations, convolutional filtering, and definition of the region of interest.

[0026] According to one aspect of the invention, the morphological parameter can be the convexity and / or roundness of the detected object. Advantageously, morphological classification parameters are less demanding than positional classification parameters because the objects extracted from the image can be composed of a fusion of colonies (rather round and convex objects) and markers, such as... Figure 4 As shown. Therefore, the sensitivity and specificity targets of this anomaly detection layer are quite different from those of traditional image-based colony detection applications. For example, consider a normalized morphological parameter, such as the convexity of the object, ranging from 0 to 1 (1 represents a perfect disk, 0 represents a crescent shape), and similarly normalized positional criteria, such as the relative distance from the center (0 if the object is at the center, and 0 if the object is at the very edge of the culture medium); then, if a model with an acceptable range is used, we would, for example, for the second stage: - For example, if the position criterion is between 0.8 and 1 (i.e., a very off-center object) and the convexity criterion is between 0.3 and 1 (a wider range is acceptable because there may be two fused colonies that do not form a disc but rather a slightly concave shape), then an anomaly detection is performed. - Perform artifact detection on any detected items that are not considered abnormal.

[0027] According to one feature of the invention, the dimensional parameters may be the diameter and / or perimeter and / or area of ​​the object being tested.

[0028] According to the features of the present invention, the intensity parameter can be the average gray level of the detected object and / or the variation of gray level within the object.

[0029] According to one feature of the invention, a score calculated from the set of extracted parameters can be used to classify targets. This score can be calculated using one of the following methods: regression analysis, decision tree, linear discriminant analysis, support vector machine, or neural network. Advantageously, depending on the different calculation criteria for the extracted targets, a discriminant model is estimated based on a restricted and specific learning library containing only colonies not detected in the first stage and all possible artifacts. This allows for the recovery of missed colonies without increasing false positives.

[0030] According to one feature of the invention, the second stage may further include a step of fusing different contrast modes and a step of converting to grayscale, such as 256 levels of 8-bit grayscale, which are performed before the object segmentation step.

[0031] According to one feature of the present invention, the present invention can be applied to fluorescence image capture.

[0032] According to one feature of the invention, when the analysis unit determines that the sample is positive, an alarm will be sent to the user regardless of the stage.

[0033] According to one feature of the invention, the settings of the first stage are optimized on a database of normal cases by excluding problematic cases (especially colonies at the edge of the culture medium).

[0034] In this application, "case" refers to a colony or artifact present in an image sequence.

[0035] The database consists of multiple inoculated or uninoculated culture media. Based on the average inoculation level (e.g., 5 colony-forming units (CFU)), and since each culture medium contains multiple cases, the database can contain approximately one thousand cases (colonies or artifacts).

[0036] Among these cases (colonies or artifacts) in the database, we categorize them into two types: (i) nominal cases; and (ii) marginal cases. Marginal cases can also be referred to as "problem cases." We only specify marginal cases; all non-marginal cases are defined as nominal cases.

[0037] These marginal cases correspond to cases of colonies or artifacts with characteristics such as incorrect prediction and false positives (the algorithm incorrectly predicts artifacts as CFUs, or conversely, the algorithm misses true CFUs and treats them as artifacts).

[0038] According to the present invention, the edge case corresponds to: - Colonies that begin to grow in shaded areas (edges of petri dishes, shaded areas, or near anchoring particles). - The shadow, which gradually increases in size across several images, could potentially be mistaken for UFC.

[0039] The nominal case is all other cases, namely colonies that do not begin to grow in the occluded area or near fixed particles, and artifacts whose appearance does not change between images.

[0040] According to one feature of the invention, the second stage further includes a step of preprocessing the last acquired image, which involves another image in the sequence, such as the first acquired image. Brief description of the attached diagram The invention will be better understood from the following description, which relates to embodiments of the invention, given by way of non-limiting examples and explained with reference to the accompanying drawings. The accompanying schematic diagrams are as follows: Figure 1 This is a view used to read the different current systems (a), (b), and (c) of the culture dish.

[0042] Figure 2 This is a general process view of a "continuous" automatic reading system.

[0043] Figure 3 This is a view of the general method for two-phase detection used in this invention.

[0044] Figure 4 This is a schematic diagram of the second stage of anomaly detection.

[0045] Figure 5 This is a diagram illustrating the method according to the present invention.

[0046] Figure 6 This is a view explaining the impact of the existence of the second stage on the optimization criteria of the first stage of the algorithm.

[0047] Figure 7 This is an illustration of a first image processing method for object detection according to the present invention.

[0048] Figure 8 This is a schematic diagram of a second image processing method for object detection according to the present invention.

[0049] Figure 9 A detection system according to the present invention is shown.

[0050] Detailed description Figure 1 The prior art for colony detection and counting systems is demonstrated. A petri dish (3) is schematically shown in the figure, in which an artifact (2) is placed and a colony (1) is placed at the edge of the petri dish (3). Furthermore, a graph is plotted to show the performance of the current methods (a), (b), and (c), with a false negative rate of approximately 2%.

[0051] (a) illustrates the traditional method, which consists of an open inspection performed by a qualified operator at the end of the culture. Disadvantages of this method include the fact that the inspection involves variable geometries and relies on manual interpretation. While effective, this traditional method is time-consuming, cumbersome, and difficult to trace.

[0052] (b) and (c) illustrate automated methods designed to maintain the performance of the traditional method (a).

[0053] (b) illustrates an automated endpoint reading method implemented by a camera-based optical system, performed at the end of the culture, with or without a cap. A drawback of this method is the simplicity of image processing for inspection: therefore, artifacts are frequently detected as colonies due to the lack of sophisticated geometry and growth checks. In examining the graphs related to method (b), it is noted that it is impossible to achieve an application-compatible trade-off between sensitivity (z) and specificity (w) (the shaded area marked 4 on the graph), where the AA axis represents the decision threshold.

[0054] (c) illustrates an automated kinetic readout method performed by a camera-based optical system throughout the culture process, with a lid (5). A significant drawback of this method is the simplicity of image sequence processing for inspection: therefore, colonies (1) at the edge of the petri dish often go undetected because their growth begins in the masked periphery (the edge of the petri dish), obscuring their appearance. When examining the figures related to method (c), a much better trade-off between sensitivity (z) and specificity (w) is found compared to (b), but some areas still require improvement: growth in the periphery, the initiation of masked growth, etc. (the shaded area marked 4 on the figure).

[0055] Figure 2 illustrates the general flow of the current detection and counting method (c). Figure 2 shows a series of images acquired during the culture process of the petri dish (3) (indicated by the arrows above). The acquired series of images will be recorded (step 1), followed by object detection (step 2), size extraction (step 3), and similar objects will be grouped (step 4). After completing these four steps, the growth analysis shown in Figure 3 (step 5) can be performed.

[0056] exist Figure 3 In the diagram, the y-axis represents the size of the detected object, and the x-axis represents time. (Back) Figure 2 According to the growth analysis, reference element (1) was identified as a colony, while reference element (2) was identified as an artifact because the size of the latter did not change over time.

[0057] Figure 4 illustrates the anomaly detection principle related to CFU.

[0058] The first image (a) is a photograph of the petri dish (3) at the endpoint. Colonies growing at the edge of the petri dish (3) are circled. The analysis unit processes this image, specifically segmenting it to highlight the object, as shown in (b). The enclosed object (2) is treated as a whole, including the colonies growing at the edge of the petri dish, because its outline is confused with the object in the petri dish. In conventional detection methods, the algorithm used by the analysis unit to determine the presence or absence of CFU treats the colony (1) as an artifact (2) because the latter is contained in the shape of the artifact, and the resulting shape does not correspond to the shape standard of the colony, nor to the growth standard of the colony. According to the invention, by means of the steps of the method according to the invention, the same object can be detected as an anomaly related to CFU even if the morphology of the same object is close to that of an artifact.

[0059] Figure 5 A detection method according to the present invention is shown. The detection method (200) according to the present invention comprises... Figure 9 The detection system (300) shown is implemented, which includes at least one incubator (301), an analysis unit (302) and at least one automatic image capture system (303).

[0060] As shown in Figure 5, the detection method includes a first stage (PI), which involves detecting and counting objects showing growth within the field of view of the image capture system (303). This first stage includes at least the following steps: (P101) - During the incubation of the sample in the incubator, multiple images of at least one object within the field of view of the image capture system (303) are acquired by the image capture system. (P102) - The presence or absence of microorganisms is determined by detecting the growth of the at least one object using the analysis unit (302): (i) if the growth of the object is detected, the analysis unit considers that at least one microbial colony is present in the sample, and the sample is positive; (ii) if the growth of the object is not detected, the analysis unit considers that no microbial colony is present in the sample, and the sample is negative. (P103) - The analysis unit counts the objects that show growth. (P300) - Notifies the user of the number / presence of samples considered positive.

[0061] In addition, such as Figure 5 As shown, the method according to the invention further includes a second stage (P2), which is applied only to samples (ii) deemed negative in the first stage (PI) of the method. The second stage (P2), based on anomaly detection performed on at least the last image acquired in the culture sequence, includes the following steps: (P201) Objects in the acquired image are detected by using a thresholding method; (P202) Extract the position and / or size and / or shape and / or intensity parameters of the detected object. (P203) A discriminative model estimated from the training database included in the analysis unit is applied, and scores calculated based on all parameters extracted from the extraction step are used to separate anomalies from artifacts. (P204) The score is compared to a threshold to determine if there is an anomaly, i.e., colony-forming units that should have been detected (case (i)) or if it is an artifact (case (ii)). In case (i), an alarm is triggered to warn the user of the presence of colony-forming units, indicating the presence of microorganisms; in case (ii), the sample is considered negative: no microorganisms are present.

[0062] Figure 6 shows the impact of the second anomaly detection phase following the first phase.

[0063] Figure 6 (a) illustrates a classic method whose performance is not ideal under continuous reading conditions. In order to ensure good segmentation of “difficult” colonies at the edge of the petri dish (1), the threshold must be adjusted, which results in the detection of “normal” colonies in the easy region (2).

[0064] Figures 6(b) and (c) illustrate the two-stage method according to the present invention. Unlike the classical method shown in (a), the model in the first stage is clearly optimized for “perfect” detection under simple condition (2), and “missed” colonies (1) will be excluded from the learning database due to their proximity to the labeled region (see (b)). The “missed” colonies (1) in the first stage (b) will be recovered in the second stage (c) because they are clearly visible in the last image of the acquired image sequence.

[0065] Figure 7 illustrates a first image processing method for object detection, comprising: performing a first image processing step, i.e., segmentation, on the acquired image (a), and then performing a second processing step, i.e., discrimination (c).

[0066] Figure 8 illustrates a second image processing method for object detection, which includes acquiring a first image (d) of the sequence and acquiring a last image (f) of the sequence. This processing method allows for determining the differences between the two images (g) and then performing a discrimination (h). Of course, the present invention is not limited to the embodiments described and illustrated in the figures. Modifications can be made without departing from the scope of the invention, particularly from the perspective of the configuration of individual elements or from the perspective of equivalent substitution techniques.

Claims

1. A method for detecting microbial colonies in a sample placed in a solid culture medium, the method being implemented by a detection system comprising at least one incubator, an analysis unit, and at least one automated image capture system, the method comprising a first stage (P1) of detecting and counting objects showing growth within the field of view of the image capture system, the first stage (P1) comprising at least the following steps: (P101) During the incubation of the sample in the incubator, a series of images of the culture medium within the field of view of the image capture system are acquired using an image capture system. (P102) Objects in the detected image are segmented using a thresholding method; Extract parameters for object size growth between consecutive images from the detected objects. The presence or absence of microorganisms is determined by detecting the growth of at least one object using an analysis unit: (i) if object growth is detected, the analysis unit considers at least one microbial colony to be present in the sample, and the sample is designated as positive; (ii) if no object growth is detected, the analysis unit considers no microbial colony to be present in the sample, and the sample is designated as negative. (P103) The analysis unit counts the objects that exhibit growth. Its features are, The method also includes a second stage (P2), which is applied only to samples that were considered negative in the first stage of the method. The second stage (P2) is based on anomaly detection performed on at least the last image acquired during the culture sequence and includes the following steps: (P201) Objects in the detected image are segmented using a thresholding method; (P202) Extract the position and / or size and / or shape and / or intensity parameters of the detected object. (P203) Apply, for example, a discriminative model estimated from a learning database contained in the analysis unit, and calculate a score based on the set of parameters extracted from the extraction step, in order to separate anomalies from artifacts.

2. The detection method as described in claim 1, wherein the second stage may include one or more additional preprocessing steps, such as fusion of different contrast modes, conversion to grayscale, convolutional filtering, and definition of the region of interest.

3. The detection method according to any one of claims 1 or 2, wherein the morphological parameters may be the convexity and / or the roundness of the object being detected.

4. The detection method according to any one of claims 1 to 3, wherein the morphological classification parameters are less stringent than the positional classification parameters.

5. The detection method according to any one of claims 1 to 4, wherein the size parameter may be the diameter and / or perimeter and / or area of ​​the object being detected.

6. The detection method according to any one of claims 1 to 5, wherein the intensity parameter may be the average gray level of the object being detected and / or the variation of gray level within the object.

7. The detection method of any one of claims 1 to 6, wherein the score calculated from the extracted set of parameters can classify the object and can be calculated using one of the following methods: regression, decision tree, linear discriminant analysis, support vector machine, neural network.

8. The detection method of any one of claims 1 to 7, wherein a discriminant model is estimated from a limited and specific learning database containing only colonies not detected in the first stage and all possible artifacts, based on different criteria calculated for the extracted objects.

9. The detection method according to any one of claims 1 to 8, wherein the second stage (P2) further includes a step of fusing different contrast modalities and a step of converting to grayscale, such as 256 levels of 8-bit grayscale, which are performed before the object segmentation step.

10. The detection method as described in any of the preceding claims, wherein the settings of the first stage are optimized based on a nominal case database by excluding problematic cases.

11. The detection method as described in any of the preceding claims, wherein the second stage further includes a step of preprocessing the last acquired image, the preprocessing involving another image in the sequence, for example, the first acquired image.

Images