A device and method for the enrichment and detection of microorganisms in a sample

Abstract

The application discloses a device and method for enriching and detecting microorganisms in a sample, which comprises an enrichment filter and an automatic detection module, wherein the filter membrane assembly is arranged obliquely in the enrichment filter, and the filter membrane assembly divides the enrichment filter into an upper space and a lower space; a sample adding port is arranged at the upper part of the enrichment filter, and a suction port is arranged at the lower end of the enrichment filter; the automatic detection module is used for acquiring the image inside the enrichment filter, automatically detecting the turbidity of the culture medium and the distribution and size of the spherical or cluster targets, and grading the microbial contamination degree of the culture medium based on the turbidity of the culture medium and the distribution and size of the spherical or cluster targets on the filter membrane assembly. In the application, the filter membrane assembly is arranged obliquely, thereby reducing the risk of filter membrane blockage; the automatic detection module replaces manual naked-eye observation through image acquisition and intelligent algorithm analysis, can accurately capture the subtle changes in the turbidity of the culture medium and the spherical or cluster target conditions, and significantly improves the reliability of the detection data.

Description

Technical Field

[0001] This invention relates to the field of sterility testing technology for pharmaceuticals or food, and particularly to a device and method for detecting microbial enrichment in samples. Background Technology

[0002] Currently, sterility testing of pharmaceuticals or food typically employs membrane filtration or direct inoculation methods. The presence of microbial growth is determined by observing changes in the turbidity of the culture medium within the enrichment filter, thus confirming the sample's suitability. When the inoculation volume is large, membrane filtration is often used for sterility testing. If microbial growth occurs during inoculation, changes in turbidity in the culture medium, or the appearance of flocculent, spherical, or other foreign matter, can be observed visually by laboratory personnel. The presence of any of these conditions indicates the presence of microorganisms within the enrichment filter.

[0003] The sterility testing of pharmaceuticals requires a significant amount of manual labor to periodically observe culture results, preventing laboratory personnel from devoting more energy to other tasks and reducing their work efficiency. The repeatability and accuracy of manual interpretation are difficult to guarantee, and the results observation relies on the operator's experience, which can sometimes lead to misjudgments or problems due to unclear changes in the culture medium. The time for periodic observation is also limited by working hours, making it impossible to quickly and timely obtain abnormal results and process production batches in a timely manner.

[0004] Besides sterility testing for pharmaceuticals, in the food industry, during qualitative microbiological testing, after samples undergo homogenization, dilution, and other processes before microbial culture, food residues can easily be misidentified as bacterial colonies. Preservatives in food can also cause false negatives. Furthermore, if enrichment filters are used for qualitative microbiological testing of food, food residues can easily clog the filter membrane. Similarly, in qualitative microbiological testing of swimming pool water, impurities in the water can be misidentified as bacterial colonies, and substances such as chlorine used for water disinfection can also cause false negatives. If enrichment filters are used for qualitative microbiological testing of water quality, some impurities can easily clog the filter membrane.

[0005] In summary, the following problems exist in the current sterility testing of pharmaceuticals or food: 1. During manual observation, there are issues such as time-consuming interpretation of results, poor timeliness in handling abnormal results, inability to guarantee accuracy and repeatability, poor traceability of results, and inability to guarantee data integrity. 2. During filtration, the filter membrane is easily clogged by impurities (especially food residues), resulting in low filtration efficiency. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art and provide a device and method for detecting microbial enrichment in samples.

[0007] The objective of this invention is achieved through the following technical solution: a microbial enrichment detection device for samples, comprising an enrichment filter and an automatic detection module. A filter membrane assembly is inclinedly arranged within the enrichment filter, dividing the interior of the filter into an upper filter membrane space and a lower filter membrane space. A sample addition port is provided at the upper part of the enrichment filter, and a filtration port is provided at the lower end of the enrichment filter. The automatic detection module is used to acquire images of the interior of the enrichment filter and automatically detect the turbidity of the culture medium and the distribution and size of spherical or clump-like targets. Based on the turbidity of the culture medium and the distribution and size of the spherical or clump-like targets on the filter membrane assembly, the microbial contamination level is rated.

[0008] Preferably, the angle of inclination between the filter membrane assembly and the horizontal plane is 45°-75°, and the filter membrane assembly is composed of a filter membrane and a supporting frame.

[0009] Preferably, the sampling port is a puncture sampling port.

[0010] Preferably, the enrichment filter has at least one observation plane on its side, through which the automatic detection module acquires an image of the interior of the enrichment filter.

[0011] Preferably, the enrichment filter is provided with an openable and closable exhaust port at its upper end, and an air filtration component is provided in the exhaust port.

[0012] A method for detecting microbial enrichment in a sample includes the following specific steps:

[0013] Step 1: Open the filtration port and pass the sample into the enrichment filter through the sample feeding port. The microorganisms in the sample are enriched in the upper space of the filter membrane by filtration. Step 2: Pass the rinsing solution into the enrichment filter through the sample addition port and rinse it; Step 3: Close the filtration port, introduce culture medium into the enrichment filter through the sample feeding port, and incubate. During the incubation process, the automatic detection module periodically acquires images of the enrichment filter and automatically detects the turbidity of the culture medium and the distribution of spherical or clump-shaped targets on the filter membrane assembly. Based on the turbidity of the culture medium and the distribution and size of the spherical or clump-shaped targets, the microbial contamination level is rated.

[0014] As a preferred embodiment, the specific detection method of the automatic detection module in step three is as follows: during the initial culture, the automatic detection module acquires the initial state image, cuts out the culture medium portion of the initial state image, scales it to the first preset image size, and uses this image as the culture medium background image; During the cultivation process, images are acquired, and the culture medium portion and filter membrane portion of the current image are cropped out respectively. The culture medium portion image is scaled to a first preset image size and used as the current frame image of the culture medium. The current frame image of the culture medium and the culture medium background image are input into the first vision algorithm module. Then, the cropped filter membrane portion is scaled to a second fixed size and used as the current frame image of the filter membrane, which is then input into the second vision algorithm module. The confidence level of culture medium turbidity is predicted by the first vision algorithm module, and the position coordinates and size of all spherical or clump-shaped targets on the filter membrane assembly are predicted by the second vision algorithm module. The turbidity confidence score output by the first vision algorithm module and the number of bounding boxes and the total area of ​​bounding boxes of spherical or clump-shaped targets output by the second vision algorithm module are input into the decision tree model to rate its microbial contamination level.

[0015] Preferably, the decision tree model is a CART classification tree with 3-dimensional input and 11-dimensional output. The priority of features is determined by calculating the Gini coefficient, the relative area ratio is calculated by the total area of ​​the bounding boxes, and corresponding weights are assigned to the number of bounding boxes, turbidity confidence, and relative area ratio. The contamination index is obtained by weighting and summing the number of bounding boxes, turbidity confidence, and relative area ratio, and the current microbial contamination level is determined based on the contamination index.

[0016] Preferably, the first vision algorithm module uses a regression network model whose output value is between 0 and 1. During training, the loss function uses the mean squared error loss, as shown in the following formula: ; In the formula, loss represents the error loss, and N is the number of samples. It is the actual value of the i-th sample. It is the predicted value of the i-th sample.

[0017] Preferably, the detection framework of the second vision algorithm module uses a single-stage multi-box detector, and the loss function for the positive and negative sample classification of the bounding box during training uses the focal loss function, as shown in the following formula:

[0018] Where loss represents the error loss for classifying a positive or negative sample, p represents the predicted value, y represents the actual value, and γ is an adjustable parameter greater than 0. The loss function used in the bounding box coordinate regression is the smoothed L1 loss function, as shown below: ; Where loss represents the error loss for the i-th sample. It is the actual value of the i-th sample. It is the predicted value of the i-th sample. Parameters used to control the smoothing range.

[0019] The beneficial effects of this invention are: 1. In this invention, the filter membrane assembly adopts an inclined arrangement design. Compared with traditional horizontal or vertical filter membranes, the advantages of the inclined arrangement are: firstly, it increases the actual filtration area; secondly, it causes the sample to flow through the filter membrane in a tangential flow. Large particulate impurities in the sample (such as food residue, water sediment, and difficult-to-filter components in cell therapy drugs) will collect at the bottom corner of the filter membrane under the action of gravity and hydrodynamics, avoiding clogging caused by impurities uniformly covering the filter membrane surface, significantly reducing the risk of filtration failure. It is especially suitable for the detection of food samples with a large number of impurities and complex water samples. The inclined arrangement of the filter membrane assembly reduces the accumulation of impurities on the filter membrane surface, ensuring the unobstructed flow of the filtration channel. Compared with traditional horizontal filter membranes, it effectively shortens the filtration time. At the same time, the vacuum-assisted design of the suction port further accelerates the liquid filtration rate, solving the problem of low filtration efficiency in the prior art.

[0020] 2. The automatic detection module incorporates a camera module. Through image acquisition and intelligent algorithm analysis, it replaces manual visual observation, accurately capturing subtle changes in culture medium turbidity and the distribution and size data of spherical and clump-like targets (such as microbial colonies and flocculent contaminants). The detection process is unaffected by subjective factors such as operator experience and visual fatigue, effectively avoiding misjudgments and omissions in manual interpretation. It also ensures the repeatability of test results across different times and operators, significantly improving the reliability of the test data. Furthermore, during detection, the automatic detection module rates the degree of microbial contamination based on culture medium turbidity data and quantitative indicators such as the distribution density and total size of spherical and clump-like targets. This upgrades the test results from a simple "pass or fail" output to a microbial contamination level output, achieving multi-level quantitative assessment of contamination levels. This makes the test results more scientific and valuable, providing more accurate data support for subsequent quality control. The automatic detection module can continuously and dynamically acquire and analyze images of the enrichment filter's interior without requiring regular manual observation and is not limited by working hours. Once an anomaly is detected (such as a sudden change in turbidity or a surge in the number of target substances), it can respond quickly and output results, solving the problem of delayed processing of abnormal results in traditional manual detection. This makes it easier for experimental personnel to promptly initiate deviation investigations and reduce the risk of production batches. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the external structure of the enrichment filter of the present invention.

[0022] Figure 2This is a schematic diagram showing the relative positions of the enrichment filter and the automatic detection module.

[0023] Figure 3 This is a cross-sectional view of the enrichment filter.

[0024] Figure 4 This is a schematic diagram of the filter membrane assembly.

[0025] Figure 5 This is a schematic diagram of the filter membrane fixation method.

[0026] Figure 6 This is a flowchart of the method of the present invention.

[0027] In the diagram: 1. Enrichment filter, 2. Filter membrane assembly, 3. Upper structure, 4. Lower structure, 5. Filter port, 6. Sample inlet, 7. Exhaust port, 8. Automatic detection module, 9. Support frame, 10. Filter membrane, 11. Bottom accumulation area. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0029] Those skilled in the art should understand that, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention.

[0030] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0031] like Figures 1 to 6As shown, a microbial enrichment detection device for samples includes an enrichment filter 1 and an automatic detection module 8. A filter membrane assembly 2 is inclinedly arranged inside the enrichment filter 1, dividing the interior of the enrichment filter 1 into an upper filter membrane space and a lower filter membrane space. A sample addition port 6 is provided at the upper part of the enrichment filter 1, and a suction port 5 is provided at the lower end of the enrichment filter 1. The automatic detection module 8 is used to acquire images of the interior of the enrichment filter 1 and automatically detect the turbidity of the culture medium and the distribution and size of spherical or clump-like targets. Based on the turbidity of the culture medium and the distribution and size of the spherical or clump-like targets on the filter membrane assembly 2, the device rates the degree of microbial contamination.

[0032] In this invention, the filter membrane assembly 2 adopts an inclined arrangement design. Compared with traditional horizontal or vertical filter membranes, the advantages of the inclined arrangement are: firstly, it increases the actual filtration area; secondly, it causes the sample to flow through the filter membrane in a tangential flow, causing large particulate impurities in the sample (such as food residue, water sediment, and difficult-to-filter components in cell therapy drugs) to collect at the bottom corner of the filter membrane under the action of gravity and hydrodynamics, avoiding clogging caused by impurities uniformly covering the filter membrane surface, and significantly reducing the risk of filtration failure. This is especially suitable for the detection of food samples with a large number of impurities and complex water samples. The inclined arrangement of the filter membrane assembly 2 reduces the accumulation of impurities on the filter membrane surface, ensuring the unobstructed flow of the filtration channel and effectively shortening the filtration time compared with traditional horizontal filter membranes. At the same time, the vacuum-assisted design of the suction port 5 further accelerates the liquid filtration rate, solving the problem of low filtration efficiency in the prior art.

[0033] The automatic detection module 8 incorporates a camera module. Through image acquisition and intelligent algorithm analysis, it replaces manual visual observation, accurately capturing subtle changes in culture medium turbidity and the distribution and size data of spherical and clump-like targets (such as microbial colonies and flocculent contaminants). Its detection process is unaffected by subjective factors such as operator experience and visual fatigue, effectively avoiding misjudgments and omissions in manual interpretation. It also ensures the repeatability of detection results across different times and operators, significantly improving the reliability of the detection data.

[0034] Secondly, during detection, the automatic detection module 8 rates the degree of microbial contamination based on the turbidity data of the culture medium and quantitative indicators such as the distribution density and total size of spherical and clump-shaped targets. This upgrades the detection results from a simple "pass or fail" output to a microbial contamination level output, achieving multi-level quantitative assessment of contamination severity. This makes the detection results more scientific and valuable, providing more accurate data support for subsequent quality control. The automatic detection module 8 can continuously and dynamically acquire and analyze images of the enrichment filter 1, eliminating the need for regular manual observation and not being limited by working hours. Once an anomaly is detected (such as a sudden change in turbidity or a surge in the number of target substances), it can quickly respond and output results, solving the problem of delayed handling of abnormal results in traditional manual detection. This facilitates timely deviation investigation by laboratory personnel, reducing the risk of production batches.

[0035] This device is not only suitable for traditional pharmaceutical sterility testing, but also effectively addresses microbial testing scenarios in the food industry involving samples with residues and water bodies containing impurities, such as swimming pool water. It can avoid interference from food residues and water impurities on test results (such as misjudging them as bacterial colonies). At the same time, the sample inlet 6 located at the top of the enrichment filter 1 is not only used for sample input, but also for introducing rinsing fluid into the enrichment filter 1. The rinsing fluid can further remove interfering components such as preservatives and chlorides from the sample enrichment, reducing the risk of false negatives, significantly broadening the applicability of the device and enhancing its practical application value.

[0036] The filter membrane assembly 2 is composed of a filter membrane 10 and a support frame 9. The filter membrane 10 is fixed on the support frame 9, and there is a certain distance between the edge of the filter membrane 10 and the edge of the support frame 9, that is, the external dimensions of the filter membrane 10 are slightly smaller than those of the support frame 9.

[0037] To accommodate the membrane assembly 2 within the enrichment filter 1, the enrichment filter 1 comprises an upper structure 3 and a lower structure 4. These three components—the membrane assembly 2, the upper structure 3, and the lower structure 4—are welded together using UV curing technology. The wall thickness of the upper structure 3 is slightly greater than that of the lower structure 4 to ensure that the welded joint between the two components does not obstruct the flow of liquid to the suction port 5. Figure 5 As shown, the filter membrane 10 is fixed between the upper structure 3 and the support frame 9 of the enrichment filter 1. The edge of the filter membrane 10 is located at 1 / 2 of the wall thickness of the upper structure 3 of the enrichment filter 1. This design can ensure that there is no dead angle at the angle between the filter membrane 10 and the upper structure 3 of the enrichment filter 1, and also ensure that the filter membrane 10 does not come into contact with the external space of the enrichment filter 1.

[0038] The tilt angle between filter membrane assembly 2 and the horizontal plane is 45°-75°. For example... Figure 3As shown, the tilt angle between the filter membrane assembly 2 and the horizontal plane is angle α in the figure, forming a bottom accumulation area 11 at the bottom of the upper side of the filter membrane assembly 2. This tilt angle range allows the test sample to form a stable and strong tangential flow when flowing through the filter membrane. Under the combined effect of gravity and hydrodynamics, large particulate impurities in the sample (such as food residue, water sediment, and difficult-to-filter components in cell therapy drugs) will quickly gather towards the bottom corner of the filter membrane, avoiding clogging caused by the uniform accumulation of impurities on the filter membrane surface, thus improving the anti-clogging effect. At the same time, the tilt angle of 45°-75° ensures that the intensity of the tangential flow is sufficient to drive the migration of impurities, while avoiding the problem of weakened tangential flow and easy retention of impurities when the angle is too small (less than 45°). It also avoids the defects of a significant reduction in the effective filtration area and decreased filtration efficiency when the angle is too large (greater than 75°), thus achieving a balance between anti-clogging and filtration area.

[0039] Furthermore, this device employs an automatic detection module 8 for automatic detection. The automatic detection module 8 is located outside the enrichment filter 1, and its tilt angle of 45°-75° makes the spherical and clump-shaped targets (such as microbial colonies) attached to the filter membrane assembly 2 more evenly distributed. It is also less likely that the targets will aggregate and obstruct due to an excessively large angle, or that the targets will be scattered and difficult to identify due to an excessively small angle. This makes it easier for the automatic detection module 8 to clearly capture the distribution and size information of the targets. Moreover, at this angle, it is beneficial for the automatic detection module 8 to be arranged directly opposite the filter membrane assembly 2, which facilitates image acquisition.

[0040] In this embodiment, the sample application port 6 is a puncture sample application port 6. The puncture sample application port 6 is normally in a closed state. Compared to a traditional open sample application port 6, it effectively isolates microorganisms in the outside air and dust, as well as contaminants in the operating environment, from entering the enrichment filter 1, avoiding interference from external contamination with the test results. This solves the problem of false positive results caused by the introduction of contaminants during traditional sample application. During sample application, the sample must be inserted into the enrichment filter 1 through the closed sample application port 6 using a puncture needle, without directly opening the enrichment filter 1 cap. This reduces the opportunity for contact between the operator's hands and tools and the internal environment of the enrichment filter 1, further reducing the risk of contamination from human operation. It is particularly suitable for pharmaceutical microbial testing scenarios where a sterile environment is extremely important.

[0041] The enrichment filter 1 has at least one observation plane on its side, through which the automatic detection module 8 acquires images of the interior of the enrichment filter 1. The curved sidewalls of traditional cylindrical enrichment filters 1 cause light refraction and reflection, resulting in distortions such as stretching and warping in the acquired culture medium images, failing to accurately reflect the state of the culture medium. In this device, the observation plane is designed as a plane, completely eliminating imaging errors caused by the curved surface. This ensures that the images acquired by the automatic detection module 8 are consistent with the actual scene inside the enrichment filter 1, providing accurate and distortion-free image data for subsequent turbidity detection and target recognition.

[0042] In this embodiment, the enrichment filter 1 has a rectangular parallelepiped structure, and all four sides of the enrichment filter 1 are planar.

[0043] An exhaust port 7 is provided at the upper end of the enrichment filter 1, and an air filtration component is installed in the exhaust port 7. When a vacuum device is connected to the external filtration port 5 for filtration, a negative pressure will be formed inside the enrichment filter 1. The exhaust port 7 can introduce outside air through the air filtration component to balance the internal and external pressures, avoiding deformation, seal failure, or abnormal filtration rate of the enrichment filter 1 due to excessive negative pressure. This ensures a stable and efficient filtration process and solves the problem of pressure imbalance during filtration in traditional closed structures. The air filtration component installed in the exhaust port 7 has a microbial interception function, which can effectively filter bacteria, fungi, and other microorganisms in the air entering the enrichment filter 1. This prevents external bacteria from entering the interior of the enrichment filter 1 through the exhaust port 7, solving the contamination risk caused by the lack of filtration function in traditional exhaust structures. This ensures that the culture medium and the intercepted microorganisms grow in a sterile environment, guaranteeing the accuracy of the test results.

[0044] Meanwhile, the exhaust port 7 is equipped with a sealing plug; when the enrichment filter 1 has completed filtration and rinsing and is in the culture stage, the exhaust port 7 can be completely sealed and blocked by the sealing plug, thereby completely isolating the inside of the enrichment filter 1 from the external environment and avoiding interference from external bacteria during the culture process.

[0045] like Figure 6 As shown, this invention discloses a method for detecting microbial enrichment in samples, comprising the following specific steps: Step 1: Open the filtration port 5 and pass the sample into the enrichment filter 1 through the sample feeding port 6. The microorganisms in the sample are enriched in the upper space of the filter membrane by filtration. Step 2: Pass the rinsing solution into the enrichment filter 1 through the sample addition port 6 and rinse it; Step 3: Close the filtration port 5, and introduce culture medium into the enrichment filter 1 through the sample feeding port 6 for incubation; during the incubation process, the automatic detection module 8 periodically acquires images of the enrichment filter 1 and automatically detects the turbidity of the culture medium and the distribution of spherical or clump-shaped targets on the filter membrane assembly 2; and rates the degree of microbial contamination based on the turbidity of the culture medium and the distribution and size of the spherical or clump-shaped targets.

[0046] In this invention, during microbial enrichment detection, the vacuum assistance of the filtration port 5, combined with the tangential flow design of the tilted filter membrane assembly 2, allows the sample to rapidly pass through the filter membrane 10 under negative pressure. Microorganisms are efficiently retained in the upper space of the filter membrane, significantly shortening the enrichment time and improving detection efficiency compared to natural filtration. The tilted filter membrane assembly 2 effectively prevents impurities in the sample (such as food residue and water sediment) from clogging the filter membrane, ensuring that microorganisms can stably adhere to the filter membrane surface, reducing microbial loss due to filter membrane clogging, and guaranteeing the accuracy and stability of enrichment. It is especially suitable for complex samples containing many impurities.

[0047] In step two, rinsing solution is introduced through the sample application port 6 to rinse the sample. This effectively removes residual preservatives (such as preservatives in food), disinfectants (such as chlorides in water), and small impurities that have not been retained, thus avoiding interference from these substances with subsequent culture and testing results and reducing the risk of false negatives. The rinsing process is carried out with the assistance of vacuum filtration. The rinsing solution flows downstream through the filter membrane 10 and is discharged from the vacuum filtration port 5, preventing the erosion and loss of microorganisms retained on the filter membrane. This ensures both the rinsing effect and the amount of microorganisms for subsequent culture, thereby ensuring the reliability of the test results.

[0048] In step three, after closing the filtration port 5, the puncture needle enters the enrichment filter 1 through the puncture and sample addition port 6, and then punctures the filter membrane assembly 2 to inject culture medium into the enrichment filter 1. This ensures that the lower structure 4 of the enrichment filter can be filled with culture medium.

[0049] The enrichment filter 1 forms a sealed culture space, which, together with the air filtration assembly of the exhaust port 7, provides a sterile and stable environment for microbial growth, ensuring normal growth and reproduction of microorganisms. This makes the changes in turbidity of the culture medium and the formation of target substances more regular, providing a good foundation for subsequent detection. During the culture process, the automatic detection module 8 periodically acquires images of the enrichment filter 1 without manual intervention or time constraints, achieving continuous dynamic monitoring of the culture process. Compared with traditional manual periodic observation, it can quickly capture subtle changes in the culture medium, improving the timeliness of handling abnormal results.

[0050] Simultaneously, the automatic detection module 8 detects both the turbidity of the culture medium and the distribution and size of spherical and clump-like targets on the filter membrane assembly 2. This dual-indicator approach comprehensively assesses microbial growth, avoiding the limitations of single-indicator detection. Combined with artificial intelligence algorithms, the detection data is analyzed to quantify and output a microbial contamination rating. The results are objective, accurate, and highly repeatable, solving the problems of reliance on experience and large errors in manual interpretation.

[0051] Specifically, in step three, the automatic detection module 8 uses the following detection method: During initial culture, the automatic detection module 8 acquires an initial state image, crops out the culture medium portion of the initial state image, scales it to a first preset image size, and uses this image as the culture medium background image; During the culture process, the automatic detection module 8 acquires an image at fixed intervals, crops out the culture medium portion and the filter membrane portion of the current image respectively, scales the culture medium portion image to a first preset image size, and uses it as the current frame image of the culture medium; The current frame image of the culture medium and the culture medium background image are input into the first vision algorithm module, and the cropped filter membrane portion image is scaled to a second fixed size and used as the current frame image of the filter membrane, which is then input into the second vision algorithm module; The first vision algorithm module predicts the confidence level of the culture medium turbidity, and the second vision algorithm module predicts the position coordinates and size of all spherical or clump-shaped targets on the filter membrane assembly 2. The turbidity confidence score output by the first vision algorithm module and the number of bounding boxes and the total area of ​​bounding boxes of spherical or clump-shaped targets output by the second vision algorithm module are input into the decision tree model to rate its microbial contamination level.

[0052] The image acquired by the automatic detection module is an RGB image. In this application, the first preset image size is 32. 32 pixels, the second preset image size is 256. 256 pixels.

[0053] Before detection, the culture medium and filter membrane portions of the image were cropped to remove interference from irrelevant areas such as the enriched filter wall and background environment. This allowed the algorithm to focus solely on the core detection object, avoiding the influence of irrelevant information on the detection results. Furthermore, the initial culture medium image was used as the background image, providing a stable comparison benchmark for subsequent turbidity change detection. This enabled the algorithm to accurately capture subtle turbidity differences in the culture medium during the cultivation process, improving the sensitivity of turbidity detection.

[0054] The first vision algorithm module can quantify the turbidity confidence level by comparing the current frame image of the culture medium with the background image, transforming the qualitative judgment of traditional manual observation into quantitative data, accurately analyzing the turbidity of the culture medium, avoiding subjective misjudgment of turbidity changes by humans, and is especially suitable for the identification of subtle turbidity changes in the culture medium.

[0055] The second vision algorithm module accurately locates the position coordinates and size of spherical or clump-shaped targets (such as microbial colonies) in the filter membrane image. It quantifies the target distribution by the number of bounding boxes and the total area, solving the problem that it is difficult for humans to distinguish between food residues and colonies, and impurities and microorganisms, thus improving the specificity and accuracy of target recognition.

[0056] During the decision-making stage, by integrating three core data categories—turbidity confidence level, number of bounding boxes, and total area of ​​bounding boxes—the decision tree model classifies microbial contamination into multi-level ratings, refining the microbial contamination levels and replacing the traditional simple qualitative judgment of "qualified or unqualified." This makes the test results more hierarchical and valuable, facilitating rapid assessment of contamination severity by laboratory personnel. The decision tree model, based on multi-dimensional data analysis, avoids the limitations of single data indicators and reduces the risk of misjudgment due to accidental factors (such as slight local turbidity changes or interference from individual impurities), making the rating results more scientific and reliable. Furthermore, by applying the decision tree model, it can be trained to adapt to the testing needs of different sample types such as pharmaceuticals, food, and water quality. The rating logic can be adjusted according to the microbial growth characteristics and impurity distribution patterns of different samples, enhancing the versatility and adaptability of the detection method.

[0057] Specifically, the decision tree model is a CART classification tree with 3-dimensional input and 11-dimensional output. It uses the Gini coefficient to determine the priority of features, calculates the relative area ratio based on the total area of ​​the bounding boxes, and assigns corresponding weights to the number of bounding boxes, turbidity confidence score, and relative area ratio. The weight of the number of bounding boxes is greater than that of the turbidity confidence score, which in turn is greater than that of the relative area ratio. The contamination index is obtained by summing the weights of the number of bounding boxes, turbidity confidence score, and relative area ratio, and the current microbial contamination level is determined based on the contamination index.

[0058] This invention utilizes a Gini coefficient selection feature, as shown in the following equation: ; In the formula, gini represents the Gini coefficient, P K This represents the frequency of the k-th category.

[0059] The decision in this invention is as follows: the weight of the number of enclosing boxes is greater than the weight of the turbidity confidence level, and the weight of the turbidity confidence level is greater than the weight of the relative area percentage. The logic behind this decision is as follows: the number of colonies (represented by spherical or clump-like targets on the filter membrane assembly) directly reflects the scale of microbial reproduction and is the core indicator of contamination, therefore this item is set with the highest weight; while the turbidity confidence level reflects the overall contamination trend of the culture medium and may be affected by non-microbial factors (such as culture medium precipitation), so it is relatively less important; and the relative area percentage is used as a supplementary indicator. By assigning corresponding weight coefficients to each indicator (number of enclosing boxes, turbidity confidence level, and relative area percentage), the contamination index is obtained by summing the weights of each indicator.

[0060] This invention classifies microbial contamination levels into 0-10. The higher the calculated contamination index, the higher the corresponding contamination level and the more severe the microbial contamination.

[0061] The first-visualization algorithm module uses a regression network model whose output values ​​are between 0 and 1, with the turbidity confidence output also between 0 and 1. During training, the mean squared error loss function is used, as shown below: ; In the formula, loss represents the error loss, and N is the number of samples. It is the actual value of the i-th sample. It is the predicted value of the i-th sample.

[0062] The regression network model uses ResNet18 as the backbone, whose residual connection structure effectively avoids the gradient vanishing problem in deep network training, ensuring the effectiveness of feature extraction. The culture medium background image and the current frame image are used as dual inputs and simultaneously fed into the ResNet18 network. The network processes these features step-by-step through convolutional and pooling layers to extract texture, brightness, and color distribution features from both. A fusion layer then performs difference operations on the two sets of features to highlight the feature differences caused by turbidity changes (such as reduced brightness and coarser texture). The fused difference features are then dimensionality-reduced through a fully connected layer to retain core features strongly correlated with turbidity changes. Finally, a BatchNorm layer standardizes the features, improving the model's adaptability to different lighting conditions and sample background colors.

[0063] During model training, a dataset of culture medium images containing different turbidity levels was constructed, with each sample labeled with its corresponding actual turbidity value (measured using a professional turbidimeter, in NTU). The actual turbidity values ​​were normalized to the 0-1 range and used as labels for model training (e.g., 0 NTU corresponds to label 0, representing no turbidity; 50 NTU corresponds to label 1, representing severe turbidity). The mean squared error loss function was used during training to measure the difference between the model's predicted values ​​and the actual labels; the optimizer could be Adam, with a learning rate set to 0.001, and the loss value was minimized through iterative training.

[0064] The trained regression network model performs inference operations on the fused features of the input background image and the current frame image, outputting a continuous value between 0 and 1, which represents the turbidity confidence score of the culture medium. The confidence score is positively correlated with the actual turbidity. For example, 0 represents the same turbidity as the initial background image (i.e., no turbidity caused by microbial growth), and 1 represents reaching the maximum turbidity level set during model training (severe contamination). Intermediate values ​​correspond to gradient turbidity states; the higher the turbidity confidence score, the more severe the microbial contamination.

[0065] The detection framework of the second vision algorithm module uses a single-stage multi-box detector. During training, the loss function for the positive and negative sample classification of the bounding box uses the focal loss function, as shown in the following formula:

[0066] Where loss represents the error loss for classifying a positive or negative sample, p represents the predicted value, y represents the actual value, and γ is an adjustable parameter greater than 0. The loss function used in the bounding box coordinate regression is the smoothed L1 loss function, as shown below: ; Where loss represents the error loss for the i-th sample. It is the actual value of the i-th sample. It is the predicted value of the i-th sample. Parameters used to control the smoothing range.

[0067] The bounding box positive and negative sample classification part is a binary classification task in the single-stage multi-box detector (SSD framework) of the convolutional neural network, which determines whether each preset anchor box corresponds to a real colony. The core is to distinguish whether the anchor box belongs to a positive sample (covering a real colony) or a negative sample (not covering a real colony). The focal loss function is used to optimize the loss calculation of this classification task, which can solve the problem of extreme imbalance between positive and negative samples in the microbial detection scenario.

[0068] In target detection of filter membrane components, the background region (no target region) is much larger than the target region, resulting in an extreme imbalance between positive and negative samples. Traditional cross-entropy loss functions tend to dominate the training process due to the excessive number of negative samples, causing the model to favor predicting negative samples and reducing target detection sensitivity. Focal loss functions assign very low weights to easily distinguishable negative samples and high weights to difficult-to-distinguish samples, effectively balancing the training contributions of positive and negative samples and preventing the model from overfitting to the background region.

[0069] In actual detection, the initial microbial colonies (small targets) or low-contrast targets (small color difference from the culture medium) on the filter membrane assembly are difficult to distinguish as positive samples, and traditional loss functions are difficult to train effectively on them. The focal loss function amplifies the loss value of such samples by modulating the coefficients, so that the model focuses on learning the features of small targets and weak features during training, which significantly improves the recognition rate of such targets and ensures that a small number of colonies in the early stage of microbial proliferation can be captured in time.

[0070] The automatic detection module 8 combines machine vision and artificial intelligence algorithms to automate the interpretation of the aseptic culture process. This process not only meets the data integrity requirements of recording image changes during the culture process and electronically outputting experimental reports, but also provides accurate and reliable algorithmic interpretation results in real time, unrestricted by manual work hours. In case of abnormal results, laboratory and production personnel can be promptly notified via multimedia communication to quickly initiate deviation investigations. This method and technology enable remote monitoring.

[0071] This invention is not limited to the preferred embodiments described above. Anyone can derive other products in various forms under the guidance of this invention. However, regardless of any changes in shape or structure, any technical solution that is the same as or similar to this application falls within the protection scope of this invention.

Claims

1. A device for detecting microbial enrichment in samples, characterized in that, The system includes an enrichment filter and an automatic detection module. The enrichment filter contains a filter membrane assembly arranged at an angle, which divides the interior of the enrichment filter into an upper filter membrane space and a lower filter membrane space. A sample inlet is located at the top of the enrichment filter, and a filtration port is located at the bottom. The automatic detection module acquires images of the interior of the enrichment filter and automatically detects the turbidity of the culture medium and the distribution and size of spherical or clump-like targets. Based on the turbidity of the culture medium and the distribution and size of the spherical or clump-like targets on the filter membrane assembly, the module rates the degree of microbial contamination.

2. The microbial enrichment detection device in a sample according to claim 1, characterized in that, The angle of inclination between the filter membrane assembly and the horizontal plane is 45°-75°, and the filter membrane assembly is composed of a filter membrane and a supporting frame.

3. The microbial enrichment detection device in a sample according to claim 1, characterized in that, The sampling port is a puncture sampling port.

4. The microbial enrichment detection device in a sample according to claim 1, characterized in that, The enrichment filter has at least one observation plane on its side, through which the automatic detection module obtains an image of the inside of the enrichment filter.

5. The microbial enrichment detection device in a sample according to claim 1, characterized in that, The enrichment filter has an openable and closable exhaust port at its upper end, and an air filtration component is installed in the exhaust port.

6. A method for detecting microbial enrichment in a sample, based on the microbial enrichment detection device according to any one of claims 1-5, characterized in that, The specific steps include the following: Step 1: Open the filtration port and pass the sample into the enrichment filter through the sample feeding port. The microorganisms in the sample are enriched in the upper space of the filter membrane by filtration. Step 2: Pass the rinsing solution into the enrichment filter through the sample addition port and rinse it; Step 3: Close the filtration port, introduce culture medium into the enrichment filter through the sample feeding port, and incubate. During the incubation process, the automatic detection module periodically acquires images of the enrichment filter and automatically detects the turbidity of the culture medium and the distribution of spherical or clump-shaped targets on the filter membrane assembly. Based on the turbidity of the culture medium and the distribution and size of the spherical or clump-shaped targets, the microbial contamination level is rated.

7. The method for detecting microbial enrichment in a sample according to claim 6, characterized in that, In step three, the specific detection method of the automatic detection module is as follows: During the initial culture, the automatic detection module acquires the initial state image, crops out the culture medium portion of the initial state image, scales it to the first preset image size, and uses this image as the culture medium background image; During the cultivation process, a cultivation image is acquired, and the culture medium portion image and the filter membrane portion image are cropped from the current image respectively. The culture medium portion image is scaled to a first preset image size and used as the current frame image of the culture medium. The current frame image of the culture medium and the culture medium background image are input into the first vision algorithm module. Then, the cropped filter membrane portion image is scaled to a second fixed size and used as the current frame image of the filter membrane, which is then input into the second vision algorithm module. The first vision algorithm module predicts the confidence level of the culture medium turbidity, and the second vision algorithm module predicts the position coordinates and size of all spherical or clump-shaped targets on the filter membrane assembly. The turbidity confidence score output by the first vision algorithm module and the number of bounding boxes and the total area of ​​bounding boxes of spherical or clump-shaped targets output by the second vision algorithm module are input into the decision tree model to rate its microbial contamination level.

8. The method for detecting microbial enrichment in a sample according to claim 7, characterized in that, The decision tree model is a CART classification tree with 3-dimensional input and 11-dimensional output. It uses the Gini coefficient to determine the priority of features, calculates the relative area ratio based on the total area of ​​the bounding boxes, and assigns corresponding weights to the number of bounding boxes, turbidity confidence, and relative area ratio. The contamination index is obtained by adding the weights of the number of bounding boxes, turbidity confidence, and relative area ratio, and the current microbial contamination level is determined based on the contamination index.

9. The method for detecting microbial enrichment in a sample according to claim 8, characterized in that, The first vision algorithm module uses a regression network model whose output value is between 0 and 1. During training, the loss function used is the mean squared error loss, as shown in the following formula: ； In the formula, loss represents the error loss, and N is the number of samples. It is the actual value of the i-th sample. It is the predicted value of the i-th sample.

10. The method for detecting microbial enrichment in a sample according to claim 8, characterized in that, The detection framework of the second vision algorithm module uses a single-stage multi-box detector. During training, the loss function for the positive and negative sample classification of the bounding box uses the focal loss function, as shown in the following formula: Where loss represents the error loss for classifying a positive or negative sample, p represents the predicted value, y represents the actual value, and γ is an adjustable parameter greater than 0. The loss function used in the bounding box coordinate regression is the smoothed L1 loss function, as shown below: ； Where loss represents the error loss for the i-th sample. It is the actual value of the i-th sample. It is the predicted value of the i-th sample. Parameters used to control the smoothing range.

Images