Gentamicin C1a titer determination method and device based on computer vision and storage medium

Abstract

The invention relates to a gentamicin C1a titer determination method and device based on computer vision and a storage medium. According to the method, after original images of fermentation liquor samples under different culture conditions are obtained, corresponding color features are extracted based on the original images; then the color features are preprocessed; and finally, on the basis of the preprocessed color characteristics, utilizing the calibrated gentamicin C1a concentration determination model to predict the titer of gentamicin C1a under different culture conditions. Compared with the prior art, the method has the advantages that the titer of gentamicin C1a in the fermentation process of different culture systems of micromonospora aculeata can be quickly, efficiently, accurately and noninvasively measured, and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of biological process detection, and in particular to a computer vision-based gentamicin C1a potency determination method, device and storage medium. Background Art

[0002] Gentamicin is a multi-component aminoglycoside antibiotic produced by microorganisms such as Micromonospora purpura and Micromonospora echinospora. It has been widely used in clinical practice due to its low price, broad antibacterial spectrum, strong antibacterial effect and stable efficacy. The most important component is gentamicin C1a, which can be used as an industrial synthetic precursor for the key drug azithromycin. Since there is no conjugated double bond (no ultraviolet absorption peak) in the molecular structure of gentamicin C1a, it must be pre-column derivatized before its potency can be detected by high-performance liquid chromatography (HPLC). That is, for the detection of gentamicin and single components, it is still necessary to perform HPLC precise quantification on the derivatized samples. However, this method must terminate the fermentation of the strain in the culture system, and the complex pretreatment process also affects its accuracy by the proficiency of the operator such as acidification and range matching.

[0003] In addition, there are studies that combine computer vision technology with machine learning and apply it to the cultivation of plant cells. Among them, computer vision technology has the advantages of being fast, non-destructive, consistent, objective, reliable and cost-effective, so it has become the main method for monitoring quality in the food industry. The target image information is acquired by the imaging system and then transmitted to the image processing system, which converts the image (color, texture, brightness, pixel distribution, etc.) into digital information, which can be further calculated, processed and analyzed to achieve the recognition, detection and control of the target object. This technology has also broadened its application areas due to its stability, flexibility, accuracy and other characteristics. In tea products or processing, computer vision combined with chemometrics has been used to determine the grade of black tea, predict catechins, theaflavins and chlorophyll in the fermentation process of black tea, and detect the moisture content in the process of tea shaping. In the field of microbial cultivation, computer vision combined with machine learning has also been used to determine plant cell morphology sorting and product detection. In addition, deep learning, as a specialized branch of machine learning, performs well in capturing complex data features and is widely used in various fields. Among them, one-dimensional convolutional neural network (1D-CNN) has gradually been incorporated into the deep learning framework for analyzing continuous data. For example, 1D-CNN has shown excellent results in measuring pH using continuous data. However, existing methods can only achieve endpoint detection (1D) and are difficult to achieve process monitoring (2D) of different culture systems.

[0004] Therefore, how to improve the efficiency and accuracy of gentamicin C1a titer determination has become a problem that needs to be solved in this field. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of low efficiency and low accuracy in the above-mentioned prior art and to provide a method, device and storage medium for determining the potency of gentamicin C1a based on computer vision.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a computer vision-based method for determining the potency of gentamicin C1a is provided, comprising the following steps: obtaining original images of fermentation broth samples under different culture conditions; extracting corresponding color features based on the original images; preprocessing the color features; and predicting the potency of gentamicin C1a under different culture conditions using a calibrated gentamicin C1a concentration determination model based on the preprocessed color features.

[0008] As a preferred technical solution, the construction, calibration and prediction process of the gentamicin C1a concentration determination model specifically includes: randomly dividing the fermentation broth samples under different culture conditions into a calibration set and a prediction set; based on the calibration set, using a preset basic machine learning algorithm or a deep learning algorithm to construct an initial gentamicin C1a concentration determination model; using the data in the calibration set to train and optimize the initial gentamicin C1a concentration determination model, and selecting a model with the best preset key indicators as the calibrated gentamicin C1a concentration determination model; using the prediction set to test and evaluate the calibrated gentamicin C1a concentration determination model to obtain the corresponding gentamicin C1a content prediction result.

[0009] As a preferred technical solution, the basic machine learning algorithm is one of the least squares method, random forest, and back-propagation artificial neural network algorithms.

[0010] As a preferred technical solution, when the deep learning algorithm adopts a one-dimensional convolutional neural network, the initial gentamicin C1a concentration determination model includes at least an input layer, two convolutional layers, two batch normalization layers, five activation function layers, two maximum pooling layers, a flattening layer, a discard layer, three fully connected layers and an output layer connected in sequence.

[0011] As a preferred technical solution, the activation function layer adopts a rectified linear unit function.

[0012] As a preferred technical solution, the different culture conditions include at least two of the following culture conditions: synthetic culture medium, compound culture medium, microplate culture, shake flask culture and bioreactor culture.

[0013] As a preferred technical solution, corresponding color features are extracted, and the specific process includes: based on the original image, selecting a target area; extracting different color components from the target area, and calculating the average value of different color components; wherein the color components include one or more of the following: red, green, blue, brightness, relative red, relative green, relative blue, hue, saturation, brightness in Lab mode, cyan, magenta, yellow and black.

[0014] As a preferred technical solution, the preprocessing method is MinMax scaling or Z-score normalization processing.

[0015] According to a second aspect of the present invention, there is provided a computer vision-based gentamicin C1a potency determination device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the method described above when executing the program.

[0016] According to a third aspect of the present invention, there is provided a storage medium having a program stored thereon, wherein the program implements the method described above when executed.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention combines multiple pretreatment methods with multiple modeling methods to achieve comprehensive monitoring of key parameters in different culture systems, which can not only enhance the versatility of the gentamicin C1a concentration determination model, but also improve the detection accuracy and application diversity, and can achieve rapid strain screening and accurate prediction of the titer of target products in multiple culture systems and the whole process, and quickly, efficiently, accurately and non-invasively perform gentamicin C1a titer determination in the fermentation process of Micromonospora aculeatus in different culture systems;

[0019] 2. The present invention constructs and calibrates a gentamicin C1a concentration determination model by combining computer vision technology with a deep learning algorithm. On the one hand, compared with traditional fermentation process detection, which usually requires multiple independent sensors to monitor different parameters, the present invention uses computer vision technology to simultaneously capture multiple parameter information through images, which can not only improve the detection efficiency, but also provide a more comprehensive fermentation process status by comprehensively analyzing the relationship between different parameters; on the other hand, the one-dimensional convolutional neural network used in the present invention is a subset of the deep learning framework, which is good at analyzing continuous data, and uses computer vision to extract the color features of the image as sequence data, and combines it with the one-dimensional convolutional neural network. Combined with neural networks, a prediction model is established for the target product in the biomanufacturing process, which can transition from single endpoint detection (1D) to comprehensive process monitoring (2D). Compared with the traditional rapid detection scheme, the method provided by the present invention can perform non-invasive detection, that is, there is no need to directly contact the fermentation system, and will not interfere with the fermentation process, so as to achieve rapid, efficient and non-invasive gentamicin C1a potency determination, reduce the risk of contamination in the fermentation process that needs to maintain a highly sterile environment, improve the purity and quality of the fermentation product, and effectively improve the detection accuracy. In the experiment, compared with the existing method, the RMSEP of the model provided by the present invention can be reduced by 58.76%;

[0020] 3. The different culture conditions of the present invention include synthetic culture medium, composite culture medium, microplate culture, shake flask culture and bioreactor culture, which provide a diverse data source and effectively improve the applicability of the gentamicin C1a concentration determination model. The adaptability of the model to various culture conditions can further ensure the accuracy and efficiency of the screening process. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic flow chart of the method provided by the present invention;

[0022] Figure 2 It is a heat map of the correlation coefficient between the potency of gentamicin C1a and the color variable in the embodiment of the present invention;

[0023] Figure 3 It is a scatter plot of PLS, RF and BP-ANN models using different preprocessing methods in the embodiment of the present invention;

[0024] Figure 4 The MSE loss trend and prediction result distribution diagram of the 1D-CNN model using different preprocessing methods in the embodiment of the present invention;

[0025] Figure 5 It is a fitting time bar chart of PLS, RF, BP-ANN and 1D-CNN models using different preprocessing methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0026] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0027] Example

[0028] like Figure 1 As shown, this embodiment provides a computer vision-based gentamicin C1a titer determination method, which first obtains the original image of the fermentation broth sample under different culture conditions; secondly, based on the original image, extracts the corresponding color features; thirdly, preprocesses the color features; finally, based on the preprocessed color features, uses a calibrated gentamicin C1a concentration determination model to predict the gentamicin C1a titer under different culture conditions. The specific implementation process is as follows:

[0029] Step S1, obtaining original images of fermentation broth samples under different culture conditions.

[0030] The different culture conditions in this step include at least two of the following culture conditions: synthetic culture medium, composite culture medium, microplate culture, shake flask culture and bioreactor culture. In this embodiment, the pre-constructed Micromonospora echinospora 49-92S-KL01 is used as the parent strain for producing gentamicin C1a. For the detailed construction process, see Publication 1 (Xu, F., Zhang, X., Liu, L., Ke, X., Wu, J., Guo, Y., Tian, ​​X., & Chu, J. (2022). Engineering the methyltransferase through inactivation of the genK and genL leads to a significant increase of gentamicin C1a production in an industrial strain of Micromonospora echinospora 49-92S. Bioprocess and biosystems engineering, 45 (10), 1693–1703). On this basis, the specific contents of each culture medium and culture conditions are as follows:

[0031] Basic solid medium (g / L): soluble starch, 10; asparagine, 0.02; CaCO3, 1; MgSO4-7H2O, 0.5; KH2PO4, 0.3; NaCl, 0.5; KNO3, 1; agar, 13; bran, 13. The solid plate was cultured at 35°C for 8-10 days.

[0032] Basic seed culture medium (g / L): soluble starch, 10; glucose, 1; corn flour, 15; soybean flour, 10; peptone, 2; KNO3, 0.5; CaCO3, 5.

[0033] Fermentation medium (g / L): soluble starch, 30; glucose, 5; corn flour, 25; soybean flour, 26; peptone, 10; KNO3, 0.5; CaCO3, 7; (NH4)2SO4, 1; CoCl, 0.03.

[0034] Fermentation synthetic medium (g / L): glucose, 20, ammonium sulfate, 3, sodium nitrate, 2, potassium chloride, 0.5, potassium dihydrogen phosphate, 1.5, ferrous sulfate heptahydrate, 0.05, magnesium sulfate, 0.05, cobalt chloride, 0.02, calcium carbonate, 2.

[0035] The seeds were cultured in shake flasks at 220 rpm and 35.5°C for 48 h. Shake flask fermentation was also carried out at the same speed and temperature for 5 days with an inoculum size of 12%.

[0036] For the synthetic culture system, the seed culture was washed three times with saline and then inoculated into the fermentation shake flask. The fermentation medium in the 5L bioreactor was the same as above (Shanghai Guoqiang Bioengineering Equipment Co., Ltd.). All media were sterilized at 115°C for 30 minutes. The initial working volume was 3L, the inoculum size was 12%, and the culture was carried out at 35.5°C for 120h. The ventilation was 0.5VVM, and the dissolved oxygen (DO) level was higher than 40% of the saturation concentration, which was maintained by gradually adjusting the stirring.

[0037] The specific process of preparing the fermentation liquid sample and acquiring the original image in this embodiment includes:

[0038] A total of 605 samples were collected from two culture systems of Micromonospora aculeatus (complex medium and synthetic medium), including 350 samples and 255 samples under the two conditions, respectively. The image acquisition system used to collect images of Micromonospora aculeatus under different culture systems mainly consists of a camera (Nikon Z-5), a bracket and a uniform light source. Since avoiding reflection effects is very important for the good operation of the system, the camera was installed on the top of a sealed space during shooting, and was configured for automatic exposure and manual white balance control. The resolution of the generated image was 4016×6016 pixels. The distance of the lens was 55 mm, the camera was operated in manual mode, and the image format used was JPG. In addition, the parameters of the camera were as follows: F3.5; shutter speed 0.2 seconds; ISO 200; manual focus. The fermentation broth at different culture stages was evenly pipetted into a 24-well plate, and then the 24-well plate was placed in the observation field of the camera.

[0039] Step S2: extracting corresponding color features based on the original image.

[0040] Specifically, Python+OpenCV was used to process the original images of fermentation broth under different culture systems to obtain a piece of digital information. Among them, the most critical steps are to effectively locate and cut the region of interest (ROI) in the 24-well plate to extract the color components and calculate the average value of each variable. Color features can be directly mapped to individual pixels, and they can be effectively used as inputs for machine learning models, thereby facilitating the rapid quantification of gentamicin C1a therein. In order to fully present the inherent color features in the fermentation broth images, different color spaces were analyzed: taking the quadrant of 500×500 pixels in the new image as the unit, the average values ​​of red (R), green (G), blue (B), brightness (L, i.e., the average value of R, G, and B), three relative colors (relative red (rr=R / L), relative green (rg=G / L), relative blue (rb=B / L)), hue (H), saturation (S), l (brightness in Lab mode), cyan (c), magenta (m), yellow (y), and black (k) were calculated, a total of 14 color components. For the feature extraction and calculation methods, please refer to public document 2 (Zhu, X., Du, C., Mohsin, A., Yin, Q., Xu, F., Liu, Z., Wang, Z., Zhuang, Y., Chu, J., Guo, M., & Tian, ​​X. (2022). An Efficient High-Throughput Screening of High Gentamicin-Producing Mutants Based on Titer Determination Using an Integrated Computer-Aided Vision Technology and Machine Learning. Analytical chemistry, 94 (33), 11659–11669).

[0041] This embodiment uses a method (see Publication 1) for detecting the potency of gentamicin using liquid chromatography ultraviolet detection (Agilent 1200, Agilent Technologies, Inc, USA) as a reference. In this method, 1 mL of gentamicin fermentation broth is acidified to pH 2.0 with 6M concentrated sulfuric acid, allowed to stand for 30 minutes, and then the sample is centrifuged at 12000 rpm for 5 minutes. 100 μL of standard and sample solution is taken, 900 μL of ultrapure water, 400 μL of derivatization reagent o-phthalaldehyde, and 1.1 mL of methanol are mixed evenly, heated in a 60 ° C water bath for 20 minutes, cooled to room temperature, filtered through a 0.22 μm organic phase filter, added to a liquid phase vial, and waited for injection. The chromatographic column used for detection is a ZORBAX Eclipse XDB-C18 chromatographic column (4.6 mm × 250 mm, 5 μm, Agilent Technologies, Inc, USA). The preparation ratio of the mobile phase was methanol: sodium heptane sulfonate: glacial acetic acid = 700:250:50, the detection wavelength was 330 nm, the injection volume was 20 μL, the flow rate was set to 1.0 mL / min, and the column temperature was 30°C.

[0042] Steps S3-S4, preprocessing the color features, and based on the preprocessed color features, using a calibrated gentamicin C1a concentration determination model to predict the gentamicin C1a titer under different culture systems.

[0043] The data preprocessing and partitioning process specifically includes:

[0044] Since the components from different color spaces in the 605 samples may cause the gradient to disappear or explode, thus hindering the convergence of the model, and also leading to an imbalance in feature weights, which has a negative impact on the generalization performance of the model, it is necessary to preprocess the color features extracted from the original image. At present, the preprocessing method commonly used in neural network training is MinMax scaling, because neural networks are very sensitive to the range of input data, and usually need to scale the data to a range between 0 and 1; while Z-score normalization is to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0045] All samples were divided into calibration and prediction sets within an appropriate range to avoid random errors and build a more robust model. For effective evaluation of gentamicin C1a concentration, 605 samples under different culture systems were randomly assigned in a ratio of 8:2, with 484 samples in the calibration set and 121 samples in the prediction set. The calibration set was used to establish the calibration model (i.e., the initial gentamicin C1a concentration determination model), and the prediction set was used to evaluate the predictive ability of the calibration model. Table 1 provides a comprehensive overview of the reference values ​​of the analyzed samples of Micromonospora aculeatus, including the range, mean, and standard deviation. The reasonable distribution of samples shows that the prediction set can be used to evaluate the performance of the model.

[0046] Table 1 Detailed information on the reference values ​​of gentamicin C1a titer and fluorescence intensity in the calibration and prediction sets

[0047]

[0048] Next, the construction, calibration and prediction of gentamicin C1a concentration determination model are introduced respectively for basic machine learning algorithm and deep learning algorithm.

[0049] (1) Using the preset basic machine learning algorithm to construct the initial gentamicin C1a concentration determination model

[0050] The average values ​​of 14 color components were used as the raw data input of the model to predict the gentamicin C1a content. Traditional machine learning models were developed using partial least squares (PLS), random forest (RF), and back propagation artificial neural network (BP-ANN) algorithms. The performance of the best model was evaluated for all samples in the calibration set through backward evaluation, and independent samples in the prediction set were tested. Among them, PLS is a widely used mathematical modeling method that can establish the relationship between features and targets, so as to achieve the purpose of predicting target information. RF regression combines the advantages of multiple decision tree algorithm predictions and is very efficient on large data sets.

[0051] BP-ANN is the most widely used and most powerful artificial neural network. The color values ​​under different culture systems are used as the input of the BP-ANN model, and the gentamicin C1a concentration value is used as the output. The parameter configuration of the network model is as follows: the transfer function from the input layer to the hidden layer and the transfer function from the hidden layer to the output layer both use the rectified linear unit (ReLU) function. When training with the calibration set, the Adam optimizer is used, the learning rate is set to 0.001, the number of training iterations is set to 10000, and the number of neurons in the optimized hidden layer is adjusted to 800. The prediction set and the optimized BP-ANN model (that is, the calibrated gentamicin C1a concentration determination model) are used to predict the gentamicin C1a content.

[0052] (2) Using the preset deep learning algorithm to construct the initial gentamicin C1a concentration determination model

[0053] This embodiment predicts the gentamicin C1a content in different culture systems of Micromonospora aculeatus by constructing a one-dimensional convolutional neural network (1D-CNN) structure. The 1D-CNN model includes an input layer, two convolutional layers, two batch normalization layers, five activation function layers, two maximum pooling layers, a flattening layer, a discard layer, three fully connected layers and an output layer. The convolutional layer is responsible for extracting feature data, with a kernel size of 3×1 and a stride of 1; the maximum pooling layer has a size of 2×1 and a stride of 2, which is used to select and filter features. A ReLU activation layer is introduced between the batch normalization layer and the maximum pooling layer to facilitate the conversion of output features. In this embodiment, a discard layer with a discard rate of 0.2 is added after the flattening layer to randomly discard nodes to prevent overfitting. Finally, the output layer consists of a fully connected dense layer with a single node for predicting the gentamicin C1a content value.

[0054] This example uses the PyTorch framework and the GTR 1066Ti GPU to train 1D-CNN on the jupyter-notebook platform. In the training phase of the 1D-CNN model, the mean square error (MSE) is used as the loss function. The change of MSE loss in continuous time can be used to evaluate the convergence of the model training process. The formula for calculating MSEloss is as follows:

[0055]

[0056] Where n is the number of samples of fermentation broth under different culture systems in the calibration set; and i are the predicted gentamicin C1a and reference gentamicin C1a values ​​of the ith sample, respectively. The MSE loss is minimized by the gradient descent algorithm provided by the adaptive moment estimation (Adam) optimizer, and L2 regularization is used to prevent the model from overfitting the training data and enhance its generalization ability.

[0057] (3) Model evaluation

[0058] The model performance is evaluated using the root mean square error (RMSE) of calibration (RMSEC) and prediction (RMSEP), as well as the calibration coefficient (Rc), prediction coefficient (Rp) and residual prediction deviation (RPD), as shown in equations (2) to (4). 2 A key metric in determining the predictive accuracy of a regression model, RMSE is a measure of the residual error, indicating how tightly the data fit around the line of best fit.

[0059]

[0060]

[0061] In summary, the method provided in this embodiment combines computer vision technology with machine learning (including basic machine learning and deep learning) algorithms, explores the relationship between the color of the fermentation broth and gentamicin C1a during the cultivation of Micromonospora aculeatus through computer vision, and uses multiple machine learning algorithms to establish a prediction model of gentamicin C1a in fermentation samples under different culture systems, thereby realizing rapid and non-invasive detection of the gentamicin C1a titer in the fermentation broth under two culture systems of Micromonospora aculeatus mutants. This method uses the color change during the fermentation of Micromonospora aculeatus to predict the titer of gentamicin C1a, which is a fast, safe, non-contact, and low-cost detection method.

[0062] Next, we compare and analyze the experimental results obtained according to the above method:

[0063] 1. Correlation between gentamicin C1a potency and image features

[0064] Figure 2 Pearson correlations between gentamicin C1a content and color components during the fermentation of Micromonospora aculeatus are shown, with red indicating strong positive correlations and blue indicating strong negative correlations. The figure shows clear correlations between different color components. In the RGB color space, the B component is positively correlated with gentamicin C1a titer (0.47), but the R and G components are negatively correlated with gentamicin C1a titer with coefficients of 0.53 and 0.09, respectively. In the HSI color space, the H and I components are positively correlated with gentamicin C1a concentration with coefficients of 0.29 and 0.29, respectively, but the S component is negatively correlated with a coefficient of 0.80. In the other color spaces, the rr and y components have strong negative correlations with gentamicin C1a content with coefficients of 0.73 and 0.77, respectively; but the rb component shows a strong positive correlation with a coefficient of 0.77. Together, these correlations indicate that gentamicin C1a concentration corresponds to the yellow color, relative color, and saturation of the fermentation broth samples. It is noteworthy that among the 14 color variables, S, y, rb, rr, m, R, and c showed strong correlations with gentamicin C1a concentrations, indicating that they made important contributions to the prediction of gentamicin C1a.

[0065] 2. Analysis of the model and comparison of preprocessing conditions

[0066] The inputs of the PLS model, RF model, and BP-ANN model were original data and preprocessed data (Z-score and MinMax), and the models were optimized by 5-fold cross validation. Figure 3The correlation between the predicted values ​​and measured values ​​of the calibration set and prediction set based on different preprocessing methods is described. Figure 3 (a) to (c) are scatter plots of PLS ​​models using original data, data preprocessed by Z-score, and data preprocessed by MinMax, respectively. Figure 3 (d) to (f) are scatter plots of the RF model using the original data, the data preprocessed by Z-score, and the data preprocessed by MinMax, respectively. Figure 3 (g) to (i) are scatter plots of the BP-ANN model using original data, data preprocessed by Z-score, and data preprocessed by MinMax, respectively.

[0067] Table 2 shows the prediction performance of PLS, RF, BP-ANN and 1D-CNN models after different preprocessing. The results show that the prediction performance of the PLS model with raw data as input and the two standardized processing is slightly different (Rc = 0.1663, RMSEC = 82.89 mg / L, Rp = 0.1070, RMSEP = 122.05 mg / L). The RF model with Z-score preprocessed data as input has the best result, Rc = 0.9217, RMSEC = 36.14 mg / L, Rp = 0.8918, RMSEP = 29.85 mg / L. The prediction performance of the BP-ANN model with raw data as input is the worst, Rc = 0.8516, RMSEC = 49.76 mg / L, Rp = 0.7733, RMSEP = 43.22 mg / L. The prediction performance of the BP-ANN model with Z-score preprocessed data as input is second, Rc = 0.9125, RMSEC = 38.21 mg / L, Rp = 0.8112, RMSEP = 39.44 mg / L. The prediction results of the BP-ANN model with MinMax preprocessed data as input are the best, Rc = 0.9359, RMSEC = 32.69 mg / L, Rp = 0.8769, RMSEP = 31.86 mg / L. c is the correlation coefficient of the calibration set; RMSE C is the calibration root mean square error; R p is the correlation coefficient of the prediction set; RMSE P is the root mean square error of prediction.

[0068] Table 2 Prediction parameters of PLS, RF, BP-ANN and 1D-CNN models after different preprocessing

[0069]

[0070] Figure 4(a) to (c) show the MSE loss trend of the 1D-CNN model based on original data, Z-score and MinMax preprocessed data. Figure 4 (d) to (f) show the distribution of prediction results generated by the 1D-CNN model based on original data, Z-score data, and MinMax preprocessed data, as well as their linear fitting values ​​and the actual values ​​of the validation set and prediction set. Figure 4 (a) in Figure 1 shows that when the original data is used as the model input, the MSE loss initially shows a significant decrease, with significant fluctuations during epochs 426-452 and 841-869. Finally, the prediction loss reaches the best convergence state at epoch 958. Figure 4 (b) in Figure 1 depicts the convergence of the MSE loss during model training based on the Z-score preprocessed data input. It initially drops sharply, then fluctuates slightly, and reaches the best convergence state of prediction loss as early as the 87th epoch. Figure 4 (c) in Figure 1 shows the convergence of the MSE loss of the input data preprocessed using MinMax scaling, showing a rapid and fluctuating decline, and the prediction loss reaches the best convergence state at the 618th epoch. Figure 4 As shown in (d) in FIG. 3 , the model trained based on the original data has Rc=0.9223, RMSEC=36.01 mg / L, Rp=0.8601, and RMSEP=33.96 mg / L. Figure 4 As shown in (e) in Figure 3, the model trained using Z-score preprocessed data showed Rc = 0.9952, RMSEC = 8.95 mg / L, Rp = 0.9862, and RMSEP = 10.67 mg / L. Figure 4 As shown in (f) in the figure, the model trained using MinMax scaling preprocessed data has Rc = 0.9568, RMSEC = 26.85 mg / L, Rp = 0.9136, and RMSEP = 26.69 mg / L.

[0071] In addition, while performing model statistics, the computation time required for model training is also calculated, such as Figure 5As shown. The comparison shows that the training time of the neural network model is significantly higher than that of the traditional model. Specifically, the training time of the 1D-CNN model without any preprocessing algorithm is 329.08s, followed by the BP-ANN, RF and PLS models, which require 188.6s, 32.1s and 15.5s respectively. It is worth noting that for the two neural network models, the time required for the training process is significantly reduced after the use of Z-score and Minmax standardization. For the two traditional machine learning models, since the required fitting time is shorter, there is no obvious time change due to the difference in preprocessing methods. For the BP-ANN model, the training process takes the shortest time of 128.9s after using the Minmax normalization method; for the 1D-CNN model, the training process takes the shortest time of 249.3s after using the Z-score normalization method. In addition to the model size, the time also depends on the data set, such as the input features and sample size. An increase in either of the two will lead to an increase in training time. The prediction time of all models is less than 5 seconds. In short, higher R 2 , RPD, and RMSEP values ​​show that 1D-CNN performs better than the BP-ANN model and significantly outperforms the traditional machine learning models (PLS and RF).

[0072] Furthermore, this embodiment also provides a gentamicin C1a potency determination device based on computer vision, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements one or more steps of the aforementioned method. The execution process is basically the same as the implementation steps of the aforementioned method, which will not be repeated here.

[0073] Furthermore, this embodiment also provides a storage medium on which a program is stored. When the program is executed, one or more steps of the aforementioned method are implemented. The execution process is basically the same as the implementation steps of the aforementioned method, and will not be repeated here.

[0074] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for determining the potency of gentamicin C1a based on computer vision, characterized in that: The following steps are involved: Get original images of fermentation broth samples under different culture conditions; Based on the original image, extract the corresponding color features; Preprocess the color features; Based on the color characteristics after pretreatment, the calibrated gentamicin C1a concentration determination model was used to predict the gentamicin C1a titer under different culture conditions.

2. The computer vision-based gentamicin C1a titer determination method according to claim 1, characterized in that: The construction, calibration and prediction process of the gentamicin C1a concentration determination model specifically includes: The fermentation broth samples under different culture conditions were randomly divided into a calibration set and a prediction set; Based on the calibration set, constructing an initial gentamicin C1a concentration determination model using a preset basic machine learning algorithm or a deep learning algorithm; Utilizing the data in the calibration set to train and optimize the initial gentamicin C1a concentration determination model, and selecting a model with the best preset key indicators as a calibrated gentamicin C1a concentration determination model; The prediction set is used to test and evaluate the calibrated gentamicin C1a concentration determination model to obtain the corresponding gentamicin C1a content prediction result.

3. The computer vision-based gentamicin C1a titer determination method according to claim 2, characterized in that: The basic machine learning algorithm is one of the least squares method, random forest, and back propagation artificial neural network algorithms.

4. The computer vision-based gentamicin C1a titer determination method according to claim 2, characterized in that: When the deep learning algorithm adopts a one-dimensional convolutional neural network, the initial gentamicin C1a concentration determination model includes at least an input layer, two convolutional layers, two batch normalization layers, five activation function layers, two maximum pooling layers, a flattening layer, a discard layer, three fully connected layers and an output layer connected in sequence.

5. The computer vision-based gentamicin C1a titer determination method according to claim 4, characterized in that: The activation function layer adopts a rectified linear unit function.

6. The computer vision-based gentamicin C1a titer determination method according to claim 1, characterized in that: The different culture conditions include at least two of the following culture conditions: synthetic culture medium, compound culture medium, microplate culture, shake flask culture and bioreactor culture.

7. The computer vision-based gentamicin C1a titer determination method according to claim 1, characterized in that: Extract the corresponding color features. The specific process includes: Based on the original image, selecting a target area; Extracting different color components from the target area and calculating average values ​​of the different color components; The color components include one or more of the following: red, green, blue, brightness, relative red, relative green, relative blue, hue, saturation, brightness in Lab mode, cyan, magenta, yellow and black.

8. The computer vision-based gentamicin C1a titer determination method according to claim 1, characterized in that: The preprocessing method is MinMax scaling or Z-score normalization processing.

9. A computer vision-based gentamicin C1a titer determination device, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.

Images