Oil production evaluation method based on schizochytrium limacinum sequential morphological analysis

Abstract

The invention discloses an oil production evaluation method based on schizochytrium limacinum sequential morphological analysis. According to the method, a unified and clear rule is established according to the cellular morphology, standardized classification is performed, and the improvement of a target detection algorithm is combined, so that local features in the cells can be captured more accurately, the relationship between the cell type and the grease yield is further deeply explored, and real-time prediction of the grease yield is realized based on a deep learning algorithm. A key tool for connecting morphological characteristics of schizochytrium limacinum and fermentation process optimization is established, and rapid and low-cost fermentation industrial quality control is realized.

Description

Technical Field

[0001] The present invention relates to an oil production evaluation method based on the chronological morphological analysis of Schizochytrium, and belongs to the field of detection technology. Background Art

[0002] Schizochytrium is a marine fungus, usually spherical, and can be produced by heterotrophic reproduction. Schizochytrium is rich in lipids, accounting for 40% - 70% of the dry cell mass, among which the content of docosahexaenoic acid (DHA) is 20% - 50%. DHA is an important n-3 polyunsaturated fatty acid, which has many benefits for human health, such as promoting the development of the brain and nervous system, protecting cardiovascular health, etc. Traditional DHA production mainly relies on fish oil extraction, but there are problems such as limited resources and low purity. The DHA oil produced by Schizochytrium has high purity, is easy to absorb, and more than 90% exists in the form of triglycerides, which is more easily absorbed by the human body. In addition, Schizochytrium has a fast growth rate and simple culture conditions, which are suitable for large-scale industrial production. Therefore, the fermentation of Schizochytrium to produce DHA oil has become a new, safer and healthier production method. In addition, the high oil content of Schizochytrium also has potential application value in the field of biodiesel production. As an efficient oil-producing strain, Schizochytrium has broad application prospects.

[0003] During the growth of Schizochytrium, the specific conditions of the culture medium and the changes in fermentation conditions will significantly affect the production efficiency of oil. A balance between its biomass and oil production needs to be achieved through fine-tuning of nutrient regulation, optimization of environmental conditions, and adjustment of metabolic pathways. However, the regulation during the fermentation production of Schizochytrium mainly relies on the timely feedback of biomass, and the feedback on oil yield is not timely enough to adjust the fermentation conditions according to the oil production situation. At present, the detection methods for the oil yield of Schizochytrium mainly include the following: (1) Chemical extraction and quantitative analysis, where the oil in the cells is extracted by organic solvent extraction and then the oil content is calculated by weighing or other methods. However, this method requires a large amount of organic solvents, has a long operation time, and there are safety risks, etc.; (2) Fluorescent staining method, where Nile red fluorescent dye binds to the oil in the cells and emits fluorescence, and the oil is quantified based on the linear relationship between the fluorescence intensity and the oil content. However, this method requires cell staining, and the fluorescence intensity may be affected by cell density and operation conditions, and the responsiveness to the oil content is not high enough; (3) Gas or liquid chromatography method, which can analyze the oil components, but has a high equipment cost, complex operation, and long detection time, and is not suitable for rapid detection of oil yield in production; (4) Cell dry weight method, where the cell dry weight is weighed and then the oil yield is calculated in combination with the oil content. However, in this method, the oil content is based on biomass, and the relationship between biomass and oil content is not positively correlated, so it cannot be used to evaluate the actual growth state of cells and thus cannot be used to guide the regulation of fermentation. At present, there is a lack of a method that can more intuitively and quickly evaluate the oil content of Schizochytrium, and it is impossible to judge whether the fermentation is in a normal state by observing the growth state of Schizochytrium in real time during the fermentation process, so the fermentation strategy cannot be adjusted in real time. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the present invention establishes a method for classifying cell morphology during the life cycle of Schizochytrium, combines cell morphology classification with object detection algorithms, improves the model according to the characteristics of the Schizochytrium dataset, and establishes a population object detection model for Schizochytrium. The purpose is to more intuitively explore the relationship between the cell morphology and oil yield of Schizochytrium, judge whether the fermentation is in a normal stage by observing the growth state of microorganisms in real time during the fermentation process, and promote the intelligent development of Schizochytrium fermentation.

[0005] The present invention is achieved through the following technical solutions:

[0006] The first object of the present invention is to provide an oil production evaluation method based on the sequential morphology analysis of Schizochytrium, including the following steps:

[0007] S1. According to the different division methods of Schizochytrium cells and the sizes of lipid droplets inside the cells, the cell morphology of Schizochytrium is divided into single vegetative cells, small lipid droplet dyads, large lipid droplet dyads, small lipid droplet triads, large lipid droplet triads, small lipid droplet tetrads, large lipid droplet tetrads, multi-nucleate cells, small lipid droplet single cells, large lipid droplet single cells, large lipid droplet-filled single cells, lipid droplet-free cells, and broken cells;

[0008] S2. Prepare a dataset, annotate the Schizochytrium cell images according to the classification in step S1, and divide the annotated image dataset into a training set and a validation set;

[0009] S3. Construct an MLC-YOLO model based on YOLOv8s, which consists of an input layer, a bottleneck layer, a detection head, a feature fusion neck, and a multi-level feature enhancement module;

[0010] S4. Use the training set prepared in step S2 to train the MLC-YOLO model to obtain the best training model; input the images in the validation set into the best training model obtained by training for validation and then save the best model;

[0011] S5. Use the best model to detect the images to be detected and judge the oil production situation.

[0012] In one embodiment of the present invention, a hybrid local channel attention module is introduced after the Upsample and Concat operations in the feature fusion neck of the MLC-YOLO model and in the C2f module in the deep refinement stage of the network.

[0013] In one embodiment of the present invention, the hybrid local channel attention module first passes the input image through local average pooling to obtain a feature map The pooling size can be adjusted by ks; where F in represents the input feature map, R represents the real number space, C in represents the number of input channels, H represents the height of the feature map, W represents the width of the feature map, and ks represents the pooling kernel size; then it is divided into two parts,

[0014] The first part passes through global average pooling to obtain a global feature map and then obtains a channel representation through one-dimensional convolution Subsequently, it is restored to the size of the original feature map through inverse average pooling UNAP It can be expressed by the following formula:

[0015] F 1out = UNAP(Conv1d(GAP(F1)))(1)

[0016] Among them, Conv1d represents one-dimensional convolution, and GAP represents global average pooling;

[0017] The second part first obtains a new feature map through feature map reconstruction Then, a new channel representation is further obtained through one-dimensional convolution Subsequently, it is reconstructed again into It can be represented by the following formula:

[0018] F 2out = RS(Conv1d(RS(F1))) (2)

[0019] Where RS represents the reconstruction operation;

[0020] Add these two parts of feature maps along the same pixel positions to obtain a fused feature map Restore it to the original feature map dimension through an inverse average pooling operation Then perform a dot product operation with the original input image at the corresponding pixel points to update the original input feature map and obtain the final attention-weighted feature map That is

[0021]

[0022] Where + represents adding corresponding channel elements, represents multiplying corresponding channel elements.

[0023] In an embodiment of the present invention, a multi-scale feature sharing pyramid convolution module is introduced in the last layer of the feature extraction network.

[0024] In an embodiment of the present invention, the multi-scale feature sharing pyramid convolution module is composed of 6 CBSs; for the input of the feature map, it is first processed by CBS to obtain an intermediate feature map F1, and then F1 is input into the dilated convolution layer; in each layer, the output feature map will be fused with the features of the next dilated convolution layer; four output feature maps F 12 、F 123 、F 1234 and F 12345 can be obtained respectively. After splicing these four feature maps, a new feature map is obtained through CBS, and its expression is:

[0025] F out = CBS(F1⊕F 12 ⊕F 123 ⊕F 1234 ⊕F 12345 )(4)

[0026] Where ⊕ represents the splicing operation, F outIt represents the final output feature map after multi-scale feature fusion.

[0027] In an embodiment of the present invention, the specific characteristics of the Schizochytrium cell morphology are as follows: the number of lipid droplets in a single vegetative cell is less than 10, with small volume and dispersed inside the cell; the small lipid droplet binary cell divides into two parts, there is a clear partition inside the cell, the liposome has a small volume and does not fill the cell; the large lipid droplet binary cell divides into two parts, there is a clear partition inside the cell, the liposome has a large volume and fills the cell; the small lipid droplet ternary cell divides into three parts, there are clear partitions inside the cell, and the liposomes are small and dense; the large lipid droplet ternary cell divides into three parts, there are clear partitions inside the cell, the liposome has a large volume and fills the cell; the small lipid droplet quaternary cell divides into four parts, there are clear partitions inside the cell, and the liposomes are small and dense; the large lipid droplet quaternary cell divides into four parts, there are clear partitions inside the cell, the liposome has a large volume and fills the cell; the multinucleate cell divides three times or more, and there are multiple partitions inside the cell; the number of lipid droplets in the small lipid droplet single cell is small but more than 10, with small volume and dispersed inside the cell; the lipid bodies in the large lipid droplet single cell continuously fuse, with a diameter greater than 0.8 μm, and the liposome does not fill the cell; the diameter of the lipid bodies in the full large lipid droplet single cell increases to 1 - 3 μm, the liposome fills the whole cell and is tightly arranged; the lipid-droplet-free cell is mostly irregular in shape, the cell boundary is not obvious, and there is no or very little liposome inside the cell; the broken cell fragments are aggregated and distributed in an irregular shape.

[0028] The second object of the present invention is to provide a computer device, which includes a processor and a memory. The memory is used to store computer instructions, and the computer instructions are loaded and executed by the processor to implement the oil production evaluation method based on the chronological morphology classification of Schizochytrium.

[0029] The third object of the present invention is to provide the application of the computer device in the screening of Schizochytrium strains. The application is to use the computer device to evaluate the oil production of the mutagenized or modified strains and screen out the strains with better oil production performance.

[0030] The fourth object of the present invention is to provide the application of the computer device in the fermentation regulation of Schizochytrium. The application is to use the computer device to quickly evaluate the oil production state of Schizochytrium during the fermentation process and regulate the fermentation parameters according to the evaluation results.

[0031] The fifth object of the present invention is to provide a method for judging whether to terminate the fermentation of Schizochytrium. The method is to use the computer device to evaluate the oil production state of Schizochytrium. If it enters the lipid catabolic phase, the fermentation of Schizochytrium is terminated.

[0032] The beneficial effects of the present invention:

[0033] The present invention establishes unified and clear rules based on cell morphology for standardized classification, and combines the improvement of the target detection algorithm to enable it to capture the local features inside the cells more precisely, further deeply explore the relationship between cell types and oil production, realize the real-time prediction of oil production based on the deep learning algorithm, establish a key tool connecting the morphological characteristics of Schizochytrium and the optimization of fermentation processes, and achieve rapid and low-cost quality control in the fermentation industry. Description of the Drawings

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0035] Figure 1 Classification standard for Schizochytrium

[0036] Figure 2 Flow chart for the recognition of Schizochytrium cell images

[0037] Figure 3 Composition of the Schizochytrium cell dataset

[0038] Figure 4 Partition of the validation set and the training set

[0039] Figure 5 Schematic diagram of the MLC-YOLO structure

[0040] Figure 6 MLCA module

[0041] Figure 7 MPSC module

[0042] Figure 8 Comparison of the average accuracy of Schizochytrium cell classification by YOLOv8s and MLC-YOLO

[0043] Figure 9 Visual analysis during the fermentation process of Schizochytrium

[0044] Figure 10 Statistics of Schizochytrium cell types under different fermentation time conditions

[0045] Figure 11 Analysis of the correlation significance between Schizochytrium cell types and oil production Detailed Embodiments

[0046] The present invention patent will be further elaborated below in combination with specific examples. These implementation cases are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0047] Materials and detection methods involved:

[0048] Schizochytrium strains:

[0049] High oil-producing mutant strain of Schizochytrium sp. (CCTCC M20231041) stored in the laboratory

[0050] Microscopic image acquisition:

[0051] Use an OLYMPUS CX31 microscope, combined with a 100× oil immersion objective lens and a 10× eyepiece, equipped with a CCD camera to observe and collect images, and store them in JPG format. CCD camera (MDX10, Mingmei, 2592×1944 resolution, pixel size 2.4μm×2.4μm, 12 bits). A total of 16,088 cell images are provided in the dataset, and the dataset is divided into a training set and a test set according to the principle of 8:2.

[0052] YOLOv5s, YOLOv8s, YOLOv10s, MLC-YOLO object detection models:

[0053] Among them, the reference of YOLOv5s is Liu, J. & Liu, Z. YOLOv5s-BC: an improved YOLOv5s-based method for real-time apple detection. Hubei Digital Manufacturing Key Laboratory, School of Mechanical and Electronic Engineering, Wuhan University of Technology, Wuhan, 430070, China Vol. 21(2024); the reference of YOLOv8s is Ma, M. & Pang, H. SP-YOLOv8s: An Improved YOLOv8s Model for Remote Sensing Image Tiny Object Detection. Changchun Univ Technol, Sch Comp Sci & Engn, Changchun 130012, Peoples R China. Vol. 13, 8161(2023); the reference of YOLOv10s is Silu Zhang, J.W., Kai Yang, Minglei Guan YOLO-ACT: an adaptive cross-layer integration method for apple leaf disease detection. Affiliations School of Computer Science and Software Engineering, University of Science and Technology Liaoning, Anshan, China. Institute of Applied Artificial Intelligence of the Guangdong-Hong Kong-Macao Greater Bay Area, Sh Vol. 15, 1451078(2024).

[0054] Schizochytrium images were obtained by sampling at different time points during the Schizochytrium fermentation process, and the images were marked and classified to form a dataset. Through the training of YOLOv5s, YOLOv8s, and YOLOv10s models, during the training process, the initial learning rate was set to 0.01, the learning rate decay coefficient was 0.0005, the SGD optimizer was used to train the parameters, the momentum value was 0.937, and the BatchSize was 32. A total of 400 Epochs were trained. To prevent overfitting during training and device consumption, after 100 Epochs of training, if the system performance did not improve, the training ended. The best model weights obtained from the training were saved and verified. In the C2f module of the YOLOv8s object detection algorithm, a Mixed Local Channel Attention (MLCA) module was introduced to adaptively adjust the attention weights of different regions and channels, enabling more accurate extraction of key features related to Schizochytrium classification. In the last layer of the Backbone, a Multi-Feature Pyramid Shared Conv (MPSC) module was introduced, and convolutional layers with different dilation rates were introduced to extract the features of Schizochytrium cells at multiple scales, resulting in the MLC-YOLO object detection algorithm.

[0055] The following specifically describes the technical solutions of the present invention in conjunction with specific embodiments. In the following embodiments, unless otherwise specified, the reagents, materials, and equipment used can be obtained from commercial sources, prepared by conventional methods, or commonly used in this industry.

[0056] Example 1:

[0057] Schizochytrium cell culture:

[0058] The Schizochytrium strain preserved at -80°C was transferred to an activation medium and cultured in a constant temperature incubator at 28°C for 3 days. A loopful was scraped with an inoculation loop and inoculated into a seed medium, and cultured at 28°C and 200 r·min -1 for 48 h. Inoculated at an inoculation amount of 10% (v·v -1 ) and transferred for 3 generations as the seed liquid. 100 mL of fermentation medium was placed in a 500 mL Erlenmeyer flask, and the seed liquid was inoculated into the fermentation medium at an inoculation amount of 1:10, and fermented and cultured at 28°C and 200 r·min -1 for 120 h.

[0059] During the Schizochytrium culture process, samples were taken every 12 h for observation. Each time, 200 microscopic images of Schizochytrium were collected (obtained by photographing with an OLYMPUS CX31 microscope), a total of 13,490 single-cell images, constituting a Schizochytrium cell image dataset at different time points.

[0060] The microscopic cell images of Schizochytrium are subdivided into 13 categories according to different cell division methods and the size of lipid droplets inside the cells. The labels, shapes, sizes, and morphological feature descriptions of each category of cell images are as Figure 1 shown.

[0061] The Schizochytrium cell images are divided into single vegetative cells (G1), which have a small number of lipid droplets (<10), small volume, and are dispersed inside the cells; small lipid droplet dyads (G2), in which the cell divides into two parts, there is a clear partition inside the cell, the liposomes are small in volume and do not fill the cell; large lipid droplet dyads (G3), in which the cell divides into two parts, there is a clear partition inside the cell, the liposomes are large in volume and fill the cell; small lipid droplet triads (G4), in which the cell divides into three parts, there are clear partitions inside the cell, and the liposomes are small and dense; large lipid droplet triads (G5), in which the cell divides into three parts, there are clear partitions inside the cell, the liposomes are large in volume and fill the cell; small lipid droplet tetrads (G6), in which the cell divides into four parts, there are clear partitions inside the cell, and the liposomes are small and dense; large lipid droplet tetrads (G7), in which the cell divides into four parts, there are clear partitions inside the cell, the liposomes are large in volume and fill the cell; multi-dyads (G8), which divide three times or more, and there are multiple partitions inside the cell; small lipid droplet single cells (G9), which have a small number of lipid droplets (>10), small volume, and are dispersed inside the cells; large lipid droplet single cells (G10), in which the intracellular liposomes continuously fuse, with a diameter greater than 0.8 μm, and the liposomes do not fill the cell; single cells filled with large lipid droplets (G11), in which the diameter of the intracellular liposomes increases to 1 - 3 μm, the liposomes fill the entire cell, and are arranged closely; lipid droplet-free cells (G12), whose cell shapes are mostly irregular, the cell boundaries are not obvious, and there are no or very few lipidosomes inside the cells; broken cells (G13), whose fragments are aggregated and distributed in irregular shapes, a total of 13 categories, including 6 categories of single cells, 2 categories of dyads, 2 categories of triads, 2 categories of tetrads, and 1 category of multi-dyads. The cell size is about 7 - 20 μm.

[0062] When recognizing cell images, the more categories there are, the more it will lead to a decrease in recognition accuracy and an increase in recognition difficulty, especially for the recognition of similar cells. However, if more systematic and refined classification is not carried out, it will be impossible to analyze the complexity of microbial cells and their growth and metabolism mechanisms. The present invention establishes unified and clear rules based on cell morphology for standardized classification, combined with the improvement of the target detection algorithm, enabling it to more accurately capture the local features inside the cells, further deeply explore the relationship between cell types and oil production, and realize the real-time prediction of oil production based on the deep learning algorithm, establishing a key tool connecting the morphological characteristics of Schizochytrium and the optimization of fermentation processes, and realizing rapid and low-cost quality control in the fermentation industry.

[0063] Example 2:

[0064] After steps such as cell culture, data collection, and dataset preprocessing, the target detection algorithms YOLOv5s, YOLOv8s, and YOLOv10s are selected to train the initial dataset. The process is as Figure 2 shown.

[0065] Among them, dataset preprocessing is to delete the blurred images in the collected Schizochytrium cell image dataset and the images with cell overlap. The number of various cells in the image dataset is as Figure 3 shown, and the dataset is divided into a training set and a validation set according to Figure 4 shown.

[0066] Table 1. Comparison of different YOLO series algorithms

[0067]

[0068] As shown in Table 1, the average accuracy mAP@0.5(%) of YOLOv5s, YOLOv8s, and YOLOv10s is compared and analyzed. The average accuracy of the YOLOv8s model is higher than that of YOLOv5s and YOLOv10s, reaching 81.8%. The average accuracy of G13 broken cells is relatively low, mainly because the broken cells are mostly lipid droplets with irregular shapes and sizes, and the background interference is relatively serious. Algorithm improvement is carried out for this problem.

[0069] Example 3:

[0070] Taking YOLOv8s as the baseline network, the MLC-YOLO network is proposed. Its structure diagram is as Figure 5As shown in the figure, the network structure diagram consists of an input layer (Input), a bottleneck layer (Bottleneck), a detection head (Detect), a feature fusion neck (Neck), and a multi-level feature enhancement module. Bottleneck: Compresses the feature dimension through multiple CBS modules to reduce the computational complexity. Detect: The detection head part, which contains the stacking of CBS and Conv2d, is used to generate preliminary detection results. Neck: The feature fusion neck, through operations such as MLCA-C2f, Concat (feature concatenation), and Upsample (upsampling), fuses multi-scale features to enhance the target localization ability. The network introduces a Mixed Local Channel Attention (MLCA) module after the Upsample and Concat operations in the Neck and in the C2f module of the deep refinement stage of the network. At the initial stage of the Neck, the MLCA-C2f module adjusts the channel attention of the concatenated multi-scale features to enhance the weight of the key channels. After further upsampling and concatenation, the MLCA-C2f module is used again to fuse higher-level semantic information and low-level detail features. After the first CBS in the deep module chain, the MLCA-C2f performs local-global attention interaction on the features to optimize the target localization ability. Between multiple CBS and Conv2d modules, the MLCA-C2f is inserted periodically to continuously enhance the expression ability of multi-scale features. And a Multi-Feature Pyramid Shared Conv (MPSC) module is introduced at the last layer of the Backbone. The MLCA module can more precisely extract the key features related to the classification of Schizochytrium by adaptively adjusting the attention weights of different regions and channels. The MPSC module can extract the features of Schizochytrium cells at multiple scales by introducing convolutional layers with different dilation rates; and the MPSC module adopts a shared convolutional layer design, which can reduce redundant calculations and improve the inference efficiency of the model. Compared with the original Spatial Pyramid Pooling Fast (SPPF) module of YOLOv8s, the MPSC extracts features through convolutional operations, which can more precisely capture the local feature information inside the cells, thereby enhancing the ability to identify cell structures.

[0071] The MLCA module is as Figure 6 shown. Its principle is to first perform local average pooling on the input image to obtain a feature map The pooling size can be adjusted by ks, where F in represents the input feature map, R represents the real number space, and C indenotes the number of input channels, i.e., the depth of the feature map, H denotes the height of the feature map, W denotes the width of the feature map, and ks denotes the pooling kernel size, i.e., the window size used in the local average pooling operation. Subsequently, it is divided into two parts. The first part undergoes global average pooling to obtain the global feature map and then obtains the channel representation through one-dimensional convolution Subsequently, it is restored to the size of the original feature map through un-average pooling (UNAP) It can be expressed by the following formula

[0072] F 1out = UNAP(Conv1d(GAP(F1))) (1)

[0073] Among them, Conv1d represents one-dimensional convolution, that is, performing a convolution operation in the channel dimension (or a specified one-dimensional direction), which is used to adjust the correlation between channels or extract local features. GAP represents global average pooling, that is, taking the average of the values at all spatial positions in each channel of the input feature map and compressing them into a scalar value

[0074] The second part first obtains a new feature map through feature map reconstruction and then further obtains a new channel representation through one-dimensional convolution Subsequently, it is reconstructed again into It can be expressed by the following formula

[0075] F 2out = RS(Conv1d(RS(F1))) (2)

[0076] Among them, RS represents the reconstruction operation

[0077] Add these two parts of feature maps along the same pixel positions to obtain the fused feature map Restore it to the dimension of the original feature map through un-average pooling operation Then perform an element-wise multiplication operation with the original input image at the corresponding pixel points to update the original input feature map and obtain the final attention-weighted feature map That is

[0078]

[0079] Among them, + represents the addition of corresponding channel elements represents the multiplication of corresponding channel elements. This structure makes full use of the spatial and position information of the input image, enabling the model to pay more attention to the cell morphology in the image

[0080] Such as Figure 7As shown, the MPSC module consists of 6 CBS (Conv Batch Normalization and SiLU). k represents the convolution kernel size, and d represents the dilation multiple. For the input feature map , it is first processed by CBS to obtain the intermediate feature map F1, and then F1 is input into the dilated convolutional layer. In each layer, the output feature map is not only local but also fused with the features of the next dilated convolutional layer. The output feature maps of the four dilated convolutions, F 12 , F 123 , F 1234 and F 12345 , can be obtained respectively. After splicing these four feature maps and passing them through CBS, a new feature map is obtained, and its expression is:

[0081] F out = CBS(F1 ⊕ F 12 ⊕ F 123 ⊕ F 1234 ⊕ F 12345 )(4)

[0082] where ⊕ represents the splicing operation. The MPSC module can not only capture local and global features at different scales but also use convolutions with multiple dilation rates to expand the receptive field, thereby improving the feature extraction ability of the model.

[0083] The experimental results of using the MLC-YOLO network to detect Schizochytrium cells are as shown in Figure 8 . The mean average precision (mAP) of MLC-YOLO reaches 84.2% at a confidence level of 0.5, which is 2.4% higher than that of the baseline network YOLOv8s. Due to the similar morphology between the two species G4 and G5, and the similar color and atypical features of G9, G10, and G12 compared to the background, as well as the small number of samples, the detection accuracy of these five types of cells is relatively low compared to other types of cells. And YOLOv8s has a very poor detection effect on G13 because of its diverse morphology and small number. However, the detection accuracy of MLC-YOLO for G13 cells has a significant improvement compared to the baseline network, with a 35.8% increase. With stronger feature extraction ability and more efficient detection ability, MLC-YOLO has a better detection effect when dealing with complex tasks and can maintain a relatively high accuracy when dealing with other types of cells. The accuracies of G5, G9, and G10 are improved by 3.7%, 0.6%, and 1% respectively.

[0084] During the cultivation of Schizochytrium, samples are taken every 12 hours for observation. Each time, 200 microscopic images of Schizochytrium are collected, totaling 13,490 single-cell images, which constitute the Schizochytrium cell image dataset at different time points. As shown in Figure 9As shown, the visual analysis of the experimental results shows that the overall detection performance of the model is excellent, and it can effectively identify non-complete cells located at the edge of the image, highlighting the robustness of the model to complex backgrounds and incomplete targets.

[0085] As Figure 10 shown, during the cell adaptation period of the entire fermentation process, single vegetative cells (G1), small lipid droplet dyads (G2), triads (G4), and tetrads (G6) dominated, and their proportions gradually decreased with the increase of fermentation time. From 36 to 108 h, the proportion of G11 showed an obvious upward trend, with the highest proportion reaching 69%, which is related to the active reproduction and lipid accumulation of cells at this growth stage. After 108 h, the proportions of lipid-free cells (G12) and broken cells (G13) gradually increased to 16%, reflecting the appearance of senescence phenomena as the cell culture time increased.

[0086] To confirm the relationship between the cell types of Schizochytrium and biomass and oil production, Pearson correlation coefficients were used for correlation significance analysis as Figure 11 shown, and it was found that cells filled with large lipid droplets (G11) were significantly positively correlated with fermentation time, biomass, and oil production. Small lipid droplet triads (G4) and small lipid droplet tetrads (G6) were significantly negatively correlated with biomass.

[0087] In application, the cell morphology of Schizochytrium is observed by a detection device based on deep learning algorithms, and the cell morphology is intelligently classified and statistically analyzed, so as to realize the real-time evaluation of oil production and cell morphology, judge the time when the cells enter the oil accumulation period and the lipid consumption period, formulate the supplementation of nutrients or the adjustment of fermentation parameters, and can also be used as a judgment criterion for terminating fermentation.

[0088] The embodiments provided above are not intended to limit the scope covered by the present invention, nor are the described steps intended to limit their execution order. Obvious improvements made by those skilled in the art in combination with the existing common general knowledge also fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for evaluating oil production based on the temporal morphological analysis of Schizochytrium, characterized in that: The steps include: S1. According to the different division modes of Schizochytrium cells and the sizes of lipid droplets inside the cells, the cell morphologies of Schizochytrium are divided into single vegetative cells, small lipid droplet dyads, large lipid droplet dyads, small lipid droplet triads, large lipid droplet triads, small lipid droplet tetrads, large lipid droplet tetrads, multidome, small lipid droplet single cells, large lipid droplet single cells, single cells filled with large lipid droplets, cells without lipid droplets, and broken cells. S2, prepare a data set, annotate the Schizochytrium cell images according to the classification in step S1, and divide the annotated image data set into a training set and a validation set; S3. Build the MLC-YOLO model based on YOLOv8s, which consists of input layer, bottleneck layer, detection head, feature fusion neck and multi-level feature enhancement module; S4, using the training set prepared in step S2 to train the MLC-YOLO model to obtain the best training model; inputting the images in the verification set into the best training model obtained by training to verify and save the best model; S5. Use the best model to detect the image to be detected and determine the oil production situation.

2. The oil production assessment method according to claim 1, characterized in that: The MLC-YOLO model introduces a hybrid local channel attention module after the Upsample and Concat operations in the feature fusion neck and in the C2f module in the deep refinement stage of the network.

3. The oil production assessment method according to claim 2, characterized in that: The hybrid local channel attention module first inputs the image The feature map is obtained by local average pooling The pooling size can be adjusted by ks; where F in represents the input feature map, R represents the real number space, C in represents the number of input channels, H represents the height of the feature map, W represents the width of the feature map, and ks represents the pooling kernel size; it is then divided into two parts, The first part obtains the global feature map through global average pooling. Then obtain the channel representation through one-dimensional convolution Then, the unaverage pooling UNAP is used to restore the original feature map size. It can be expressed by the following formula: F 1out NAP(Conv1d(GAP(F1))) (1) Among them, Conv1d represents one-dimensional convolution, and GAP represents global average pooling; The second part first reconstructs the feature map to obtain a new feature map Then, a new channel representation is obtained by further one-dimensional convolution. Then reconstruct It can be expressed by the following formula: F 2out =RS(Conv1d(RS(F1))) (2) Where RS represents the reconstruction operation; Add these two feature maps along the same pixel position to get the fused feature map Restore the original feature map dimension through the inverse average pooling operation Then perform a dot multiplication operation with the original input image at the corresponding pixel points to update the original input feature map and obtain the final attention-weighted feature map Right now Where + means adding the corresponding channel elements. Represents the multiplication of corresponding channel elements.

4. The oil production assessment method according to claim 1, characterized in that: A multi-scale feature sharing pyramid convolution module is introduced in the last layer of the feature extraction network.

5. The oil production assessment method according to claim 4, characterized in that: The multi-scale feature sharing pyramid convolution module consists of 6 CBSs; for the input The feature map of the first dilated convolution layer is processed by CBS to obtain the intermediate feature map F1, and then F1 is input into the dilated convolution layer; in each layer, the output feature map is fused with the feature of the next dilated convolution layer; four dilated convolution output feature maps F 12 、F 123 、F 1234 and F 12345 , these four feature maps are concatenated and then a new feature map is obtained through CBS, and its expression is: in represents the splicing operation, F out Represents the final output feature map after multi-scale feature fusion.

6. The oil production assessment method according to claim 1, characterized in that: The specific characteristics of the cell morphology of Schizochytrium are that the number of lipid droplets in a single nutrient cell is less than 10, the volume is small, and it is dispersed inside the cell; the small lipid droplet binary cell is divided into two parts, there is a clear partition inside the cell, the liposome is small and does not fill the cell; the large lipid droplet binary cell is divided into two parts, there is a clear partition inside the cell, the liposome is large and fills the cell; the small lipid droplet triplicate cell is divided into three parts, there is a clear partition inside the cell, the liposome is small and dense; the large lipid droplet triplicate cell is divided into three parts, there is a clear partition inside the cell, the liposome is large and fills the cell; the small lipid droplet tetrad cell is divided into four parts, there is a clear partition inside the cell, the liposome is small and dense ; The large lipid droplet tetrad cells divide into four parts, with clear partitions inside the cells, and the liposomes are large and fill the cells; the multidome divides three or more times, with multiple partitions inside the cells; the small lipid droplet single cell has a small number of lipid droplets, but more than 10, with a small volume and dispersed inside the cells; the liposomes inside the large lipid droplet single cell continuously fuse, with a diameter greater than 0.8 μm, and the liposomes do not fill the cells; the diameter of the liposomes inside the single cell filled with large lipid droplets increases to 1-3 μm, the liposomes fill the entire cell and are tightly arranged; the lipid droplet-free cells are mostly irregular in shape, with unclear cell boundaries, and no or very few liposomes inside the cells; the broken cell fragments are aggregated and distributed in irregular shapes.

7. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory is used to store computer instructions, and the computer instructions are loaded and executed by the processor to implement the oil production assessment method based on the temporal morphological classification of Schizochytrium as described in any one of claims 1 to 6.

8. Use of the computer device according to claim 7 in screening Schizochytrium strains, characterized in that: The application is to use the computer device to evaluate the oil production of the induced or modified strains, and screen out strains with better oil production performance.

9. Application of the computer device according to claim 7 in the regulation of Schizochytrium fermentation, characterized in that: The application is to use the computer device to quickly evaluate the oil production status of Schizochytrium during the fermentation process, and to regulate the fermentation parameters according to the evaluation results.

10. A method for determining whether to terminate Schizochytrium fermentation, characterized in that: The method comprises using the computer device of claim 7 to evaluate the oil production state of Schizochytrium, and terminating the fermentation of Schizochytrium if the lipid reverse consumption period is reached.

Images