A method and system for evaluating the activity of stem cells throughout the entire culture process.

Abstract

This invention discloses a method and system for evaluating the activity of stem cells throughout the entire process of stem cell culture, belonging to the field of stem cell culture technology. The technical solution involves periodically acquiring stem cell microscopic images using a microscopic imaging system to obtain images covering the 0–100% confluence range, followed by preprocessing. Weakly supervised segmentation is performed on the preprocessed microscopic images to obtain morphological segmentation maps. The preprocessed microscopic images are then recursively decomposed into texture feature maps using two-dimensional discrete wavelet transform. The segmentation maps and texture feature maps are input into a convolutional neural network for convolution calculation and then flattened to obtain morphological feature vectors and texture feature vectors. The morphological feature vectors and texture feature vectors are concatenated to obtain a fused feature vector. The fused feature vector is then input into an evaluation network to output an activity score. The beneficial effects of this invention are: it provides a method and system for evaluating the activity of stem cells throughout the entire process of stem cell culture.

Description

Technical Field

[0001] This invention belongs to the field of stem cell culture technology, and specifically relates to a method and system for evaluating the activity of stem cells throughout the entire process of stem cell culture. Background Technology

[0002] In vitro expansion and culture of stem cells is a core technology in regenerative medicine, where cell viability directly impacts therapeutic efficacy. However, current manual assessment methods suffer from limitations such as low efficiency, subjective bias, and inconsistent standards, severely hindering large-scale application. The stem cell culture process comprises three stages: seeding, expansion culture, and collection, with expansion culture being the core stage. During this stage, continuous cell proliferation leads to a significant change in confluence from sparse to dense, requiring high-frequency monitoring of stem cells to adjust the culture strategy in a timely manner. However, existing automated technologies struggle to adapt to the complex changes throughout the entire culture process, and their development faces the following key challenges:

[0003] Significant dynamic changes in cell confluence pose a great challenge to the automated assessment of stem cell activity. The main difficulty lies in the distinct challenges of image analysis and feature extraction brought about by the two stages of low and high confluence.

[0004] Under low confluence conditions, cells are sparsely distributed on the culture substrate, resulting in limited effective visual information per unit area, with the background region dominating. This distribution characteristic leads to the following problems: First, the number of morphological and texture features available for determining cell state is insufficient, making it difficult to support reliable feature modeling or classification decisions; second, noise, impurities, or uneven illumination in non-cellular regions of the image are relatively prominent interference factors, easily misleading cell detection algorithms and increasing the risk of false alarms or missed detections. Figure 1 As shown in (a) and (b), cells are sparsely distributed in low-cohesion images, with low effective information density and high background complexity. The discrimination ability of traditional image processing methods in such scenarios is significantly reduced.

[0005] Under high confluence conditions, cell proliferation leads to a sharp increase in density, resulting in large-scale overlap and adhesion between cells, triggering a series of complex problems: First, the boundaries between individual cells become blurred or even disappear due to physical contact, making it difficult for morphological segmentation algorithms to accurately define the outline of individual cells, causing under-segmentation or over-segmentation; second, cell stacking causes aliasing and distortion of morphological features, rendering previously clear morphological discrimination indicators ineffective; furthermore, high density leads to significant occlusion between cells, causing the complete loss of some cell information, further increasing the uncertainty of the evaluation. Figure 1 As shown in (c) and (d) in the images, cells in high-confluence images are clustered together with blurred and indistinguishable boundaries and highly complex morphology and structure, which greatly increases the difficulty of accurately identifying individual cells and assessing their activity.

[0006] In summary, there is an urgent need for an image-based activity assessment technology for the entire stem cell culture process that can simultaneously utilize the complementary advantages of morphology and texture under limited annotation conditions, achieve adaptive processing for different confluence stages, and output standardized quantitative scores. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for evaluating the activity of stem cells throughout the entire process of stem cell culture that integrates morphological and textural features.

[0008] This invention is achieved through the following measures:

[0009] In a first aspect, this embodiment provides a method for evaluating the activity of stem cells throughout the entire process of stem cell culture, characterized by comprising:

[0010] During the stem cell seeding period, logarithmic growth phase, and plateau phase, stem cell microscopic images were periodically acquired using a microscopic imaging system to obtain stem cell microscopic images covering the 0–100% confluence range and then preprocessed.

[0011] Weakly supervised segmentation is performed on the preprocessed microscopic image to obtain a morphological segmentation map;

[0012] Texture feature maps are obtained by recursively multi-level decomposition of the preprocessed microscopic images using two-dimensional discrete wavelet transform.

[0013] After inputting the segmentation map and texture feature map into the convolutional neural network for convolution calculation, they are flattened to obtain the morphological feature vector and texture feature vector.

[0014] The morphological feature vector and the texture feature vector are concatenated to obtain a fused feature vector; the fused feature vector is input into the evaluation network to output an activity score.

[0015] Furthermore, the preprocessing includes:

[0016] Each RGB microscopic image is converted into a grayscale image according to a preset weight;

[0017] Perform illumination normalization and enhancement processing on grayscale images, including:

[0018] Gray-level histogram matching is performed using a reference image as a baseline to eliminate brightness differences caused by shooting at different times.

[0019] Sharpen and enhance the contrast of the image to highlight cell boundaries and intercellular textures.

[0020] Furthermore, weakly supervised segmentation is performed on the preprocessed microscopic image to obtain a morphological segmentation map, including:

[0021] A segmentation network was constructed based on the U-Net architecture, and its encoder was pre-trained on unlabeled stem cell images using the Barlow Twins self-supervised learning method, so that the network could obtain feature representations that are invariant to changes in illumination and geometric transformations.

[0022] On a small number of images with cell boundary annotations, the pre-trained U-Net is fine-tuned under weak supervision. The following segmentation loss is used during training:

[0023] Weighted binary cross-entropy loss is used to address the problem of a severe imbalance in the number of cell pixels relative to the background pixels, by assigning higher weights to cell pixels.

[0024] Size constraint loss: Based on the statistical range of stem cell diameters, a penalty is applied to the equivalent diameter of each connected region to ensure it is within the range specified in the standard. Within the range, thus suppressing noise debris and non-physiological adhesion areas;

[0025] Shape regularity constraint loss: Based on the length-to-width ratio of physiologically normal cells, minimize the deviation between the length-to-width ratio of each connected region and the ideal length-to-width ratio, thereby avoiding unreasonable cell shapes that are too long or too flat.

[0026] Boundary smoothness constraint loss: using the cosine of the angle between the vectors formed by adjacent boundary points as an indicator, it suppresses abrupt changes in local boundary angles and avoids non-physiological jagged boundaries.

[0027] By training U-Net using the segmentation loss described above, the network can obtain high-precision morphological segmentation maps with limited labeled data.

[0028] Furthermore, the preprocessed microscopic image is recursively decomposed into multiple layers using two-dimensional discrete wavelet transform to obtain texture feature maps, including:

[0029] A two-dimensional discrete wavelet transform is performed on the preprocessed image, and multi-level decomposition is carried out according to the preset number of decomposition layers. After one level, a low-frequency approximate sub-band LL and three high-frequency detail sub-bands LH, HL and HH are generated.

[0030] High-frequency detail subbands of each decomposition layer are preserved, and bilinear upsampling is used to unify the high-frequency detail subbands of different levels to match the original image or target features. Figure 1 The required space dimensions;

[0031] The upsampled LH, HL, and HH subbands at each scale are spliced ​​together along the channel dimension to obtain multi-scale, multi-directional texture feature maps.

[0032] Furthermore, the feature vectors will be fused. The input evaluation network outputs an activity score, including:

[0033] fuse feature vectors The input evaluation network, through fully connected layers and a sigmoid activation function, normalizes the model output to a preset activity score range, thus obtaining the activity score of the corresponding stem cell image.

[0034] Furthermore, the evaluation network is trained using expert ratings as supervision labels, employing root mean square error as the regression loss, and combining it with the segmentation loss to form a joint loss for training and optimization.

[0035] Secondly, this embodiment provides an activity evaluation system for the entire process of stem cell culture, characterized in that it includes:

[0036] The system comprises the following modules: an image acquisition and preprocessing module for acquiring and preprocessing stem cell microscopic images covering the 0–100% confluence range; a weakly supervised segmentation module for performing weakly supervised segmentation on the preprocessed microscopic images to obtain morphological segmentation maps; a texture analysis module for recursively decomposing the preprocessed microscopic images using two-dimensional discrete wavelet transform to obtain texture feature maps; a statistical feature calculation module for inputting the segmentation maps and texture feature maps into a convolutional neural network for convolution calculation and then flattening to obtain morphological feature vectors and texture feature vectors; a fusion module for concatenating the morphological feature vectors and texture feature vectors to obtain a fused feature vector; and a scoring output module for outputting an activity score based on the fused feature vector.

[0037] Thirdly, this embodiment provides an electronic device, characterized in that it includes: a processor and a memory, wherein the processor is used to execute a program stored in the memory for an activity evaluation method for the entire process of stem cell culture, so as to realize the activity evaluation method for the entire process of stem cell culture.

[0038] The beneficial effects of the technical solution provided by the embodiments of this invention are as follows: This application proposes a dual-branch network that integrates morphological and textural features, realizing dynamic, accurate, and robust automated evaluation of the entire culture process. This method extracts cell morphological features through weakly supervised segmentation under morphological constraints, captures texture information by combining wavelet multi-scale texture analysis, and finally achieves complementary heterogeneous features through dynamic adaptive fusion, significantly improving the performance consistency of the model under different convergence conditions. Attached Figure Description

[0039] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings listed below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1These are images with different degrees of convergence in the background technology;

[0041] Figure 2 A flowchart of a method for evaluating the activity of stem cells throughout the entire process of stem cell culture, as described in this embodiment of the invention;

[0042] Figure 3 A comparison of Dice coefficients with and without constraints;

[0043] Figure 4 Comparison of average boundary distances with and without constraints;

[0044] Figure 5 To compare the model evaluation results with the human evaluation results;

[0045] Figure 6 A graph showing the root mean square error between the predictive activity of the network and the expert evaluation benchmark;

[0046] Figure 7 To evaluate the performance variation curves (accuracy variation curves) of the network and three variants across the full convergence range.

[0047] Figure 8 To evaluate the performance variation curves (RMSE value variation curves) of the network and three variants across the full convergence range. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] Example 1:

[0050] This embodiment provides a method for evaluating the activity of stem cells throughout the entire process of stem cell culture, characterized by comprising:

[0051] S1. During the stem cell seeding period, logarithmic growth phase and plateau phase, periodically acquire stem cell microscopic images using a microscopic imaging system to obtain stem cell microscopic images covering the 0–100% confluence range and perform preprocessing.

[0052] The preprocessing includes:

[0053] Each RGB microscopic image is converted to a grayscale image according to a preset weight. First, the three-channel RGB image is converted to a single-channel grayscale image. The conversion process uses a linear combination based on the weights perceived by the human eye. The specific formula is as follows:

[0054]

[0055] Where R, G, and B are the intensity values ​​of the red, green, and blue channels of the original image, respectively; This transformation, which preserves the structural information of cell morphology and texture, effectively reduces computational complexity while retaining the values ​​of individual pixels.

[0056] Perform illumination normalization and enhancement processing on grayscale images, including:

[0057] Gray-level histogram matching is performed using a reference image as a baseline to eliminate brightness differences caused by shooting at different times.

[0058] Sharpen and enhance the contrast of the image to highlight cell boundaries and intercellular textures.

[0059] S2. Perform weakly supervised segmentation on the preprocessed microscopic image to obtain a morphological segmentation map, specifically including:

[0060] A segmentation network was constructed based on the U-Net architecture, and its encoder was pre-trained on unlabeled stem cell images using the Barlow Twins self-supervised learning method, so that the network could obtain feature representations that are invariant to changes in illumination and geometric transformations.

[0061] On a small number of images with cell boundary annotations, the pre-trained U-Net is fine-tuned under weak supervision. The following segmentation loss is used during training:

[0062] Weighted binary cross-entropy loss is used to address the problem of a severe imbalance in the number of cell pixels relative to the background pixels. It assigns higher weights to cell pixels, enhancing the ability to identify cells. Its definition is as follows:

[0063]

[0064] Where M represents the total number of pixels in the image. Represents pixels The true label, with a background of 0 and a cell of 1; The model predicts the probability that the pixel belongs to a cell; This is a coefficient used to balance the weights of cell pixels to compensate for their small number. This loss function provides pixel-level supervision signals, enabling the model to accurately distinguish cells from the background.

[0065] Size constraint loss: Based on the statistical range of stem cell diameters, a penalty is applied to the equivalent diameter of each connected region to ensure it is within the range specified in the standard. Within a certain range, noise debris and non-physiological adhesion areas are suppressed, as defined below:

[0066]

[0067] in, The equivalent diameter of the i-th connected region is obtained by converting the region area to the circle diameter, and the unit is micrometers. Its value is calculated by converting the area of ​​the connected region to the equivalent area of ​​the circle. and These represent the biologically reasonable minimum and maximum cell diameter thresholds, respectively; N is the total number of connected regions. This constraint suppresses the occurrence of noise fragments and non-physiological fusion regions by penalizing regions that exceed the threshold range.

[0068] Shape regularity constraint loss: Based on the aspect ratio of a physiologically normal cell, minimize the deviation between the aspect ratio of each connected region and the ideal aspect ratio, thereby avoiding unreasonable cell shapes that are too long or too flat. Its definition is as follows:

[0069]

[0070] in, Let be the aspect ratio of the i-th connected region. Let be the ideal aspect ratio constant, and N be the total number of connected regions. This function minimizes the difference between the actual and ideal aspect ratios, promoting segmentation results that better reflect biological morphology.

[0071] Boundary smoothness constraint loss: Using the cosine of the angle between the vectors formed by adjacent boundary points as an indicator, it suppresses abrupt changes in local boundary angles and avoids non-physiological jagged boundaries. Its definition is as follows:

[0072]

[0073] In the formula The angle between two vectors formed by three consecutive points on the cell boundary is used to measure the degree of local turning; K is the total number of boundary points. By maximizing the average value of the cosine of the angle between the direction vectors of adjacent boundary points, the segmentation result is made to maintain the smooth and continuous characteristics that real cells generally possess.

[0074] The segmentation loss integrates all the above loss terms and is used to train U-Net, enabling the network to obtain high-precision morphological segmentation maps with limited labeled data. Its definition is:

[0075]

[0076] in , and These are hyperparameters used to adjust the weights of various morphological constraints.

[0077] S3. Texture feature maps are obtained by recursively decomposing the preprocessed microscopic images using two-dimensional discrete wavelet transform. The two-dimensional discrete wavelet transform, with its multi-scale decomposition capability and direction selectivity, can adaptively capture the spatial structural features of cell populations at different degrees of convergence, thus providing robust texture representation for the entire culture process. Specifically, this includes:

[0078] The preprocessed image is subjected to a two-dimensional discrete wavelet transform and decomposed into multiple levels according to a preset number of decomposition layers. After one level, a low-frequency approximate sub-band LL (used to characterize the global contour of the image) and three high-frequency detail sub-bands LH, HL, and HH (capturing texture information in the horizontal, vertical, and diagonal directions, respectively) are generated.

[0079] The low-frequency subbands can be further decomposed to extract higher-level semantic features. High-frequency detail subbands at each decomposition level are preserved, and bilinear upsampling is used to unify the high-frequency detail subbands at different levels with the original image or target features. Figure 1 The required space dimensions;

[0080] The upsampled LH, HL, and HH subbands at each scale are spliced ​​together along the channel dimension to obtain multi-scale, multi-directional texture feature maps.

[0081] S4. Input the segmentation map and texture feature map into a convolutional neural network for convolution calculation, and then flatten them to obtain the morphological feature vector. and texture feature vector ;

[0082] S5. Transfer the morphological feature vector With texture feature vector The fused feature vector is obtained by concatenating the features. S6. Fuse the feature vectors The input evaluation network outputs an activity score, which specifically includes the fused feature vectors. The input is fed into the evaluation network, which uses fully connected layers and a sigmoid activation function to normalize the model output to a preset activity score range, thus obtaining the activity score of the corresponding stem cell image. The evaluation network employs an MLP (Multi-Level Processing) network.

[0083] Furthermore, the evaluation network is trained using expert ratings as supervision labels and root mean square error as the regression loss. and the segmentation loss Combining losses The evaluation network is then trained and optimized.

[0084] The combined loss is:

[0085]

[0086] in Hyperparameters to balance the importance of the two tasks.

[0087] The root mean square error (RMSE) is used to measure the absolute deviation between the model's predicted activity score and the expert evaluation. The formula is:

[0088]

[0089] Where n is the test sample size. The average score, obtained through independent evaluation by three researchers, serves as the standard for assessing stem cell activity. This refers to the model's predicted activity score.

[0090] Considering that model predictions and expert evaluations are unlikely to be completely consistent, this paper sets the following criterion: if the difference between the model prediction score and the expert evaluation score is within 5 points, the evaluation result is considered correct; otherwise, it is considered incorrect. To further verify the model's adaptability to continuous changes in confluence during cell expansion, this paper designs a confluence robustness verification method: the cell confluence range of 0–100% is divided into 10 intervals, and the local RMSE is calculated in each interval. The smaller the RMSE, the closer the model prediction result is to the expert evaluation.

[0091] Example 2:

[0092] Building upon Example 1, this example involves periodically acquiring stem cell microscopic images using a microscopic imaging system during the stem cell seeding, logarithmic growth, and plateau phases. These images cover the 0–100% confluence range and are then preprocessed. To ensure a balanced distribution of cell confluence, images are acquired every 3 hours. The acquired images are divided into a training set (3200 images) and a test set (700 images).

[0093] Low, medium, and high confluence images comprise 30%, 40%, and 30% of the total training set, respectively. Among these, 480 images have fine annotations, with cell boundary segmentation and annotation completed image by image using the LabelMe tool. The test set includes 500 normal cell images, 100 extremely high confluence cell images, and 100 pathological morphology images.

[0094] The evaluation network is trained using the training set and validated using the test set, including cell segmentation performance evaluation and cell viability evaluation accuracy evaluation.

[0095] Cell segmentation performance was comprehensively evaluated using two dimensions: geometric accuracy and confluence adaptability. Geometric accuracy was quantified using the Dice similarity coefficient (DSC) and average boundary distance (ABD).

[0096]

[0097] Where X and Y represent the pixel sets of the predicted segmentation region and the ground truth labeled region, respectively. The DSC value range is [0,1], and a higher value indicates better consistency of the segmentation boundary.

[0098] The average boundary distance is calculated based on the improved Hausdorff distance formula:

[0099]

[0100] In the formula, Y represents the set of true boundary points, and y represents the true boundary points. For predicting boundary points, d is the Euclidean distance operator, and Z is the total number of predicted boundary points. This index directly reflects the boundary positioning accuracy; the smaller the ABD value, the higher the boundary positioning accuracy.

[0101] like Figure 3 and Figure 4 The quantitative results show the performance of the evaluation network across different convergence intervals. Overall, it achieved an average Dice coefficient of 0.839 and an average boundary distance of 1.3 pixels on the test set, indicating acceptable consistency between the segmentation results and human annotations. Notably, the segmentation performance in each convergence interval is better than that without constraints after adding segmentation loss constraints.

[0102] Furthermore, the performance difference between manual evaluation and evaluation network prediction was compared using the average value of expert evaluation results as a benchmark. Three evaluators independently completed the image evaluations; all evaluators had three or more years of experience in stem cell projects. To comprehensively evaluate the overall performance of the evaluation network in stem cell activity assessment tasks, validation was conducted from multiple dimensions. Figure 5 A subset of test data was selected to demonstrate the similarity between the network evaluation and three human evaluations, based on the average of expert assessment results. Both the network evaluation and the three human evaluations were conducted on the same test dataset (e.g., ...). Figure 5 As shown, the closest 100 units to the outer edge indicate a greater similarity to the baseline. Figure 5 The results clearly demonstrate that the accuracy and stability of the evaluation network are significantly higher than those of the manual evaluation. The solid blue line represents the evaluation network's results, while the dashed line represents the manual evaluation results.

[0103] The root mean square error (RMSE) between the predictive activity of the evaluation network and the expert evaluation benchmark was 2.2, significantly lower than the average of 3.9 for the three human evaluators. On a 0-100 scale, this RMSE represents a relative error of only 2.2%, indicating that the evaluation network's predictions closely approximate the expert benchmark, achieving an excellent level of accuracy. In contrast, the RMSE among human evaluators was 3.9, representing a relative error of 3.9%, which is within an acceptable range and reflects typical subjective differences among evaluators. However, the evaluation network's error is significantly lower than that of human evaluations. This result not only confirms its superior accuracy compared to human evaluations but also demonstrates the consistency and reliability of its output, providing ample evidence for the practical application of automated evaluation systems.

[0104] Example 3:

[0105] Based on Example 2, the convergence adaptability of the evaluation network was verified. Figure 6 Grouped bar charts were used to compare and display the RMSE values ​​and accuracy of the evaluation network across different confluence ranges. The results show that the model maintains optimal performance within the 30%-80% confluence range; performance only slightly decreases in the low confluence range of 0%-30%; and while performance decreases somewhat in the high confluence range of 90%-100%, it still maintains high accuracy. This performance variation mainly stems from the differences in the reliability of information modalities at different confluence levels: under low confluence conditions, fewer cells lead to decreased confidence in texture information, making evaluation more reliant on morphological features; under high confluence conditions, increased cell overlap makes morphological feature extraction difficult, making evaluation more reliant on texture information; neither modality alone can achieve high-precision evaluation. However, in the medium confluence range of 30%-80%, both morphological and texture features have high confidence, effectively complementing each other and significantly improving the model's judgment accuracy. These results fully demonstrate the strong adaptability of this evaluation network to dynamic changes in confluence during culture.

[0106] Example 4:

[0107] This embodiment provides a system for evaluating the activity of stem cells throughout the entire process of stem cell culture. It is characterized by comprising: an image acquisition and preprocessing module for acquiring and preprocessing microscopic images of stem cells covering the 0–100% confluence range; a weakly supervised segmentation module for performing weakly supervised segmentation on the preprocessed microscopic images to obtain a morphological segmentation map; a texture analysis module for recursively decomposing the preprocessed microscopic images using two-dimensional discrete wavelet transform to obtain a texture feature map; and a statistical feature calculation module for performing convolutional encoding on the segmentation map and the texture feature map respectively and calculating the Gram matrix to obtain a morphological feature vector. and texture feature vector The gated fusion module inputs the segmentation map and texture feature map into a convolutional neural network for convolutional computation, and then flattens them to obtain morphological feature vectors. and texture feature vector ;

[0108] morphological feature vectors With texture feature vector The fused feature vector is obtained by concatenating the features. ;

[0109] The scoring output module is used to output scores based on the fused feature vector. Output activity score.

[0110] The functions of each module are described in the description of the activity evaluation method for the entire process of stem cell culture in Example 1, and will not be repeated here.

[0111] Example 5:

[0112] Based on Example 1 or Example 4, the contribution of the core module is verified through three variant model systems: Variant 1: Retaining texture-related features and removing morphology-related features. This verifies the importance of texture feature columns and multi-scale analysis for cell adhesion scenarios.

[0113] Variant 2: Retain morphological features and remove texture features to verify the key role of cell morphology information in evaluating the network.

[0114] Variant 3: Retain texture and morphological features, but remove the constraints of the segmentation loss to verify the effect of explicit knowledge constraints on segmentation accuracy.

[0115] All variants were tested under the same experimental conditions in 10 uniform segments within the 0-100% cell confluence range.

[0116] Figure 7 and Figure 8 The performance variation curves of this evaluation network and three variants are shown across the full convergence range. Figure 7 This is the accuracy change curve. Figure 8 This is a curve showing the change in RMSE values.

[0117] Retaining morphological features while removing texture features (i.e., removing the wavelet texture analysis module) significantly degrades the network's performance, with the decline becoming more pronounced as cell confluence increases. Accuracy drops to 67.0% and RMSE rises to 5.4 in the 90-100% confluence range. This indicates that texture features play a crucial role in the evaluation process, especially in multi-cell adhesion scenarios. As cell confluence increases, cell boundaries become blurred and overlap intensifies; texture analysis can provide macroscopic information on cell distribution and arrangement, compensating for the limitations of morphological analysis alone.

[0118] Texture-related features are retained, while morphology-related features are removed. This variant performs poorly in the low confluence range, with an accuracy of only 72%-78%, indicating that in sparse cell environments, the lack of accurate morphological information severely restricts model performance. The morphological features of individual cells are the most direct indicator for assessing cell state, while texture features struggle to provide equally valuable discriminative information in this context. As confluence increases, the contribution of texture features becomes more apparent, leading to a slight improvement in the variant's performance. However, it remains below the standard model across the entire confluence range, highlighting the importance of morphological features in evaluation.

[0119] The variant that removes morphological prior constraints exhibits the most stable performance characteristics, second only to the full evaluation network under all test conditions. The accuracy difference between this variant and the standard evaluation network is within 2%-3% across the full convergence range, validating the important role of morphological prior knowledge injection in improving model evaluation capabilities.

[0120] Ablation experiments revealed a complementary mechanism between texture and morphological features in different cell confluence environments. Texture analysis plays an indispensable role under high confluence conditions, effectively addressing the recognition challenges posed by cell adhesion and overlap; while morphological analysis forms the foundation of the entire cell recognition system, providing the most direct evidence for cell state assessment.

[0121] It is worth noting that the contributions of the two features are not simply additive, but rather exhibit a synergistic enhancement effect. The model achieves significant performance improvements by organically combining texture analysis and morphological information, particularly reaching optimal recognition results in the medium convergence range. This synergistic effect indicates that the population distribution information provided by texture features and the individual cell information provided by morphological features can mutually verify and complement each other, jointly constructing a more robust and accurate cell state assessment system.

[0122] Example 6:

[0123] This embodiment provides an electronic device, characterized in that it includes a processor and a memory, wherein the processor is used to execute a program stored in the memory for a fault sample generation method based on seismic data from oil and gas fields, so as to realize the activity evaluation method for the entire process of stem cell culture described in the above embodiment.

[0124] An electronic device includes at least one processor, memory, at least one network interface, and other user interfaces. The various components of the electronic device are coupled together via a bus system. It is understood that the bus system is used to enable communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0125] The user interface may include a display, keyboard, or clicking device (e.g., mouse, trackball, touchpad, or touchscreen). It is understood that the memory in this embodiment may be volatile memory or non-volatile memory, or may include both.

[0126] In this embodiment of the invention, the processor executes the method steps provided in each method embodiment by calling a program or instruction stored in the memory, specifically a program or instruction stored in an application program.

[0127] In some implementations, the memory stores elements such as executable units or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0128] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. The program implementing the method of this invention can be included in the application programs.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the activity of stem cells throughout the entire process of culturing the same, characterized by, The method comprises the following steps: During the stem cell inoculation period, the logarithmic growth period and the plateau period, periodic acquisition of stem cell microscopic images is performed by a microscopic imaging system, and the stem cell microscopic images covering the interval of 0-100% confluence are obtained and preprocessed; Weakly supervised segmentation is performed on the preprocessed microscopic images to obtain a morphological segmentation map; Recursive multi-layer decomposition is performed on the preprocessed microscopic images by two-dimensional discrete wavelet transform to obtain a texture feature map; After the segmentation map and the texture feature map are input into a convolutional neural network for convolution calculation, flattening is performed to obtain a morphological feature vector and a texture feature vector; The morphological feature vector and the texture feature vector are spliced to obtain a fusion feature vector; The fusion feature vector is input into an evaluation network to output an activity score; Weakly supervised segmentation is performed on the preprocessed microscopic images to obtain a morphological segmentation map, comprising: A segmentation network is constructed based on a U-Net architecture, and the encoder of the segmentation network is pre-trained on unlabeled stem cell images by using a Barlow Twins self-supervised learning method; The pre-trained U-Net is weakly supervised fine-tuned on a small number of images with cell boundary annotations, and the following segmentation loss is used during training, comprising: A weighted binary cross-entropy loss is used to give higher weight to cell pixels, and the definition is as follows: wherein M represents the total number of pixels in the image, represents the true label of the pixel point , 0 for background and 1 for cell; is the probability that the model predicts the pixel belongs to the cell; is a coefficient for balancing the weight of cell pixels; Size constraint loss: according to the statistical range of stem cell diameter, the equivalent diameter of each connected region is punished to be within the range, which is defined as follows: wherein, D i represents the equivalent diameter of the i-th connected region, which is converted from the area of the connected region into the diameter of a circle, and is in units of micrometers, and is calculated by converting the area of the connected region into the area of a circle; and D min and D max are the minimum and maximum cell diameter thresholds, respectively, which are biologically reasonable; and N is the total number of connected regions. A shape regularity constraint loss is used to minimize the deviation between the aspect ratio of each connected region and the ideal aspect ratio according to the aspect ratio of a normal cell in physiology, and the definition is as follows: wherein, is the aspect ratio of the i-th connected region, is the ideal aspect ratio constant, and N is the total number of connected regions; A boundary smoothness constraint loss is used to use the vector angle cosine formed by adjacent boundary points as an indicator to suppress local boundary angle mutations, and the definition is as follows: wherein represents the angle between two vectors formed by three consecutive points on the cell boundary, which is used to measure the degree of local turning; K is the total number of boundary points; The segmentation loss integrates all the above loss terms to train the U-Net, so that the network obtains a high-precision morphological segmentation map under limited labeled data, and the definition is as follows: wherein , and are hyperparameters for adjusting the weights of the various shape constraints. Recursive multi-layer decomposition is performed on the preprocessed microscopic images by two-dimensional discrete wavelet transform to obtain a texture feature map, comprising: Two-dimensional discrete wavelet transform is performed on the preprocessed images, and multi-level decomposition is performed according to a preset decomposition level, and one low-frequency approximation subband LL and three high-frequency detail subbands LH, HL and HH are generated by one level; The high-frequency detail subbands of each decomposition layer are retained, and a bilinear upsampling method is used to unify the high-frequency detail subbands of different levels to the spatial size consistent with the original image or the target feature map; The upsampling L, HL and HH subbands at each scale are spliced in the channel dimension to obtain a multi-scale and multi-direction texture feature map.

2. The method of claim 1, wherein the active evaluation is performed by a processor. The preprocessing comprises: Each RGB microscopic image is converted into a grayscale image according to a preset weight; Illumination normalization and enhancement processing is performed on the grayscale image, comprising: Gray histogram matching is performed based on a reference image to eliminate the brightness difference caused by shooting at different time points; Sharpening and contrast enhancement are performed on the image to highlight the cell boundaries and intercellular textures.

3. The method of claim 1, wherein the active evaluation is performed by a processor. fuse feature vectors The input evaluation network outputs an activity score, including: The fusion feature vector is input into the evaluation network, the model output is normalized to a preset activity score interval through a fully connected layer and a Sigmoid activation function, and an activity score corresponding to the stem cell image is obtained.

4. The method of claim 1, wherein the active evaluation is performed by a processor. The training of the evaluation network takes expert scores as supervision labels, adopts root mean square error as a regression loss, and is trained and optimized with the segmentation loss to form a joint loss.

5. An activity evaluation system for the entire process of stem cell culture, characterized by, The method comprises the following steps: An image acquisition and preprocessing module is configured to acquire stem cell microscopic images covering a 0-100% confluence interval and perform preprocessing; A weakly supervised segmentation module is configured to perform weakly supervised segmentation on the preprocessed microscopic images to obtain a morphological segmentation map; A texture analysis module is configured to perform recursive multi-layer decomposition on the preprocessed microscopic images through two-dimensional discrete wavelet transform to obtain a texture feature map; A statistical feature calculation module is configured to input the segmentation map and the texture feature map into a convolutional neural network for convolution calculation, and then flatten to obtain a morphological feature vector and a texture feature vector; A fusion module is configured to concatenate the morphological feature vector and the texture feature vector to obtain a fusion feature vector; A score output module is configured to output an activity score based on the fusion feature vector. The weakly supervised segmentation on the preprocessed microscopic images obtains a morphological segmentation map, including: A segmentation network is constructed based on a U-Net architecture, and the encoder of the segmentation network is pre-trained on unannotated stem cell images using a Barlow Twins self-supervised learning method; wherein M represents the total number of pixels in the image, represents the true label of the pixel point , 0 for background and 1 for cell; is the probability that the model predicts that the pixel belongs to the cell; is a coefficient for balancing the weight of cell pixels; Size constraint loss: according to the statistical range of stem cell diameter, the equivalent diameter of each connected region is punished to be within the range, which is defined as follows: wherein, D i represents the equivalent diameter of the i th connected region, which is converted from the area of the connected region into the diameter of a circle, and is in units of microns, and is calculated by converting the area of the connected region into the area of a circle; and D min and D max are the minimum and maximum cell diameter thresholds, respectively, which are biologically reasonable; and N is the total number of connected regions. The pre-trained U-Net is weakly supervised fine-tuned on a small number of images with cell boundary annotations, and the following segmentation loss is used during training, including: wherein, is the aspect ratio of the i-th connected region, is the ideal aspect ratio constant, and N is the total number of connected regions; A weighted binary cross-entropy loss, which gives higher weight to cell pixels, is defined as follows: wherein represents the angle between two vectors formed by three consecutive points on the cell boundary, which is used to measure the degree of local turning; K is the total number of boundary points; A shape regularity constraint loss: according to the aspect ratio of a normal cell in physiology, the deviation between the aspect ratio of each connected region and the ideal aspect ratio is minimized, which is defined as follows: wherein , and are hyperparameters for adjusting the weights of the various shape constraints. A boundary smoothness constraint loss: the vector angle cosine formed by adjacent boundary points is used as an indicator to suppress local boundary angle mutations, which is defined as follows: The segmentation loss integrates all the above loss terms to train the U-Net, so that the network obtains a high-precision morphological segmentation map under limited annotated data, which is defined as follows: The texture feature map is obtained by performing recursive multi-layer decomposition on the preprocessed microscopic images through two-dimensional discrete wavelet transform, including: Performing two-dimensional discrete wavelet transform on the preprocessed image, performing multi-level decomposition according to the preset decomposition level, generating one low-frequency approximation subband LL and three high-frequency detail subbands LH, HL and HH through one level; 6. An electronic device, comprising: The high-frequency detail subbands of each decomposition layer are retained, and a bilinear upsampling method is used to unify the high-frequency detail subbands of different levels to the spatial size consistent with the original image or the target feature map; The upscaled LH, HL and HH subbands are concatenated in the channel dimension to obtain a multi-scale, multi-direction texture feature map. The method comprises the following steps: A processor and a memory, wherein the processor is configured to execute a program stored in the memory for an activity evaluation method for the whole process of stem cell culture to realize the activity evaluation method for the whole process of stem cell culture according to any one of claims 1-4.

Images