Cell image intelligent analysis method with multi-dimensional adaptive evolution capability

Abstract

This invention relates to the fields of computer vision and life science image analysis technology, and particularly to an intelligent cell image analysis method with multi-dimensional adaptive evolution capabilities. The method includes: acquiring cell images; selecting an optimal model based on the imaging mode and number of channels of the cell images; using the optimal model to infer performance indicators for monitoring; determining whether the performance indicators exceed a preset normal range, or whether the number of newly added annotations reaches a preset threshold, or whether the time interval between the optimal model's last training and the current training period reaches a preset cycle; if so, selecting high-value images from unlabeled cell images based on uncertainty, sample diversity, and marginal cases, acquiring weak annotations for the high-value images and generating dense pseudo-labels, and outputting labeled images; initiating incremental training to update the optimal model, and returning to the step of obtaining prediction results using the optimal model. This invention solves the problem of cell image analysis models being difficult to continuously evolve.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and life science image analysis technology, and in particular relates to an intelligent cell image analysis method with multi-dimensional adaptive evolution capabilities. Background Technology

[0002] In life science research, accurately and efficiently detecting cells and calculating confluence or transfection efficiency from microscope images is a crucial step in experimental analysis. However, existing image analysis tools generally have significant shortcomings: on the one hand, general-purpose software such as ImageJ and CellProfiler rely on manually designed features and parameter adjustments, which is extremely unfriendly to non-professional users, while commercial software such as Harmony and MetaXpress is expensive and has a rigid workflow; on the other hand, existing AI models mostly use fixed weights, and once the cell type, microscope model, lighting conditions, or staining method changes, the recognition accuracy will drop significantly, resulting in poor robustness.

[0003] Furthermore, existing deep learning-based image analysis models (such as convolutional neural networks, CNNs) generally employ a static weight paradigm. While this significantly improves recognition accuracy on specific datasets, the parameters remain fixed once the model is trained and deployed in software. This leads to the following technical drawbacks: (1) The cell types, culture densities, imaging equipment and lighting conditions used during training often deviate from the data distribution in actual user applications (domain drift). Since the model cannot adapt dynamically, when new cell morphologies appear, different fluorescent dyes are used, or the microscope light source ages slightly, the model's recognition accuracy will drop significantly, failing to meet the diverse needs of real scientific research scenarios and exhibiting poor robustness.

[0004] (2) In daily use, researchers manually correct a small number of images under difficult or new conditions (generating valuable labeled data). However, existing tools cannot automatically integrate this incremental labeled data back into the model to achieve continuous learning and performance evolution. Every time a new task or new experimental condition is encountered, users need to collect a large number of labeled samples and retrain the model from scratch, which makes it difficult to accumulate knowledge and consumes a lot of resources.

[0005] In summary, there is still a lack of cell image analysis solutions with self-learning and continuous evolution capabilities in this field, which makes it difficult to meet the urgent need for intelligent and adaptive analysis tools in modern high-throughput life science experiments. Summary of the Invention

[0006] In view of this, the present invention aims to provide a cell image intelligent analysis method with multi-dimensional adaptive evolution capabilities, so as to solve the technical problem that cell image analysis models are difficult to continuously evolve under the dual constraints of variable imaging conditions and high annotation costs.

[0007] To achieve the above objectives, the technical solution created by this invention is implemented as follows: In a first aspect, the present invention provides a cell image intelligent analysis method with multi-dimensional adaptive evolutionary capabilities, comprising: S10. Obtain at least one cell image; S20. Determine the analysis task based on the imaging mode and number of channels of the cell image, and select the optimal model based on the analysis task; S30. Using the optimal model, reason about the cell image to obtain a prediction result, and extract or calculate the performance index for monitoring from the prediction result; S40. Determine whether the performance index exceeds a preset normal range, or whether the number of newly added annotations reaches a preset threshold, or whether the time interval between the optimal model and the last training reaches a preset period; if so, then S50. Based on uncertainty, sample diversity, and marginal cases, select high-value images from unlabeled cell images, obtain weak annotations for the high-value images, generate dense pseudo-labels, and output labeled images. S60. Start incremental training to update the optimal model, and return to step S30. The mixed training dataset for incremental training includes the labeled images.

[0008] Optionally, step S10 further includes detecting the quality of the cell image, and if the quality of the cell image is unqualified, then the cell image is discarded or replaced.

[0009] Optionally, the number of newly added annotations in step S40 includes at least one of the following: the number of cell images annotated by the user, the percentage of the area of ​​annotated cell pixel regions in a single cell image, the number of cell images with dense pseudo-labels, and the number of externally imported annotated cell images.

[0010] Optionally, after step S20, the method further includes the following step: The optimal model is evaluated on the validation set according to the preset evaluation cycle. When multiple consecutive evaluations show that the key indicators decrease one after another and the cumulative decrease exceeds the preset decay threshold, the process proceeds to step S50.

[0011] Optionally, between steps S50 and S60, the following steps are included: Determine whether the distribution difference between the cell image and the training set of the optimal model exceeds a preset difference threshold; if yes, start the lightweight domain adaptation process; if no, proceed to step S60.

[0012] Optionally, the lightweight domain adaptation process includes the following steps: S71. Add a domain adaptation layer to the optimal model, and freeze all parameters of the backbone network in the optimal model to obtain the adaptation model structure. S72. Construct a loss function based on the adapted model structure; S73. Using the adapted model structure as the initial state and the loss function as the optimization objective, perform lightweight training, automatically align the feature distribution through adversarial training or style transfer techniques, and output the trained adapted layer parameters. S74. Integrate the trained adaptation layer parameters into the adaptation model structure to generate the adaptation model; S75. Replace the optimal model with the adapted model.

[0013] Optionally, step S60 includes: S61. Freeze the backbone network of the optimal model to obtain a trainable network structure; S62. Load all the weights of the optimal model into the network structure to obtain the initial model; S63. Representative old samples are sampled proportionally from historical data and added to the hybrid training dataset through a memory playback mechanism. S64. Calculate the Fisher information matrix of historical important parameters and construct the EWC regularization loss function; S65. The initial model is incrementally fine-tuned using the hybrid training dataset to obtain the fine-tuned model; S66. Evaluate the performance of the fine-tuned model on an independent validation set, and determine whether to replace the optimal model with the fine-tuned model.

[0014] Optionally, if the optimal model is replaced with the fine-tuned model, then after step S66, the following steps are included: saving the fine-tuned model as a new version, with the version number automatically incrementing according to a preset rule, and automatically generating a version record. The version record includes at least one of the following: version number, training time, triggering reason, training data volume, training duration, performance comparison between the old and new models, performance improvement, forgetting rate, training parameters, and model size.

[0015] In a second aspect, the present invention provides a computer device, characterized in that it comprises: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the intelligent cell image analysis method with multi-dimensional adaptive evolution capability described in this invention.

[0016] Secondly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to cause the computer to execute the intelligent cell image analysis method with multi-dimensional adaptive evolution capability described in the present invention.

[0017] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability created by the present invention allows laboratory personnel to build a high-precision model without coding knowledge through weak annotation and intelligent guidance. The model update is triggered by three conditions, realizing rapid adaptation to changes in imaging conditions and cell state.

[0018] (2) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability described in this invention integrates elastic weight solidification and memory playback strategy in the mixed training dataset of incremental training. By constraining historical important parameters and mixing representative old samples, catastrophic forgetting is avoided, and the model is ensured to remain reliable under the accumulation of multiple batches of data.

[0019] It eliminates the need for extensive user annotation and automatically adapts to cross-domain differences such as different microscope models, lighting, staining, and backgrounds. Through unsupervised adversarial training or style transfer to align feature distribution, it significantly improves the robustness of the model in real experimental scenarios.

[0020] (3) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability described in this invention is based on a sample screening method that combines uncertainty, sample diversity, and edge case joint ranking. It combines weak annotation and semi-supervised learning to prioritize pushing blurry cells, low-confidence regions, and cross-batch difference samples for users to annotate, thereby maximizing model performance improvement with minimal manual input.

[0021] (4) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability described in this invention actively detects the performance degradation of the model. When the performance of the model degrades, it can automatically initiate training without waiting for the user to add new data, ensuring a rapid response to changes in imaging conditions, cell state, background texture, etc.

[0022] (5) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability described in this invention can continuously monitor model status and environmental changes, automatically trigger retraining and safely accumulating historical knowledge, and achieve stable and high-precision analysis across devices and conditions.

[0023] (6) The cell image intelligent analysis method with multi-dimensional adaptive evolution capability created by the present invention significantly reduces the threshold for using AI image analysis, realizes a user-specific evolution model that is "more accurate and more stable with use", and significantly improves the efficiency and reproducibility of cell biology research. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic flowchart of the intelligent cell image analysis method with multi-dimensional adaptive evolution capability described in the embodiments of the present invention; Figure 2 A schematic diagram of the structure of the computer device described in the embodiment of the present invention.

[0025] Explanation of reference numerals in the attached figures: 62. Computer equipment; 64. External devices; 66. Processing unit; 68. Bus; 70. Network adapter; 72. Input / output (I / O) interface; 74. Display; 78. System memory; 80. Random access memory (RAM); 82. Cache memory; 84. Storage system; 90. Program / utility; 92. Program module. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and do not constitute a limitation thereof. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined to form various implementations. Furthermore, the order of the steps or actions in the method description can be changed or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various orders in the specification and drawings are merely for the clear description of a particular embodiment and do not imply a mandatory order, unless otherwise stated that a particular order must be followed.

[0028] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. The term "based on" should be understood as "at least partially based on." Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more, and the term "including" means "including but not limited to." Various embodiments of the present invention may exist in the form of a range; it should be understood that the description in the form of a range is merely for convenience and brevity and should not be construed as a rigid limitation on the scope of the invention; therefore, it should be considered that the range description has specifically disclosed all possible sub-ranges and single numerical values ​​within that range; for example, it should be considered that the range description from 1 to 6 has specifically disclosed sub-ranges, such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and single numbers within the range, such as 1, 2, 3, 4, 5, and 6, regardless of the range. Furthermore, whenever a numerical range is referred to herein, it means including any referenced number (fraction or integer) within the range referred to.

[0029] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0030] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] Example 1 like Figure 1 As shown, this embodiment provides a cell image intelligent analysis method with multi-dimensional adaptive evolution capabilities, including the following steps: S10. Obtain at least one cell image; It can acquire at least one cell image, supporting imaging modes such as brightfield, fluorescence, confocal, and phase contrast, and supports single-channel or multi-channel image input. Users can upload images in batches through the image import interface, or connect to microscope equipment in real time to acquire images.

[0032] Preferably, after acquiring the cell image, the quality of the cell image is checked. If the quality of the cell image is unqualified, the cell image is discarded or replaced (the user is prompted to replace or re-acquire the image). The above quality check includes the following three criteria; failure to meet any one criterion results in the image being deemed unqualified: (1) Blur detection: Calculate the Laplacian variance of the image. If it is less than 0.3 times the average variance of the global image library, it is judged as blurry. (2) Exposure detection: If the proportion of pixels with a brightness value greater than 240 in a bright field image exceeds 15%, or the proportion of pixels with a brightness value greater than 240 in a fluorescence image exceeds 20%, it is judged as overexposed; (3) Contamination detection: Use a pre-trained foreign object detection model to identify non-cellular regions. If a continuous block is detected with an area greater than 10% of the total image area and there is no corresponding structure in the cell channel, it is determined that there is contamination.

[0033] If the quality detection result of the cell image is qualified, the cell image is retained and the process proceeds to step S20.

[0034] Furthermore, before proceeding to step S20, annotations for cell images that have passed quality inspection can be received. Specifically, the user performs weak annotations on cell images that have passed quality inspection through interactive image annotation operations.

[0035] In this invention, weak annotation of cell images supports the following three specific forms: (1) Point labeling: The user clicks on the mark at the center of the cell or a key location, and the coordinates of the point are obtained as the foreground seed point. This method is suitable for scenarios where the center of the cell is clearly distinguishable, such as cell nucleus segmentation and well-separated single cells, and each cell only needs to be clicked once.

[0036] (2) Doodle annotation: The user roughly paints the cell area with a brush, or draws a line or several dots through the target area. At this time, all pixels in the painted area are obtained as foreground seed points, and the background area can be selectively annotated. This method is suitable for irregular cells, adhered cell clusters, or cells with blurred boundaries. Each cell cluster only requires 2 to 5 doodles.

[0037] (3) Box annotation: The user draws the bounding rectangle of the cell or cell cluster with a rectangle, and the system records the coordinates of the upper left and lower right corners of the rectangle. This method is suitable for densely populated areas of cells or scenarios where a large number of cells need to be quickly labeled, and only one rectangle needs to be drawn for each cell cluster.

[0038] The point annotation above uses the marked points as foreground seed points. A graph cut algorithm is used to expand these seed points to the entire cell region, generating a complete cell segmentation mask. Doodle annotation uses pixels within the doodle area as foreground seed points, with the area covered by the doodle as the initial constraint. A graph cut algorithm is used to optimize the boundaries, generating a fine-grained segmentation mask. Box annotation uses the area within the rectangular box as the foreground candidate region, with the box boundary as the spatial constraint. Combined with the model's initial predictions, a graph cut algorithm is used to finely segment the cell boundaries within the box.

[0039] This invention designates labeled cell images as labeled images and employs a semi-supervised learning strategy to generate dense pseudo-labels corresponding to these cell images. The semi-supervised learning strategy utilizes user-provided weak annotations as seed points and automatically labels adjacent or similar unlabeled regions through self-training, pseudo-label generation, or consistency regularization algorithms to generate dense pseudo-labels.

[0040] S20. Determine the analysis task based on the imaging mode and number of channels of the cell image, and select the optimal model based on the analysis task; The analysis task (such as confluence calculation, cell counting, transfection efficiency, subcellular localization, cell viability, 3D cell nucleus segmentation, colocalization analysis, etc.) is determined based on the imaging mode and number of channels of the cell image. The corresponding optimal model is selected according to the analysis task. The available optimal models include: four-layer U-Net (LightweightVersion with 4 Layers), Mask R-CNN, YOLO (You Only LookOnce), DeepLabv3+ (Deep Learning for Semantic Image Segmentation 3rd Edition Enhanced Version), U-Net with Morphological Classification Head, and 3D U-Net.

[0041] In practical implementation, the analysis task and the corresponding optimal model can be determined based on Table 1 below: Table 1

[0042] Of the analytical tasks described above, confluence calculation is often the most basic (requiring only cell region segmentation) and can be performed independently. Cell counting requires the detection of individual cells and typically involves instance separation based on segmentation (which may depend on confluence results). Transfection efficiency requires the simultaneous classification of positive and negative cells and depends on cell detection or segmentation results.

[0043] Therefore, when deciding on the analysis task, if the user does not specify, the confluence calculation will be performed first by default, and then the user will be guided to perform cell counting or transfection efficiency analysis. However, the user can choose arbitrarily, and in this case, the necessary intermediate steps will be automatically supplemented (for example, when selecting transfection efficiency, cell segmentation will be performed automatically first).

[0044] S30. Using the optimal model, reason about the cell image to obtain a prediction result, and extract or calculate the performance index for monitoring from the prediction result; This performance metric is determined based on the analysis task. More specifically, the analysis task, prediction results, and corresponding performance metrics are shown in Table 2 below: Table 2

[0045] S40. Determine whether the performance index exceeds the preset normal range, or whether the number of newly added annotations reaches the preset threshold, or whether the time interval between the optimal model and the last training reaches the preset period; if yes, proceed to step S50; if no, directly output the prediction result.

[0046] The present invention employs a triple-condition joint triggering method, which will automatically trigger model update (i.e., steps S50 and S60) as long as any one of the conditions is met.

[0047] For the first condition, namely determining whether the performance indicator exceeds a preset normal range, this preset normal range can be determined comprehensively based on task type, application scenario requirements, historical performance statistics, industry standards, and user-defined parameters; this invention does not impose any limitations on it. For indicators where smaller values ​​are better (such as MAE, relative error, and forgetting rate), a value higher than a preset upper threshold is considered "yes"; for indicators where larger values ​​are better (such as mAP, IoU, R², F1, and accuracy), a value lower than a preset lower threshold is considered "yes".

[0048] The second condition involves determining whether the number of newly added annotations reaches a preset threshold. This number includes at least one of the following: the number of user-annotated cell images, the percentage of labeled cell pixel regions in a single cell image, the number of cell images with dense pseudo-labels, and the number of externally imported labeled cell images. Each of these newly added annotation quantities has a corresponding preset threshold. If any one of these newly added annotation quantities reaches its corresponding preset threshold, the second condition is deemed satisfied. The aforementioned number of newly added annotations and their corresponding preset thresholds are as follows: The number of cell images annotated by the user is the number of cell images annotated through weak annotation, with a preset threshold of 10~20.

[0049] The percentage of the area of ​​labeled cell pixel regions in a single cell image refers to the percentage of the area of ​​labeled pixel regions to the total area of ​​the image for each labeled cell image; the corresponding preset threshold is 4% to 7%, for example, 5%.

[0050] The number of cell images with dense pseudo-labels refers to the number of cell images corresponding to the dense pseudo-labels automatically generated through step S50 or semi-supervised learning; the corresponding preset threshold is 3~10, for example, 5.

[0051] The number of externally imported labeled cell images refers to the number of images imported from external sources that contain complete annotation information. The corresponding preset threshold is 3~10, for example, 5.

[0052] The third condition involves determining whether the time interval between the optimal model's last training iteration and its current online state has reached a preset period. "Last training iteration" refers to the point in time when the currently online optimal model most recently completed training and was deployed, including both initial training and incremental training scenarios. This preset period can be designed according to requirements, such as 24 hours or one week.

[0053] If all three conditions are met negatively, the prediction result will be output directly. The prediction result includes cell count, confluence, transfection efficiency, and confidence interval.

[0054] Cell counting involves counting the instance segmentation masks (such as Mask R-CNN) or detection boxes (YOLO) output by the model, with each independent instance counted as one cell. If cells are stuck together and cannot be separated, a watershed algorithm or a distance transform-based post-segmentation counting method is used.

[0055] The confluence is calculated as the sum of the pixel areas occupied by all cells / the total pixel area of ​​the image × 100%. If multiple cell types are involved (such as transfected positive and negative cells), the percentage of each type is calculated separately.

[0056] The confidence interval adopts the model uncertainty quantification method - Monte Carlo Dropout: during inference, the same image is forward-propagated 10 times (20% of neurons are randomly dropped each time) to obtain 10 segmentation / detection results, and the standard deviation of each cell count and confluence is calculated.

[0057] S50. Based on uncertainty, sample diversity, and marginal cases, select high-value images from unlabeled cell images, obtain weak annotations for the high-value images, generate dense pseudo-labels, and output labeled images. This invention uses multi-level active learning to automatically select the most valuable images worthy of annotation from a massive amount of unlabeled cell images, prioritizes pushing them to users for weak annotation, and expands the weak annotations into dense pseudo-labels through a semi-supervised learning strategy, thereby maximizing the contribution of newly annotated data to the improvement of model performance while minimizing the workload of manual annotation.

[0058] The aforementioned multi-level active learning includes the first layer, the second layer, the third layer, and the output layer.

[0059] The first layer is a multi-dimensional joint ranking layer, which calculates scores for unlabeled cell images based on three dimensions: uncertainty, sample diversity, and marginal cases. Uncertainty is measured using the minimum confidence level: U(x) = 1 - max(p(y|x)); where p(y|x) is the model's class probability distribution for sample x, max(p(y|x)) is the maximum class probability (confidence level), and U(x) is the uncertainty. A larger U(x) value indicates greater uncertainty about the model for that sample. The model is the optimal model selected in step S20. It should be noted that the sample in this invention is a cell image, including labeled cell images (annotated images) and unlabeled cell images.

[0060] Sample diversity is measured using the cosine similarity of the feature space, and its expression is: , Where, x i For unlabeled cell images, f i For unlabeled cell images x i eigenvectors; The feature centers of the labeled cell image. This represents sample diversity; the larger the value, the higher the diversity.

[0061] Edge case flags are identified using a combination of batch distribution differences and confidence scores. More specifically, the Maximum Mean Discrepancy (MMD) is calculated between the feature distribution of the current batch and the feature distribution of historical batches. If MMD > a threshold (default 0.15), a significant distribution shift is considered in the current batch. In this case, cell images in the batch with a prediction confidence < 0.7 are marked as edge cases, denoted by the edge case flag E(x) = 1. The edge case flag E(x) is a binary variable, E(x) ∈ {0, 1}, where: E(x) = 1 indicates that sample x is marked as an edge case; E(x) = 0 indicates that sample x is not an edge case.

[0062] The second layer is the high-value screening layer, which selects the top-K samples as high-value images. The images are ranked based on a weighted total score calculated from the three factors: uncertainty, sample diversity, and marginal case markers. The weighted total score is calculated as follows: Weighted Total Score = 0.5 × Uncertainty + 0.3 × Sample Diversity + 0.2 × Marginal Case Marker. Unlabeled cell images are then sorted from highest to lowest weighted total score, and the top K unlabeled cell images are selected as high-value images. The value of K ranges from 5 to 20, for example, 10.

[0063] It is evident that high-value images possess the following characteristics: the model is most uncertain about them, they differ the most from the labeled set, or they are on the edge of cross-batch distributions. Labeling these samples can achieve the greatest improvement in model performance with minimal manual input.

[0064] The third layer is the annotation and pseudo-label generation layer, which uses a semi-supervised learning strategy to generate dense pseudo-labels on the annotated images; specifically, it includes the following steps: S501. Use the optimal model to infer the labeled image and generate an initial segmentation mask.

[0065] Specifically, the labeled image is input into the optimal model, and the optimal model outputs a probability map of each pixel belonging to the foreground (cell). The initial binary mask is obtained by thresholding (the threshold is 0.5 by default).

[0066] S502. Construct constraints based on the type of weak annotations in the annotated image.

[0067] For point annotations, the marked point and its nearest neighbor region (with a radius of 3 to 5 pixels) are used as foreground seed points; for graffiti annotations, the pixels within the graffiti area are used as foreground seed points, and if the user also annotates the background area, it is used as the background seed point; for bounding box annotations, the area inside the bounding box is used as the foreground candidate area, and the area outside the bounding box near the boundary is used as the background seed point.

[0068] S503: The graph cut algorithm is used to optimize the segmentation boundary and output a fine binary mask.

[0069] The energy function of the graph cut algorithm consists of a data term and a smoothing term: the data term is based on the initial prediction probability and weak labeling constraints, forcing foreground seed points to be foreground and background seed points to be background; the smoothing term is based on the color differences between pixels, encouraging adjacent similar pixels to have the same label. By minimizing the energy function, the optimized fine segmentation boundary is obtained.

[0070] S504. The optimized fine binary mask is used as a dense pseudo-label for semi-supervised training.

[0071] The output layer outputs an annotated image consisting of high-value images and dense pseudo-labels. It can simultaneously record the following statistical dimensions: the number of user-annotated cell images, the percentage of the area of ​​an annotated cell pixel region in a single cell image relative to the total area of ​​the image, and the number of cell images with dense pseudo-labels.

[0072] S60. Start incremental training to update the optimal model, and return to step S30 with the updated optimal model. The mixed training dataset for incremental training includes the labeled images.

[0073] Step S60 includes: S61. Freeze the backbone network of the optimal model to obtain a trainable network structure; Specifically, all parameters of the optimal model backbone are frozen so that it does not participate in gradient updates during subsequent training. The backbone is responsible for extracting general features of the cell images (such as edges, textures, and shapes). Freezing the backbone preserves the knowledge already learned by the model and prevents catastrophic forgetting during incremental training. Simultaneously, only the top-level networks relevant to the analysis task are unfrozen.

[0074] S62. Load all the weights of the optimal model into the network structure to obtain the initial model; S63. Representative old samples are sampled proportionally from historical data and added to the hybrid training dataset through a memory playback mechanism. When constructing the hybrid training dataset, newly added labeled images can be merged with representative old samples sampled proportionally from historical data through a memory playback mechanism.

[0075] The representative old sample ratio mixing ratio r=N in the memory playback mechanism replay / N new (N) replay N represents the number of representative old samples. new The number of newly labeled images can be determined based on the task similarity σ, with the specific expression being:

[0076] Where, r base =1.0, task similarity Among them, MMD (Maximum Mean Discrepancy) is a statistic used to measure the degree of difference between two data distributions. A larger MMD value indicates a greater difference between the old and new feature distributions. Task Similarity The similarity between the new and old tasks is represented by r, ranging from 0 to 1. When the new and old tasks are significantly different (σ close to 0), r increases to approximately 2.0, indicating that more historical samples need to be replayed (or even doubled) to prevent forgetting. When the new and old tasks are highly similar (σ close to 1), r decreases to approximately 0.2, indicating that only a small number of old samples need to be replayed, allowing the model to learn more from the new data. The adjusted r value is truncated between 0.2 and 2.0 to ensure that the proportion of replayed samples is neither too low (leading to forgetting) nor too high (leading to insufficient learning of new data).

[0077] Furthermore, based on the mixing ratio r of representative old samples and new samples in the memory replay mechanism determined by task similarity σ, the current mixing ratio r can be further adjusted according to the risk of forgetting.

[0078] first, For the old category, Define a forgetting risk index F for the current category. risk =max{ D That is, among all the old categories (cell types that have been learned in previous training), find the category with the largest decrease in IoU (Intersection over Union), and use this maximum value as the current forgetting risk indicator to quantify the degree to which the model has forgotten old knowledge.

[0079] ; in, This represents the intersection-union ratio (IUU) of the current optimal model on class c after training on the old task. This represents the intersection-over-union ratio (IoU) of the current optimal model when testing class c again after learning a new task. This indicates taking the maximum value.

[0080] The rules for adjusting the mixing ratio are as follows: If there is no risk of forgetting (F) risk If the value is less than 0.05, it indicates that the current optimal model maintains good performance for the old class. In this case, a low replay strategy is adopted, and r = 0.5 is set, which means that the number of replay samples is half of the number of new samples. If significant forgetting (F) exists risk If the value is greater than 0.15, it indicates that the current optimal model has significantly forgotten some old categories. In this case, it is necessary to increase the replay ratio and set r = min(2.0, r...). prev +0.2), where r prev This indicates the current playback ratio, which will gradually increase by 0.2 from the current playback ratio, but will not exceed the upper limit of 2.0; If the risk of forgetting is at an intermediate level (0.05 ≤ F) riskIf the concentration is ≤0.15, then the current mixing ratio remains unchanged.

[0081] Furthermore, the representative old samples in the memory replay mechanism are obtained through a typical sample selection algorithm. Specifically, K-means clustering is performed on each cell morphology feature, with the number of clusters K determined based on the number of cell categories to be analyzed in the specific task (usually set to 3 to 5). The image corresponding to the center of each cluster is taken as the typical sample. Every time 10 new annotations are added, clustering is re-executed and the typical sample set is updated to ensure that the cached samples always maintain representativeness of the current data distribution. Replay samples are randomly sampled in each training batch, ensuring the balance of old categories and avoiding excessive or ignored replay of certain categories.

[0082] S64. Calculate the Fisher information matrix of historical important parameters and construct the EWC (Elastic Weight Fixed) regularization loss function; The constructed EWC regularization loss function is: ; Where λ is the regularization intensity hyperparameter; For the current optimal model, the i The values ​​of the parameters; The optimal model saved before incremental training i The values ​​of the parameters; The diagonal element of the Fisher information matrix represents the first element. i The importance of these parameters to the old task.

[0083] The definition of is: ,in, Let be the partial derivative, representing the value of the th element of the optimal model saved before incremental training. i Differentiate each parameter; The logarithm (log-likelihood) of the optimal model prediction probability saved before incremental training. Let be the mathematical expectation.

[0084] In this invention, the regularization intensity hyperparameter λ is adjusted as follows: , Where, λ base =1000, Forgetting Risk Index The value range is [0,1]. (·) represents the hyperbolic tangent function. When the risk of forgetting is high (ρ approaches 1), λ increases to approximately 2000, imposing stronger constraints on historically important parameters to prevent forgetting; when there is no forgetting (ρ approaches 0), λ decreases to approximately 500, allowing the model to learn new knowledge more flexibly. The adjusted λ value is truncated between 500 and 2000 to ensure that the regularization strength is neither too weak (leading to forgetting) nor too strong (hindering the learning of new knowledge).

[0085] Building upon this, a knowledge stabilization and accumulation mechanism is also included to ensure that the model maintains a high recognition rate for cell morphology and imaging features learned early on under multiple batches and conditions of accumulated data. The knowledge stability S is defined as follows. stab for: ; Where T is the time window of the three most recent model updates. For the IoU on the fixed historical test set after the t-th update, The IoU on this test set after the first learning is given. This fixed historical test set consists of N most representative images from the early period, where N is 150 to 300, for example, 200.

[0086] The precipitation criterion is: when three consecutive updates satisfy S stab When the IoU is ≥ 0.95 (i.e., IoU fluctuation is less than 5%), and the IoU of the most recent update has decreased by more than 3% compared to the previous update where no category appeared, the knowledge is considered to have been stably accumulated. At this point, the elastic weight is reduced to solidify the regularization strength (λ is halved), and the memory replay ratio is relaxed (r is reduced to 0.2) to accelerate the absorption of new knowledge. When the accumulation standard is not met, a higher regularization strength and replay ratio are maintained to consolidate historical knowledge.

[0087] S65. The initial model is incrementally fine-tuned using the hybrid training dataset to obtain the fine-tuned model; During fine-tuning, hyperparameters such as batch size and learning rate are automatically recommended based on the current dataset size. Loss curves and validation metrics are displayed in real time. An early stopping strategy is implemented based on validation set performance, and the best model is automatically saved as the fine-tuning model based on optimal conditions across multiple metrics. In specific implementation, step S65 includes the following steps S651 to S653: S651. Automatically recommend hyperparameters based on the current dataset size; In this embodiment, recommended hyperparameters include batch size and learning rate.

[0088] The recommended batch size takes into account factors such as GPU memory size, analysis task type, model complexity, and dataset size. Specifically, it automatically detects available GPU memory and limits the batch size to no more than 80% of the available memory to avoid memory overflow. Depending on the analysis task, it recommends smaller batch sizes (2-8) or larger batch sizes (8-32). Lightweight models such as U-Net recommend batch sizes of 4-16, while heavyweight models such as Mask R-CNN recommend batch sizes of 2-8. Large datasets can use slightly larger batch sizes to accelerate convergence, while small datasets use slightly smaller batch sizes to prevent overfitting.

[0089] The recommended learning rate takes into account factors such as the current training stage and the size of the dataset. If the current training stage is the first training, a larger learning rate (0.001) should be used to achieve fast convergence; if the current training stage is for incremental fine-tuning, a smaller learning rate (0.0001~0.0005) should be used to avoid destroying learned knowledge; a larger learning rate can be used for large datasets, and a smaller learning rate can be used for small datasets.

[0090] The recommendation algorithm is based on a Bayesian optimized pre-trained model. Input features include dataset size, image size, number of model parameters, and GPU model. Outputs the recommended batch size and learning rate. Users can also manually override the recommended values ​​to meet specific experimental needs.

[0091] S652, adopting a four-indicator combined early stop; During training, the validation set loss, validation set IoU, learning rate, and number of training epochs are monitored in real time. Training is automatically stopped when all four of the following conditions are met: Condition ①: The validation set loss does not decrease for 5 consecutive epochs, and the relative decrease is less than 0.001, indicating that the model has converged on the loss function; Condition ②: The validation set IoU does not improve for 5 consecutive epochs, and the improvement is less than 0.005, indicating that the model has saturated in terms of segmentation accuracy; Condition ③: The learning rate has been reduced to 1 / 100 of the initial learning rate, indicating that the optimization process has converged sufficiently and the benefits of continuing training are limited; Condition 4: The minimum number of training epochs has been completed (default 20 epochs) to ensure that the model has undergone at least sufficient iterations and to avoid underfitting due to premature stopping.

[0092] In addition, a mandatory early stopping exception rule is set: if the validation set loss does not decrease for 10 consecutive epochs, training will be forcibly stopped even if the IoU is still slowly increasing, to prevent model overfitting. Users can enable or disable the early stopping function with one click through the interface, and can customize parameters such as the number of epochs for early stopping, the threshold, and the minimum number of training epochs.

[0093] S653, Automatically saves the best model based on multi-index Pareto optimality; During training, after each epoch, the current model's values ​​on three metrics are calculated: validation set IoU, validation set mAP, and inference time (or model size).

[0094] If a new model outperforms all historical models on at least one metric and is no worse than the historical best on other metrics (i.e., the Pareto frontier), it is saved as a candidate best model.

[0095] The final model saved is the one with the highest overall score in the Pareto frontier (overall score = IoU + 0.5×mAP - 0.01×inference time).

[0096] Users can manually choose whether to prioritize optimizing accuracy or speed.

[0097] S66. Evaluate the performance of the fine-tuned model on an independent validation set, and determine whether to update the optimal model accordingly.

[0098] The performance metrics evaluated on the validation set include mAP, IoU, MAE, R², Dice, accuracy, and F1; the specific performance metrics vary depending on the analysis task. The criterion for deciding whether to update the optimal model can be determined based on whether the performance gain exceeds the update threshold; different performance metrics have different update thresholds.

[0099] If the performance improvement meets expectations, the optimal model is updated to the fine-tuned model, and the process returns to step S30. If the performance improvement does not meet expectations (e.g., IoU gain <5% or still below the user-defined target), the current optimal model is not replaced, and the process returns to step S50 to reacquire labeled images. If the process returns to step S50 for three consecutive rounds, but the performance improvement still does not meet expectations or is still unsatisfactory, the process returns to step S20 to reselect the optimal model.

[0100] Preferably, if the optimal model is replaced with the fine-tuned model, then after step S66, the following steps are included: saving the fine-tuned model as a new version, with the version number automatically incrementing according to a preset rule, and automatically generating a version record. The version record includes at least one of the following: version number, training time, triggering reason, training data volume, training duration, performance comparison between the old and new models, performance improvement, forgetting rate, training parameters, and model size.

[0101] After each incremental training iteration, if the fine-tuned model passes validation and is deployed, the model is automatically saved as a new version. This includes recording metadata such as version number, training time, triggering reason, training data volume, and training duration, as well as a performance snapshot, including key metrics such as mean AP, IoU, and inference time. Inference time is recorded as the average inference time (milliseconds) for a single image. During testing, the model is run 10 times on 10 fixed representative images, and the average value is taken to exclude the cold start overhead of the first run. Image resolution (e.g., 1024×1024) is also recorded to compare time performance at different resolutions.

[0102] Version management employs a dual limit strategy: the most recent 20 versions are retained by default, with a total storage space limit of 5GB. The smaller of these two limits is used as the actual limit. When either the number of versions or the storage space limit is exceeded, an automatic cleanup process is triggered: the worst-performing and oldest versions are deleted first. The performance score is calculated as: Score = (IoU - 0.5) × 2 + (mAP - 0.3) × 1.5, where mAP (mean Average Precision) is the average precision. Versions with a score below 0.2 and a retention time exceeding 30 days are automatically cleaned up. Furthermore, users can mark important versions (e.g., add a star); marked versions are permanently retained and do not participate in any automatic cleanup processes.

[0103] This method also supports version comparison and rollback operations: users can view a list of historical versions in the interface, compare performance metrics (such as IoU, mAP, and inference time change curves) of different versions, and select any historical version to perform a one-click rollback operation. During rollback, the model weight file of that version is loaded from the version repository, replacing the current online service model, and the version record is updated, marking the time, reason, and target version number of the rollback operation.

[0104] The following specific examples further illustrate this embodiment: User A photographed HEK293 cells using a bright-field microscope. The initial model confluence error was 15%. After annotating 5 images, the user fine-tuned the error to 3%. Two weeks later, a change in the culture medium brand caused a change in background texture. The system detected that the average prediction confidence dropped from 0.92 to 0.68 (below the threshold of 0.7), automatically triggering incremental training. The user only needed to annotate 2 edge case images, and the model error recovered to 3.5%. Simultaneously, due to the activation of elastic weight fixation, the model could still accurately identify cell morphology under the old culture medium, without any forgetting.

[0105] Example 2 Based on Example 1, this example provides a cell image intelligent analysis method with multi-dimensional adaptive evolution capabilities, including steps S10 to S60, which are the same as in Example 1 and will not be repeated here. In this example, after step S20, the method further includes the following step: The optimal model is evaluated on the validation set according to the preset evaluation cycle. When multiple consecutive evaluations show that the key indicators decrease one after another and the cumulative decrease exceeds the preset decay threshold, the process proceeds to step S50.

[0106] The validation set employs a dynamic update strategy. Specifically, 10% of all historical labeled images are randomly sampled and used as a fixed validation set. Every time five new labeled images are added, one image is randomly selected from this sample and added to the validation set. The validation set is automatically updated every time five new labeled images are added, ensuring that the validation set represents the latest data distribution.

[0107] The key performance indicators (KPIs) mentioned above are weighted composite scores, with weights preset according to the type of analysis task and also allowing users to manually adjust them. The composite score for the segmentation task is: Score = 0.7 × mIoU + 0.3 × pixel accuracy; the composite score for the detection task is: Score = 0.6 × mAP + 0.4 × F1-score; and the composite score for the convergence and counting tasks is: Score = 0.8 × (1 - MAPE) + 0.2 × average confidence level. Where MAPE is the mean absolute percentage error.

[0108] In this embodiment, the key indicators are evaluated three times consecutively, and the overall score decreases each time, with the cumulative decrease reaching or exceeding 5%. When this condition is met, it is determined that the model's performance is continuously deteriorating, and the process proceeds to step S50.

[0109] This step enables proactive detection of performance degradation of the current best model. When the performance of the best model degrades, training is initiated automatically without waiting for the user to add new data, ensuring a rapid response to changes in imaging conditions, cell state, background texture, etc.

[0110] Example 3 Based on Embodiment 1 or 2, this embodiment provides a cell image intelligent analysis method with multi-dimensional adaptive evolution capabilities, including steps S10 to S60, which are the same as in Embodiment 1 and will not be repeated here. Between steps S50 and S60, the following is included: Determine whether the distribution difference between the cell image and the training set of the optimal model exceeds a preset difference threshold; if yes, initiate the lightweight domain adaptation process; if no, proceed to step S60. The preset difference threshold is 0.12~0.18, for example, 0.15.

[0111] More specifically, a joint detection method using Multi-Kernel Maximum Mean Difference (MK-MMD) and Fréchet Inception Distance (FID) is employed to more accurately determine whether the distributional differences between newly acquired cell images and the training set are significant. MK-MMD extracts features from multiple intermediate layers of Inception-v3, calculates the distance between source and target domain features in the Hilbert space of the regenerating kernel, and sets a trigger threshold of 0.2 (normalized). This threshold is determined by bootstrap sampling of the source domain itself (randomly splitting the source domain into two groups, calculating the MMD distribution, and taking the 95th percentile). FID calculates the Wasserstein distance between the Gaussian distribution of the source and target domain features in the Inception-v3 pooling layer, and sets a trigger threshold of 1.5 times FID_ref, where FID_ref is the average FID of 10-fold cross-validation within the source domain (typically 15 to 30). The joint judgment uses "OR" logic: when MK-MMD>0.2 or FID>1.5 × FID_ref, a significant distribution shift is determined, triggering the lightweight domain adaptation process; otherwise, the lightweight domain adaptation process is skipped, and incremental training is directly initiated.

[0112] The lightweight domain adaptation process includes the following steps: S71. Add a domain adaptation layer to the optimal model, and freeze all parameters of the backbone network in the optimal model to obtain the adaptation model structure. A domain adaptation layer is added between the feature extractor of the optimal model and the task head. The domain adaptation layer employs at least one of deformable convolutional layers or affine transformation layers: the deformable convolutional layer adaptively adjusts the sampling position by learning offsets to adapt to changes in cell morphology under different imaging conditions; the affine transformation layer transforms the feature map using a learnable scaling parameter γ and translation parameter β.

[0113] The number of parameters in this adaptation layer is less than 1% of the total parameters of the model, ensuring a lightweight and fast adaptation process. With this configuration, the backbone network remains fixed, and only the adaptation layer participates in parameter updates, thereby quickly adapting to new imaging conditions while retaining the original feature extraction capabilities.

[0114] S72. Construct a loss function based on the adapted model structure; The loss function includes the task loss calculated on the source domain labeled data, and the multi-kernel maximum mean difference loss that minimizes the difference in feature distributions between the source and target domains. Calculating the task loss only on the source domain labeled data ensures that the model's performance on the original task does not degrade. Depending on the task type, segmentation tasks use cross-entropy loss or Dice loss, while detection tasks use the sum of classification and regression losses.

[0115] The multi-kernel maximum mean difference loss uses the multi-kernel maximum mean difference (MK-MMD) as the domain alignment loss to calculate the feature distribution difference between the source domain and the target domain in the adaptation layer output space.

[0116] S73. Using the adapted model structure as the initial state and the loss function as the optimization objective, perform lightweight training, automatically align the feature distribution through adversarial training or style transfer techniques, and output the trained adapted layer parameters. The training parameters are configured with lightweight parameters: 200 to 500 iterations, 1e-4 learning rate, and Adam optimizer.

[0117] There are two specific ways to implement feature alignment: Method 1 (Adversarial Training): Employing a Gradient Reversal Layer (GRL). A domain discriminator is added after the feature extractor. The gradient reversal layer prevents the domain discriminator from distinguishing the source of the feature extractor's output, thus achieving feature distribution alignment. During training, the gradient of the domain discriminator's loss is inverted when it backpropagates through the GRL, allowing the feature extractor to learn domain-invariant feature representations.

[0118] Method 2 (Style Transfer): A CycleGAN or Adaptive Instance Normalization (AdaIN) is used to convert the target domain image into the source domain style while preserving cell structure and morphology. After training, the image is then input into the main model for inference.

[0119] Both methods mentioned above prioritize adversarial training by default. When a significant style shift is detected (e.g., FID > 50) and the target domain image is unlabeled, they automatically switch to a recurrent generative adversarial network for unsupervised style alignment. The scenarios suitable for these two methods in specific implementations are shown in Table 3 below: Table 3

[0120] Dynamic evaluation and early stopping: Every 50 iterations, the average confidence score of the model predictions is calculated on the target domain validation set as a surrogate metric (precise accuracy cannot be calculated because the target domain is unlabeled). If the average confidence score no longer improves in three consecutive evaluations (i.e., 150 consecutive iterations), an early stopping mechanism is triggered to terminate subsequent iterations and avoid overfitting and invalid computation.

[0121] S74. Integrate the trained adaptation layer parameters into the adaptation model structure to generate the adaptation model; S75. Replace the optimal model with the adapted model, and proceed to step S60.

[0122] The following specific examples further illustrate this embodiment: User B moved from lab A to lab B, and switched microscopes from Leica to Nikon, resulting in significant differences in lighting conditions. This caused the distribution differences between the acquired cell images and the original training set to exceed a preset difference threshold. The system automatically initiated a lightweight domain adaptation process, requiring no new annotations, and improved the IoU of the segmentation of the initially acquired Nikon images from 0.45 (original model) to 0.82. Subsequently, the system actively selected 5 cross-batch difference samples through multi-level learning for the user to quickly annotate. After a second incremental training, the IoU increased to 0.93, with the total annotation time less than 10 minutes.

[0123] Accordingly, according to embodiments of the present invention, the present invention also provides a computer device, a readable storage medium, and a computer program product.

[0124] Figure 2 This is a schematic diagram of the structure of a computer device 62 provided in an embodiment of the present invention. Figure 2 A block diagram of an exemplary computer device 62 suitable for implementing embodiments of the present invention is shown. Figure 2 The computer device 62 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0125] like Figure 2 As shown, computer device 62 is represented in the form of a general-purpose computing device. Computer device 62 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0126] The components of computer device 62 may include, but are not limited to: one or more processors or processing units 66, system memory 78, and bus 68 connecting different system components (including system memory 78 and processing unit 66).

[0127] Bus 68 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. Examples of these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0128] Computer device 62 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 62, including volatile and non-volatile media, removable and non-removable media.

[0129] System memory 78 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 80 and / or cache memory 82. Computer device 62 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 84 may be used to read and write non-removable, non-volatile magnetic media (…). Figure 2 Not shown; usually referred to as a "hard drive"). Although Figure 2 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 68 via one or more data media interfaces. System memory 78 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0130] A program / utility 90 having a set (at least one) of program modules 92 may be stored, for example, in system memory 78. Such program modules 92 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 92 typically perform the functions and / or methods described in the embodiments of the present invention.

[0131] Computer device 62 can also communicate with one or more external devices 64 (e.g., keyboard, pointing device, display 74, etc.), and with one or more devices that enable a user to interact with computer device 62, and / or with any device that enables computer device 62 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via input / output (I / O) interface 72. Furthermore, computer device 62 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 70. As shown, network adapter 70 communicates with other modules of computer device 62 via bus 68. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with computer device 62, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0132] The processing unit 66 executes various functional applications and data processing by running programs stored in the system memory 78, such as implementing the intelligent cell image analysis method with multi-dimensional adaptive evolution capabilities provided in the embodiments of the present invention.

[0133] This invention also provides a non-transitory computer-readable storage medium storing computer instructions, on which a computer program is stored. When the program is executed by a processor, it implements the intelligent cell image analysis method with multi-dimensional adaptive evolution capability provided in all embodiments of this invention.

[0134] The computer storage medium of this invention can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0135] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0136] The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. The computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0137] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described intelligent cell image analysis method with multi-dimensional adaptive evolution capabilities.

[0138] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0139] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A cell image intelligent analysis method with multi-dimensional adaptive evolutionary capability, characterized in that, include: S10. Obtain at least one cell image; S20. Determine the analysis task based on the imaging mode and number of channels of the cell image, and select the optimal model based on the analysis task; S30. Using the optimal model, reason about the cell image to obtain a prediction result, and extract or calculate the performance index for monitoring from the prediction result; S40. Determine whether the performance index exceeds the preset normal range, or whether the number of newly added annotations reaches the preset threshold, or whether the time interval between the optimal model and the last training reaches the preset period. If so, then S50. Based on uncertainty, sample diversity, and marginal cases, select high-value images from unlabeled cell images, obtain weak annotations for the high-value images, generate dense pseudo-labels, and output labeled images. S60. Start incremental training to update the optimal model, and return to step S30. The mixed training dataset for incremental training includes the labeled images.

2. The intelligent cell image analysis method with multi-dimensional adaptive evolution capability according to claim 1, characterized in that: Step S10 further includes: detecting the quality of the cell image, and if the quality of the cell image is unqualified, removing or replacing the cell image.

3. The intelligent cell image analysis method with multi-dimensional adaptive evolutionary capability according to claim 1, characterized in that, The number of newly added annotations in step S40 includes at least one of the following: the number of cell images annotated by the user, the percentage of the area of ​​annotated cell pixel regions in a single cell image, the number of cell images with dense pseudo-labels, and the number of annotated cell images imported from outside.

4. The intelligent cell image analysis method with multi-dimensional adaptive evolutionary capability according to claim 1, characterized in that, Following step S20, the following step is also included: The optimal model is evaluated on the validation set according to the preset evaluation cycle. When multiple consecutive evaluations show that the key indicators decrease one after another and the cumulative decrease exceeds the preset decay threshold, the process proceeds to step S50.

5. The intelligent cell image analysis method with multi-dimensional adaptive evolution capability according to claim 1, characterized in that, Between steps S50 and S60, the following is included: Determine whether the distribution difference between the cell image and the training set of the optimal model exceeds a preset difference threshold; if yes, start the lightweight domain adaptation process; if no, proceed to step S60.

6. The intelligent cell image analysis method with multi-dimensional adaptive evolutionary capability according to claim 5, characterized in that, The lightweight domain adaptation process includes the following steps: S71. Add a domain adaptation layer to the optimal model, and freeze all parameters of the backbone network in the optimal model to obtain the adaptation model structure. S72. Construct a loss function based on the adapted model structure; S73. Using the adapted model structure as the initial state and the loss function as the optimization objective, perform lightweight training, automatically align the feature distribution through adversarial training or style transfer techniques, and output the trained adapted layer parameters. S74. Integrate the trained adaptation layer parameters into the adaptation model structure to generate the adaptation model; S75. Replace the optimal model with the adapted model.

7. The intelligent cell image analysis method with multi-dimensional adaptive evolutionary capability according to claim 1, characterized in that, Step S60 includes: S61. Freeze the backbone network of the optimal model to obtain a trainable network structure; S62. Load all the weights of the optimal model into the network structure to obtain the initial model; S63. Representative old samples are sampled proportionally from historical data and added to the hybrid training dataset through a memory playback mechanism. S64. Calculate the Fisher information matrix of historical important parameters and construct the EWC regularization loss function; S65. The initial model is incrementally fine-tuned using the hybrid training dataset to obtain the fine-tuned model; S66. Evaluate the performance of the fine-tuned model on an independent validation set, and determine whether to replace the optimal model with the fine-tuned model.

8. The intelligent cell image analysis method with multi-dimensional adaptive evolutionary capability according to claim 7, characterized in that, If the optimal model is replaced with the fine-tuned model, then after step S66, the following steps are included: saving the fine-tuned model as a new version, with the version number automatically incrementing according to a preset rule, and automatically generating a version record. The version record includes at least one of the following: version number, training time, triggering reason, training data volume, training duration, performance comparison between the old and new models, performance improvement, forgetting rate, training parameters, and model size.

9. A computer device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the cell image intelligent analysis method with multi-dimensional adaptive evolution capability as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the cell image intelligent analysis method with multi-dimensional adaptive evolution capability as described in any one of claims 1 to 8.

Images