Cell microscopic image automatic classification method based on deep learning

Abstract

The invention provides a cell microscopic image automatic classification method based on deep learning, relates to the technical field of deep learning and pattern recognition, and aims to extract deep features through self-supervised characterization learning and mask modeling in the step S2, so that the dependence on manual annotation is remarkably reduced. By utilizing the uncertainty index Kzs and B1 and B2 threshold grading evaluation in the step S4, precise screening of low-confidence and rare pathological cells is realized, and manual labeling deviation is relieved; the recognition gain Gr and an incremental learning strategy are combined, and the recognition depth of the rare type is improved with few labels; meanwhile, through the step S1 equipment identification and the step S3 cross-domain style standardization, the color shift caused by the hardware model and the dyeing process is eliminated; model evolution is dynamically triggered by using the stability coefficient Sc and the efficiency index Xz, the robustness and generalization ability of the system in a heterogeneous data environment are enhanced, and the diagnosis and treatment requirements of complex clinical scenes are effectively met.

Description

Technical Field

[0001] This invention relates to the field of deep learning and pattern recognition technology, specifically to a method for automatic classification of cell microscopic images based on deep learning. Background Technology

[0002] Traditional methods heavily rely on manual microscopic examination by pathologists, which is not only time-consuming and labor-intensive but also highly subjective, susceptible to fatigue and experience differences, and ill-suited for handling the massive amounts of data generated by modern high-throughput screening. Early computer-aided diagnostic techniques, while achieving automation, largely depended on manually designed morphological features (such as perimeter and area), often exhibiting insufficient generalization ability when dealing with images of overlapping cells, varied morphologies, and complex backgrounds. In recent years, with the rapid development of artificial intelligence, deep learning-based image classification methods have emerged. These methods, by constructing deep neural network models, can automatically extract multi-dimensional deep semantic features from massive images, significantly improving the accuracy and robustness of cell classification. This technology not only reduces labor costs but also provides strong technical support for achieving objectivity and standardization in medical diagnosis, and has become a research hotspot in the field of bioinformatics.

[0003] In existing technologies, training deep learning models requires tens of thousands of cell images precisely annotated by senior pathologists. However, the manual annotation process is extremely time-consuming and labor-intensive, and there are subjective differences in cognition among different experts, resulting in extremely high costs for obtaining standardized, high-precision "gold standard" data. This greatly limits the learning effect of the model on rare cell types or specific clinical samples, leading to the technical drawback of being highly dependent on large-scale and high-quality manually annotated datasets. Based on the aforementioned deep reliance on specific labeled data, existing technologies exhibit another drawback: insufficient cross-domain generalization ability and robustness. Since models are often trained on datasets collected under specific experimental conditions (such as specific staining processes, light intensities, or microscope models), subtle background noise or color shifts in the images can cause the extracted features to fail once the application scenario changes. This data-sensitive generalization bottleneck makes automatic classification systems highly susceptible to significant drops in classification accuracy when faced with heterogeneous data from different laboratories, making it difficult to meet the complex and ever-changing needs of clinical applications. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic classification method for cell microscopic images based on deep learning, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based automatic classification method for cell microscopic images includes the following steps: Step S1: Collect raw cell microscopic images from different sources, perform image standardization preprocessing, and construct a label-free cell image library containing various experimental conditions; Step S2: Perform self-supervised representation learning to automatically extract deep morphological and texture features of cells from an unlabeled cell image library and construct a pre-trained morphological representation model to reduce reliance on manually labeled data; Step S3: Perform cross-domain style normalization processing, calculate the color offset and noise distribution between different image sources, and map the heterogeneous image features to a unified metric space; Step S4: Perform real-time classification monitoring based on active learning, calculate the cell classification uncertainty index Kzs during the classification process, and preset the first threshold B1 and the second threshold B2 to grade and evaluate the uncertainty index Kzs and generate manual intervention instructions. Step S5: Perform classification efficiency evaluation and dynamic optimization, collect classification process data to construct efficiency evaluation set, calculate cell classification efficiency index Xz, and generate model evolution strategy based on evaluation results.

[0006] Furthermore, in step S1, high-resolution images of cells at different growth stages and pathological states are acquired through multi-mode imaging technology; then, device feature information in the images is identified, including light source color temperature and magnification, and adaptive histogram equalization technology is used to eliminate the influence of uneven illumination.

[0007] Further, step S2 specifically includes: applying masked image modeling technology to randomly mask a portion of the cell image, using a deep neural network to predict the pixel distribution of the masked portion, thereby forcing the model to learn the intrinsic structural relationships between the cell nucleus, cytoplasm, and cell membrane; extracting morphological parameters of the cell, including nuclear-cytoplasmic ratio, nuclear membrane roughness, and chromatin aggregation, and converting them into high-dimensional morphological feature vectors.

[0008] Furthermore, in step S3, style features of different laboratory samples are extracted using generative adversarial training; a cross-domain transformation function is constructed by calculating the color histogram difference and texture contrast deviation between the source and target domains to achieve standardized alignment of cell image features under different staining processes.

[0009] Further, in step S4, the logic for calculating the cell classification uncertainty index Kzs is as follows: the probability distribution entropy value of the predicted category is obtained through the softmax layer, the variance of the model prediction is obtained by combining Monte Carlo sampling, and the distance between the sample to be classified and the known category center in the feature space is calculated. Finally, the prediction entropy value, prediction variance and spatial distance are weighted and summed to obtain the uncertainty index Kzs that reflects the difficulty of sample identification.

[0010] Furthermore, in step S4, the first threshold B1 is greater than the second threshold B2, and the criteria for classifying and evaluating the uncertainty index Kzs are as follows: When the uncertainty index Kzs ≤ the second threshold B2, the classification result is determined to have high confidence, and the classification result is automatically output. When the second threshold B2 < uncertainty index Kzs ≤ first threshold B1, it is judged as a low confidence sample, and an auxiliary verification instruction is generated to prompt manual confirmation of key feature points; When the uncertainty index Kzs > the first threshold B1, it is determined to be a suspected rare pathological cell or excessive noise interference, and an expert intervention instruction is generated for precise annotation by pathology experts.

[0011] Furthermore, step S4 also includes a human-computer interaction guidance step, which generates a heat map on the microscopic image through class activation mapping technology, and displays in real time the model’s attention weight to local cell features, including nucleolar atypia and cytoplasmic vacuolation, when making classification decisions, so as to provide physicians with clinical traceability analysis.

[0012] Furthermore, in step S5, the calculation logic of the cell classification efficiency index Xz is as follows: obtain the classification accuracy coefficient, the recall consistency coefficient of cross-domain testing, the sensitivity coefficient of rare samples, and the time cost of model running in the classification dataset, and calculate the total efficiency index Xz through a multi-factor weighted formula.

[0013] Furthermore, in step S5, the specific evaluation content of the classification efficiency index Xz includes: By extracting feature vectors from images of the same type of cells collected by different types of microscopes, calculating their distribution offset distance in the feature space, and converting them into a cross-hardware feature expression stability coefficient Sc; under the condition of only containing a preset number of labeled samples, comparing the improvement in the recognition accuracy of the current model and the benchmark model for rare cell types, and converting it into a recognition gain coefficient Gr; by weighting the stability coefficient Sc and the recognition gain coefficient Gr with preset weights, the cell classification efficiency index Xz is calculated. A preset classification performance benchmark threshold Qz is used as a critical indicator to determine whether the system needs to initiate an optimization strategy; If the cell classification efficiency index Xz ≥ the classification efficiency benchmark threshold Qz, then the current method is deemed to have sufficient robustness and generalization ability when facing heterogeneous hardware environments and rare samples, and the current model parameters are maintained. If the cell classification efficiency index Xz is less than the classification efficiency benchmark threshold Qz, the current method is determined to have failed to meet the standard in feature extraction stability or weakly supervised learning efficiency, triggering the model evolution strategy described in step S5.

[0014] Furthermore, the model evolution strategy generated in step S5 includes: When Xz falls below a preset threshold, an incremental learning mechanism is automatically triggered, adding high-value samples verified by experts to the training set for fine-tuning; adjusting the loss function of the deep neural network and increasing the penalty term for cross-domain noise; and using knowledge distillation technology to transfer the discriminative power of large and complex models to lightweight models to improve the operational efficiency of actual clinical deployment.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Through the self-supervised representation learning mechanism described in step S2, deep morphological and texture features are automatically extracted from the unlabeled cell image library constructed in S1 using masked image modeling technology, fundamentally reducing the system's strong dependence on large-scale expert-annotated data. Combined with the cell classification uncertainty index Kzs calculated in step S4, and with the preset first threshold B1 and second threshold B2 for hierarchical evaluation, accurate screening and manual intervention guidance for low-confidence samples and suspected rare pathological cells are achieved, greatly mitigating the negative impact of time-consuming and labor-intensive manual annotation and differences in expert subjective perception. Using the recognition gain coefficient Gr defined in step S9, the recognition performance of rare cell types can be evaluated under the condition of a very small number of labeled samples. Combined with the incremental learning strategy in step S10, the learning depth of the model on specific clinical samples is significantly improved while ensuring high classification accuracy, solving the defect of existing technologies that limit model performance due to high annotation costs.

[0016] By using device feature recognition and standardized preprocessing in step S1, combined with cross-domain style normalization and cross-domain transformation functions in step S3, the differences in color shift and noise distribution caused by different microscope models, light source color temperatures, magnification, and staining processes are effectively eliminated, mapping heterogeneous image features to a unified metric space. Using the cross-hardware feature expression stability coefficient Sc and the total classification efficiency index Xz calculated in step S9, the consistency of feature expression of the model when facing new acquisition devices can be quantitatively monitored. By comparing the quantitative comparison with the classification efficiency benchmark threshold Qz, the model evolution strategy and loss function penalty mechanism described in step S10 are dynamically triggered. This method significantly enhances the robustness of the automatic classification system when facing heterogeneous data from different laboratory sources, effectively overcomes the generalization bottleneck caused by environmental sensitivity, and meets the stringent requirements for cross-domain generalization ability and adaptability to complex scenarios in clinical applications. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Example 1: Please see Figure 1 This invention provides a technical solution: an automatic classification method for cell microscopic images based on deep learning, the specific steps of which include: Step S1: Collect raw cell microscopic images from different sources, perform image standardization preprocessing, and construct a label-free cell image library containing various experimental conditions; Step S2: Perform self-supervised representation learning to automatically extract deep morphological and texture features of cells from an unlabeled cell image library and construct a pre-trained morphological representation model to reduce reliance on manually labeled data; Step S3: Perform cross-domain style normalization processing, calculate the color offset and noise distribution between different image sources, and map the heterogeneous image features to a unified metric space; Step S4: Perform real-time classification monitoring based on active learning, calculate the cell classification uncertainty index Kzs during the classification process, and preset the first threshold B1 and the second threshold B2 to grade and evaluate the uncertainty index Kzs and generate manual intervention instructions. Step S5: Perform classification efficiency evaluation and dynamic optimization, collect classification process data to construct efficiency evaluation set, calculate cell classification efficiency index Xz, and generate model evolution strategy based on evaluation results.

[0021] In this embodiment, by performing image standardization preprocessing and constructing a label-free cell image library through step S1, high-resolution image data from different sources can be effectively integrated and the effects of uneven illumination can be eliminated by utilizing device feature information. The self-supervised representation learning in step S2 is performed and high-dimensional morphological feature vectors such as nuclear-cytoplasmic ratio, nuclear membrane roughness and chromatin aggregation are extracted using masked image modeling technology. This significantly reduces the dependence on manual annotation while enhancing the model's semantic perception of the cell's internal structure. By utilizing the cross-domain style normalization process and the construction of the cross-domain transformation function in step S3, the differences in staining offset and noise distribution between different laboratories can be accurately quantified and eliminated, ensuring the standardized alignment and cross-domain generalization of heterogeneous image features in a unified metric space. The cell classification uncertainty index Kzs, which integrates predicted entropy, predicted variance and spatial distance, is calculated in real time through the active learning mechanism in step S4. The grading evaluation logic is executed in conjunction with the preset first threshold B1 and second threshold B2, which realizes the precise scheduling of automated output, manual confirmation and expert intervention instructions, and significantly improves the security of complex clinical sample identification. Step S5 calculates the cell classification efficiency index Xz, which includes the cross-hardware feature expression stability coefficient Sc and the recognition gain coefficient Gr. This quantitatively determines whether the system's robustness meets the standard. Based on the quantitative comparison between Xz and the classification efficiency benchmark threshold Qz, the system dynamically triggers incremental learning and other model evolution strategies, ensuring continuous closed-loop optimization of the system's performance in long-term clinical applications.

[0022] Example 2: In step S1, high-resolution images of cells at different growth stages and pathological states are acquired through multi-mode imaging technology; then, device feature information in the images is identified, including light source color temperature and magnification, and adaptive histogram equalization technology is used to eliminate the influence of uneven illumination.

[0023] Step S2 specifically includes: applying masked image modeling technology to randomly mask a portion of the cell image, using a deep neural network to predict the pixel distribution of the masked portion, thereby forcing the model to learn the intrinsic structural relationships between the cell nucleus, cytoplasm, and cell membrane; extracting morphological parameters of the cell, including nuclear-cytoplasmic ratio, nuclear membrane roughness, and chromatin aggregation, and converting them into high-dimensional morphological feature vectors.

[0024] In step S3, style features of different laboratory samples are extracted using generative adversarial training; a cross-domain transformation function is constructed by calculating the color histogram difference and texture contrast deviation between the source and target domains to achieve standardized alignment of cell image features under different staining processes.

[0025] In step S4, the logic for calculating the cell classification uncertainty index Kzs is as follows: the probability distribution entropy value of the predicted category is obtained through the softmax layer, the variance of the model prediction is obtained by combining Monte Carlo sampling, and the distance between the sample to be classified and the known category center in the feature space is calculated. Finally, the prediction entropy value, prediction variance and spatial distance are weighted and summed to obtain the uncertainty index Kzs that reflects the difficulty of sample identification.

[0026] In step S4, the first threshold B1 is greater than the second threshold B2, and the criteria for classifying and evaluating the uncertainty index Kzs are as follows: When the uncertainty index Kzs ≤ the second threshold B2, the classification result is determined to have high confidence, and the classification result is automatically output. When the second threshold B2 < uncertainty index Kzs ≤ first threshold B1, it is judged as a low confidence sample, and an auxiliary verification instruction is generated to prompt manual confirmation of key feature points; When the uncertainty index Kzs > the first threshold B1, it is determined to be a suspected rare pathological cell or excessive noise interference, and an expert intervention instruction is generated for precise annotation by pathology experts.

[0027] Step S4 also includes a human-computer interaction guidance step, which uses class activation mapping technology to generate a heat map on the microscopic image to display in real time the model’s attention weight to local cell features, including nucleolar atypia and cytoplasmic vacuolation, when making classification decisions, so as to provide physicians with clinical traceability analysis.

[0028] In step S5, the calculation logic of the cell classification efficiency index Xz is as follows: obtain the classification accuracy coefficient, the recall consistency coefficient of cross-domain testing, the sensitivity coefficient of rare samples, and the time cost of model running in the classification dataset, and calculate the total efficiency index Xz through a multi-factor weighted formula.

[0029] In step S5, the specific evaluation content of the classification efficiency index Xz includes: By extracting feature vectors from images of the same type of cells collected by different types of microscopes, calculating their distribution offset distance in the feature space, and converting them into a cross-hardware feature expression stability coefficient Sc; under the condition of only containing a preset number of labeled samples, comparing the improvement in the recognition accuracy of the current model and the benchmark model for rare cell types, and converting it into a recognition gain coefficient Gr; by weighting the stability coefficient Sc and the recognition gain coefficient Gr with preset weights, the cell classification efficiency index Xz is calculated. A preset classification performance benchmark threshold Qz is used as a critical indicator to determine whether the system needs to initiate an optimization strategy; If the cell classification efficiency index Xz ≥ the classification efficiency benchmark threshold Qz, then the current method is deemed to have sufficient robustness and generalization ability when facing heterogeneous hardware environments and rare samples, and the current model parameters are maintained. If the cell classification efficiency index Xz is less than the classification efficiency benchmark threshold Qz, the current method is determined to have failed to meet the standard in feature extraction stability or weakly supervised learning efficiency, triggering the model evolution strategy described in step S5.

[0030] The model evolution strategy generated in step S5 includes: When Xz falls below a preset threshold, an incremental learning mechanism is automatically triggered, adding high-value samples verified by experts to the training set for fine-tuning; adjusting the loss function of the deep neural network and increasing the penalty term for cross-domain noise; and using knowledge distillation technology to transfer the discriminative power of large and complex models to lightweight models to improve the operational efficiency of actual clinical deployment.

[0031] In this embodiment, the cell classification uncertainty index Kzs is set to a value distribution in the range [0, +∞), which quantitatively characterizes the risk level of the model's identification of the current cell sample to be classified. When Kzs approaches its theoretical minimum, the sample to be classified is extremely close to the known category cluster center constructed by steps S2 and S3 in the feature space, and the model's predicted probability distribution exhibits an extremely low entropy state, indicating that the technical state is in the high-confidence automated processing zone, and the system can achieve high-precision automatic classification. When Kzs approaches its maximum value, it means that there is a serious distribution shift in the sample features, manifested as a sharp increase in predicted entropy or uncontrolled Monte Carlo sampling variance, indicating that the technical state is in the cognitively abnormal zone, usually corresponding to rare pathological morphology or unidentified heterogeneous interference. At this time, the system must lock the classification output and trigger the highest level of expert intervention.

[0032] The calculation of the Kzs index is driven by three key input parameters: the entropy of the probability distribution of the predicted class, the model prediction variance, and the sample distance in the feature space. The entropy is positively correlated with Kzs, ensuring the system's immediate perception of ambiguity in classification decisions; the prediction variance obtained through Monte Carlo sampling captures the cognitive uncertainty of model weights, identifying the instability of feature representations when facing noisy, heterogeneous images; and the feature space distance is non-linearly positively correlated with Kzs, ensuring the system's sensitivity to rare morphological samples. This multi-factor weighting logic solves the "blind confidence" problem that occurs in existing technologies when models face morphologically unrestricted samples, laying a solid mathematical foundation for subsequent hierarchical evaluation.

[0033] In the practical application of output interval division, this invention employs an objective threshold derivation methodology, mapping the continuously changing Kzs exponent to an effectiveness index space composed of the combined cost factor Cf of false alarms and missed diagnoses. Analysis reveals a typical S-shaped logistic regression curve relationship between Kzs and Cf; that is, Cf grows slowly within a low uncertainty interval, but once a specific critical point is crossed, the risk cost increases exponentially. This causal relationship model based on data patterns proves that the interval division is not arbitrarily set, but rather rooted in the nonlinear dynamic process of the model's classification error rate evolving with feature shift.

[0034] To accurately define the interval boundaries, this invention utilizes second-order derivative analysis to search for the curvature maxima of the Kzs-Cf correlation curve, thereby mathematically identifying two key inflection points as the first threshold B1 and the second threshold B2. The smaller inflection point B2 is defined as the performance saturation boundary for automated processing, representing the limit at which the model can maintain maximum efficiency without intervention; the larger inflection point B1 is defined as the critical instability point of the model's knowledge boundary, representing the threshold at which the model's discriminative ability fails. This method of extracting thresholds from the correlation function using mathematical means ensures that the hierarchical evaluation criteria have rigorous logical justification and physical meaning.

[0035] The three finally determined practical application intervals and intervention strategies achieve an accurate cost-benefit mapping: in the high-confidence automation interval (Kzs ≤ B2), the system performs fully automated and rapid classification to maximize the efficiency of clinical examinations; in the low-confidence assisted verification interval (B2 < Kzs ≤ B1), the system initiates the human-computer interaction guidance described in step S4, displays a heat map through class activation mapping technology for doctors to check, so as to hedge the diagnostic risk at a relatively low labor cost; in the extremely high-risk expert intervention interval (Kzs > B1), the system forcibly introduces expert annotations, which not only avoids the risk of serious misdiagnosis, but also reserves high-value morphological samples for the incremental learning mechanism described above, realizing the continuous evolution of the model classification performance.

[0036] It should be noted that: all calculation formulas in this application document adopt regression analysis including but not limited to machine learning algorithms to deeply analyze the relevant parameters collected, identify their natural trends and interrelationships. Using professional software, such as the Scikit-learn library of Python or the R language, a mathematical model matching the data is automatically generated. Then, the performance of the model is objectively evaluated through methods such as cross-validation, and combined with continuous feedback and optimization, to ensure that the created formula truly reflects the internal laws of the data, thereby ensuring its effectiveness and accuracy. In all calculation formulas of this application, the parameters in each formula are processed by dimensionless normalization within a consistent range to ensure that different physical quantities are compared on the same scale; the dimensionless technical means include but not limited to min-max normalization, Z-Score standardization; The technical solution of the present invention can be embodied in the form of a software product in essence or the part that contributes to the prior art. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disc of a computer, etc., including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0037] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatic classification of cell microscopic images based on deep learning, characterized in that, The specific steps include: Step S1: Collect raw cell microscopic images from different sources, perform image standardization preprocessing, and construct a label-free cell image library containing various experimental conditions; Step S2: Perform self-supervised representation learning to automatically extract deep morphological and texture features of cells from an unlabeled cell image library and construct a pre-trained morphological representation model to reduce reliance on manually labeled data; Step S3: Perform cross-domain style normalization processing, calculate the color offset and noise distribution between different image sources, and map the heterogeneous image features to a unified metric space; Step S4: Perform real-time classification monitoring based on active learning, calculate the cell classification uncertainty index Kzs during the classification process, and preset the first threshold B1 and the second threshold B2 to grade and evaluate the uncertainty index Kzs and generate manual intervention instructions. Step S5: Perform classification efficiency evaluation and dynamic optimization, collect classification process data to construct efficiency evaluation set, calculate cell classification efficiency index Xz, and generate model evolution strategy based on evaluation results.

2. The method for automatic classification of cell microscopic images based on deep learning according to claim 1, characterized in that: In step S1, high-resolution images of cells at different growth stages and pathological states are acquired through multi-mode imaging technology; then, device feature information in the images is identified, including light source color temperature and magnification, and adaptive histogram equalization technology is used to eliminate the influence of uneven illumination.

3. The automatic classification method for cell microscopic images based on deep learning according to claim 1, characterized in that: Step S2 specifically includes: applying masked image modeling technology to randomly mask a portion of the cell image, using a deep neural network to predict the pixel distribution of the masked portion, thereby forcing the model to learn the intrinsic structural relationships between the cell nucleus, cytoplasm, and cell membrane; extracting morphological parameters of the cell, including nuclear-cytoplasmic ratio, nuclear membrane roughness, and chromatin aggregation, and converting them into high-dimensional morphological feature vectors.

4. The automatic classification method for cell microscopic images based on deep learning according to claim 1, characterized in that: In step S3, style features of different laboratory samples are extracted using generative adversarial training; a cross-domain transformation function is constructed by calculating the color histogram difference and texture contrast deviation between the source and target domains to achieve standardized alignment of cell image features under different staining processes.

5. The method for automatic classification of cell microscopic images based on deep learning according to claim 1, characterized in that: In step S4, the logic for calculating the cell classification uncertainty index Kzs is as follows: the probability distribution entropy value of the predicted category is obtained through the softmax layer, the variance of the model prediction is obtained by combining Monte Carlo sampling, and the distance between the sample to be classified and the known category center in the feature space is calculated. Finally, the prediction entropy value, prediction variance and spatial distance are weighted and summed to obtain the uncertainty index Kzs that reflects the difficulty of sample identification.

6. The method for automatic classification of cell microscopic images based on deep learning according to claim 1, characterized in that: In step S4, the first threshold B1 is greater than the second threshold B2, and the criteria for classifying and evaluating the uncertainty index Kzs are as follows: When the uncertainty index Kzs ≤ the second threshold B2, the classification result is determined to have high confidence, and the classification result is automatically output. When the second threshold B2 < uncertainty index Kzs ≤ first threshold B1, it is judged as a low confidence sample, and an auxiliary verification instruction is generated to prompt manual confirmation of key feature points; When the uncertainty index Kzs > the first threshold B1, it is determined to be a suspected rare pathological cell or excessive noise interference, and an expert intervention instruction is generated for precise annotation by pathology experts.

7. The method for automatic classification of cell microscopic images based on deep learning according to claim 1, characterized in that: Step S4 also includes a human-computer interaction guidance step, which uses class activation mapping technology to generate a heat map on the microscopic image to display in real time the model’s attention weight to local cell features, including nucleolar atypia and cytoplasmic vacuolation, when making classification decisions, so as to provide physicians with clinical traceability analysis.

8. The method for automatic classification of cell microscopic images based on deep learning according to claim 1, characterized in that: In step S5, the calculation logic of the cell classification efficiency index Xz is as follows: obtain the classification accuracy coefficient, the recall consistency coefficient of cross-domain testing, the sensitivity coefficient of rare samples, and the time cost of model running in the classification dataset, and calculate the total efficiency index Xz through a multi-factor weighted formula.

9. The method for automatic classification of cell microscopic images based on deep learning according to claim 8, characterized in that: In step S5, the specific evaluation content of the classification efficiency index Xz includes: By extracting feature vectors from images of the same type of cells collected by different types of microscopes, calculating their distribution offset distance in the feature space, and converting them into a cross-hardware feature expression stability coefficient Sc; under the condition of only containing a preset number of labeled samples, comparing the improvement in the recognition accuracy of the current model and the benchmark model for rare cell types, and converting it into a recognition gain coefficient Gr; by weighting the stability coefficient Sc and the recognition gain coefficient Gr with preset weights, the cell classification efficiency index Xz is calculated. A preset classification performance benchmark threshold Qz is used as a critical indicator to determine whether the system needs to initiate an optimization strategy; If the cell classification efficiency index Xz ≥ the classification efficiency benchmark threshold Qz, then the current method is deemed to have sufficient robustness and generalization ability when facing heterogeneous hardware environments and rare samples, and the current model parameters are maintained. If the cell classification efficiency index Xz is less than the classification efficiency benchmark threshold Qz, the current method is determined to have failed to meet the standard in feature extraction stability or weakly supervised learning efficiency, triggering the model evolution strategy described in step S5.

10. The method for automatic classification of cell microscopic images based on deep learning according to claim 8, characterized in that: The model evolution strategy generated in step S5 includes: When Xz falls below a preset threshold, an incremental learning mechanism is automatically triggered, adding high-value samples verified by experts to the training set for fine-tuning; adjusting the loss function of the deep neural network and increasing the penalty term for cross-domain noise; and using knowledge distillation technology to transfer the discriminative power of large and complex models to lightweight models to improve the operational efficiency of actual clinical deployment.

Images