A method for microscopic imaging-based pathological feature recognition and negative exclusion

Abstract

The application discloses a pathological feature recognition and negative exclusion method based on microscopic imaging, comprising the following steps: S1, collecting pathological section images and digitally generating original microscopic image data; S2, preprocessing the original microscopic image; S3, constructing a pathological image recognition network that fuses a transformation coder and a gated dynamic receptive field mechanism, and outputting a pathological feature vector; S4, modeling the context through attention guidance and category perception decoder, and outputting an image classification result; S5, constructing a discriminant boundary separation model based on positive and negative sample embedding, and judging the negative exclusion; S6, combining uncertainty and boundary distance for confidence weighted evaluation, setting a dynamic threshold, and screening out low-confidence samples; S7, encoding the classification result and the negative label into structured data, and sending to a diagnosis assistance system. The application fuses multi-scale modeling and negative screening mechanism, realizes intelligent recognition of pathological images and reliable diagnosis output.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and computational pathology, and in particular to a method for identifying pathological features and excluding negative results based on microscopic imaging. Background Technology

[0002] Against the backdrop of the rapid development of digital pathology, microscopic image analysis has gradually become an important tool for assisting pathological diagnosis, playing a crucial role, especially in tasks such as tumor screening, tissue classification, and morphological feature recognition. Traditional pathological diagnosis relies on manual observation of microscopic sections, which is highly subjective, inefficient, and easily affected by factors such as physician fatigue and experience differences when dealing with large-scale image data, making it difficult to guarantee the accuracy and consistency of diagnostic results. In recent years, with the widespread application of deep learning in the field of medical imaging, more and more studies have attempted to use convolutional neural networks (CNNs) for automatic feature extraction and classification of microscopic images. However, existing methods still have significant limitations in handling problems such as multi-scale structures, non-uniform staining, and complex tissue boundaries in pathological images.

[0003] Existing technologies typically use CNN structures with fixed receptive fields to extract features from pathological images, making it difficult to simultaneously capture local morphological features at the cellular level and global spatial relationships at the tissue level. This leads to instability in model performance when identifying lesion boundaries or minute lesions. Furthermore, existing methods often suffer from insufficient reliability or misclassification of negative samples when dealing with ambiguous distributions or imbalanced sample sizes among pathological image categories, particularly in distinguishing between normal tissue and low-risk lesions, where effective boundary modeling and uncertainty control mechanisms are often lacking. In addition, most existing research focuses on improving classification accuracy, with less consideration given to how to systematically identify and exclude low-risk or negative samples to optimize the reading burden for pathologists and improve screening efficiency.

[0004] To address the aforementioned issues, this invention proposes a pathological feature recognition and negative result exclusion method based on microscopic imaging. It employs a recognition network that integrates a fusion transformer coding module and a gated dynamic receptive field mechanism, achieving collaborative modeling of multi-scale structural features and global contextual information, significantly improving the discrimination capability under complex tissue structures. Simultaneously, a discrimination boundary separation model and a confidence-weighted evaluation mechanism are constructed to effectively address the shortcomings of existing technologies in negative result exclusion and high-uncertainty sample control. Dynamic threshold control enables rapid screening of high-confidence images and review suggestions for low-confidence samples, establishing a complete closed-loop recognition process from image acquisition to structured result output, fundamentally improving the intelligent processing efficiency and diagnostic assistance capabilities of digital pathological images.

[0005] Therefore, how to provide a method for identifying pathological features and excluding negative results based on microscopic imaging is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for pathological feature recognition and negative sample exclusion based on microscopic imaging. The invention integrates a gated dynamic receptive field and a transformer coding mechanism to achieve multi-scale pathological feature extraction, introduces a discrimination boundary separation and confidence weighting strategy to improve the accuracy of negative sample exclusion, and constructs a structured output and system interface linkage process to achieve intelligent recognition, reliable screening and efficient diagnostic assistance of pathological images.

[0007] A method for identifying pathological features and excluding negative results based on microscopic imaging according to an embodiment of the present invention includes the following steps:

[0008] S1. Acquire high-resolution images of pathological tissue sections using a microscopic imaging device, and digitize the high-resolution images to obtain raw microscopic image data;

[0009] S2. Perform image preprocessing on the original microscopic image data, including noise suppression, color normalization, tissue boundary enhancement, and foreground region segmentation, to generate a processed image with clear structure.

[0010] S3. Construct a pathological image recognition network that integrates a transformer coding module and a gated dynamic receptive field mechanism. Take the processed image as input, extract multi-scale spatial semantic features through feature alignment, channel dynamic control and global context modeling, and generate pathological feature vectors.

[0011] S4. Input the pathological feature vector into the classification decoding module, use a multi-layer attention-guided feature fusion structure to perform contextual association modeling of features in different regions, and use a category-aware decoder to generate the classification judgment result of the pathological image.

[0012] S5. Construct a discrimination boundary separation model in the feature compression space. Based on the positive and negative sample embedding distribution, perform negative exclusion judgment on the preliminary classification judgment result, screen samples whose feature performance is close to the negative template and output them as negative results.

[0013] S6. Perform a confidence-weighted evaluation on the negative exclusion results. Combine the uncertainty distribution output by the discrimination boundary separation model with the class boundary distance, set a dynamic threshold, screen out low-confidence samples, and retain high-confidence images for subsequent processing.

[0014] S7. Encode the pathological classification results and negative exclusion labels according to a unified structural format, write them into a structured dataset, and send them synchronously to the diagnostic assistance system for doctors to access or for analysis by the remote consultation system.

[0015] Optionally, S1 specifically includes:

[0016] S11. Fix the stained pathological tissue sections onto a glass slide carrier and place them in a fully automated digital microscope platform to prepare for image acquisition.

[0017] S12. Control the fully automatic digital microscope platform to automatically scan the tissue section region by region according to the set scanning magnification and spatial resolution to obtain local field-of-view images;

[0018] S13. Stitch together multiple local view images in spatial order to generate a complete high-resolution slice image;

[0019] S14. Perform a unified format conversion on the high-resolution slice image to generate raw microscopic image data that meets the requirements of subsequent processing.

[0020] Optionally, S2 specifically includes:

[0021] S21. Perform image denoising processing on the original microscopic image data, and use edge-preserving filtering to remove background noise and imaging artifacts while preserving tissue structure details.

[0022] S22. Perform color normalization processing on the staining in the original microscopic image, and unify the color distribution of the image based on the target staining template to improve the consistency between images.

[0023] S23. Perform tissue structure edge enhancement operation on the boundary region of the original microscopic image, and enhance the cell and tissue contours through gradient operation to highlight local structural features;

[0024] S24. Use a foreground extraction algorithm to segment the original microscopic image, remove the background region and extract the tissue foreground region to generate a processed image with clear structure and well-defined boundaries.

[0025] Optionally, S3 specifically includes:

[0026] S31. Input the processed image obtained in step S2 into the shallow feature extraction module. The shallow feature extraction module consists of multiple standard two-dimensional convolutional layers, batch normalization layers and nonlinear activation functions, which are used to extract basic visual features of the image such as low-level edges, textures and color distribution to generate a primary feature map.

[0027] S32. Input the primary feature map into the gated dynamic receptive field module. The gated dynamic receptive field module consists of multiple parallel convolutional branches. Each branch has a different convolutional kernel size and stride parameter, which is used to extract multi-scale structural features. The gated unit adaptively allocates the weights of each branch according to the channel response value and fuses and outputs a multi-scale feature map.

[0028] S33. Input the multi-scale feature map into the feature enhancement module. The feature enhancement module includes two parts: channel attention mechanism and spatial attention mechanism. It weights and enhances high-response channels to strengthen the spatial expression ability of the lesion area and outputs the enhanced feature map.

[0029] S34. The enhanced feature map is input to the transformer encoding module. The transformer encoding module consists of multiple stacked encoder units. Each encoder unit includes a multi-head self-attention computing layer, a feedforward network and a residual connection structure. At the same time, relative position encoding is introduced to maintain the spatial semantics of the pathological structure, extract the context information between distant structures, and output the context-encoded feature map.

[0030] S35. Construct a cross-scale interactive attention module, align the scales of the encoded feature maps of different depths and establish a cross-attention mapping, calculate the similarity weighting value between scales through the feature matching function, fuse them to generate the contextual relationship between high-expression regions and low-expression regions, and output a cross-scale fused feature map.

[0031] S36. Input the cross-scale fused feature map into the structural semantic separation unit to separate the main path features for structure preservation and the auxiliary path features for semantic enhancement. The main path uses a position attention mechanism to enhance structural stability, and the auxiliary path uses a context enhancement mechanism to strengthen the semantic representation of the lesion region. Then, the features are weighted and combined through the path fusion module.

[0032] S37. Perform a dimension compression mapping operation on the fused feature map, and convert the multi-dimensional image features into a vector representation of a unified dimension through global average pooling and linear transformation to form the final pathological image feature vector and output it.

[0033] Optionally, S4 specifically includes:

[0034] S41. Receive the pathological image feature vector output in step S3, and expand the feature vector into a two-dimensional spatial feature map through a linear projection layer for subsequent structural reconstruction and classification reasoning.

[0035] S42. Input the two-dimensional spatial feature map into a multi-level classification decoding network. The multi-level classification decoding network consists of multiple stacked attention fusion decoding units. Each attention fusion decoding unit includes an attention fusion module, a convolution restoration module, and a channel fusion module.

[0036] S43. In each attention fusion module, a joint modeling mechanism of channel attention and spatial attention is introduced. The highly correlated feature channels are dynamically enhanced through global pooling and local response adjustment strategies. Spatial location information is used to calculate the attention map of salient regions to achieve spatially guided feature focusing.

[0037] S44. Input the attention-enhanced feature map into the context-aware fusion module. The context-aware fusion module includes a local context aggregation submodule and a non-local relation capture submodule. The local context aggregation submodule calculates the semantic consistency of the neighborhood based on the sliding window strategy, and the non-local relation capture submodule calculates the global semantic correlation between distant regions based on pixel-level feature similarity.

[0038] S45. Input the feature map after context-aware fusion into the category-aware decoder. The category-aware decoder includes a category embedding layer, a feature matching layer and a category response generation layer. The category embedding layer maps the preset pathological tissue type into a high-dimensional embedding vector. The feature matching layer calculates the correlation score between each pixel or region feature and the category embedding. The category response generation layer outputs a preliminary classification probability map based on the correlation score.

[0039] S46. Based on the output of the category-aware decoder, a residual enhancement path is introduced to guide the fusion of deep semantic features and shallow structural features, enhance the discrimination power of classification boundaries, and at the same time preserve the details of tissue morphology to prevent feature information from being overly smoothed.

[0040] S47. Adjust the scale consistency of the preliminary classification probability map and introduce a multi-resolution fusion strategy to upsample, normalize and weight the output features of different spatial levels to form a multi-channel classification map with a unified scale.

[0041] S48. Input the multi-channel classification image into the integrated decision module. This module includes multiple parallel classification heads. Each classification head performs independent discrimination based on different feature branches and generates the final classification decision vector using a weighted method.

[0042] S49. The final classification decision vector is processed by Softmax normalization to output the final classification result of the pathological image, which is used as the input basis data for the subsequent negative exclusion module.

[0043] Optionally, S5 specifically includes:

[0044] S51. Receive the pathological image classification judgment result and its corresponding pathological image feature vector generated in step S4, and input the pathological image feature vector into the feature compression mapping module. The feature compression mapping module includes a fully connected layer, a normalization layer and a dimension reduction transformation layer, which are used to map the original high-dimensional feature vector to a low-dimensional discrimination space.

[0045] S52. Construct positive sample embedding distribution and negative sample embedding distribution in a low-dimensional discriminant space. The positive sample embedding distribution is clustered based on the characteristics of known positive cases, and the negative sample embedding distribution is constructed based on the characteristics of known negative cases to create a boundary template.

[0046] S53. Based on the embedding distribution of the positive and negative samples, a discrimination boundary separation model is constructed. The discrimination boundary separation model includes an inter-class spacing constraint mechanism and a boundary direction enhancement mechanism, which are used to form distinguishable feature region boundaries in a low-dimensional space.

[0047] S54. The feature points of the image to be judged are matched with the negative sample template in the discrimination space. The distance weight function is used to evaluate the closeness of the image to the negative distribution and generate a discrimination similarity score.

[0048] S55. Input the discrimination similarity score into the negative probability scoring module, and generate a multi-factor evaluation vector by combining the category boundary separation strength and discrimination confidence information;

[0049] S56. Input the multi-factor evaluation vector into the negative exclusion decision module. The negative exclusion decision module uses a decision tree strategy to logically combine the decision conditions, identify samples that meet the negative distribution characteristics, and output them as negative results.

[0050] S57. Mark the image samples that are determined to be negative as "excluded" and send them to the downstream module for archiving or recording, so that they will not enter the subsequent in-depth analysis process.

[0051] S58. Image samples that are determined to be non-negative are retained for subsequent confidence assessment or manual review to ensure that the complete diagnostic process for high-risk or uncertain samples is carried out.

[0052] Optionally, S6 specifically includes:

[0053] S61. Receive the negative exclusion results output in step S5 and their corresponding discrimination similarity scores and multi-factor evaluation vectors, and construct a confidence vector representation for each sample.

[0054] S62. Input the confidence vector into the confidence weighted evaluation module. The confidence weighted evaluation module consists of a boundary proximity evaluation unit, an intra-class consistency scoring unit, and a model uncertainty measurement unit. It calculates the shortest boundary distance between the sample and the negative template, its local distribution consistency score within the class, and the stability index of its prediction output, respectively.

[0055] S63. The three types of confidence scores are weighted and fused, and a single comprehensive confidence score is generated by using a feature weighted aggregation mechanism. The range of the single comprehensive confidence score is normalized to a fixed interval, which serves as the final representation of the sample's credibility.

[0056] S64. Construct a confidence-boundary coupling mapping model to model the relationship between the comprehensive confidence score and the boundary position of the sample in the discrimination space using bivariate modeling, and obtain the uncertainty distribution state of the sample.

[0057] S65. Based on the uncertainty distribution and sample category labels, dynamically adjust the threshold setting strategy. The threshold setting strategy is based on category density estimation and confidence fluctuation range to determine the minimum credible judgment standard for different categories.

[0058] S66. Compare the dynamic threshold with the comprehensive confidence score, mark samples below the threshold as low-confidence samples, and output them to the manual review processing path.

[0059] S67. Mark samples that exceed the dynamic threshold as high-confidence samples and directly enter the subsequent processing stage or output them to the diagnostic report module.

[0060] S68. Record all parameters and scoring results involved in threshold determination, write them to the evaluation log file, and use them as the data source for continuous learning and parameter tuning of the system model.

[0061] S69. Perform confidence-based hierarchical classification on all retained samples and establish three output levels: high confidence, medium confidence, and low confidence, for subsequent path control and processing priority allocation strategies.

[0062] Optionally, S7 specifically includes:

[0063] S71. Receive the pathological image classification judgment results and negative exclusion results output in steps S4 and S6 respectively, and generate a comprehensive information vector for each image sample, including category label, judgment confidence, negative status mark and evaluation log path.

[0064] S72. Input the comprehensive information vector into the result encoding module. The result encoding module includes a field mapping unit, a format standardization unit, and a structure generation unit. The field mapping unit extracts the required information according to the predefined field template. The format standardization unit unifies the format of each data item to the target encoding rule. The structure generation unit establishes a standard structured record row based on the primary key.

[0065] S73. Write the structured results into the pathological diagnosis database. The pathological diagnosis database adopts a multi-table joint structure. The main table records the unique identifier of the image and the final classification result, and the supplementary tables record the evaluation parameters, negative judgment details and historical identification process metadata.

[0066] S74. The written data is synchronously sent to the diagnostic auxiliary system interface, which adopts a standard data transmission protocol and supports data interaction with hospital information systems, image workstations, manual review terminals or remote collaboration platforms.

[0067] S75. Introduce encryption mechanisms and authentication processes during data transmission to ensure that image recognition results are not tampered with or leaked during transmission.

[0068] S76. Generate an automatic recognition record summary for each classification result and negative exclusion mark, and push it to the doctor's work interface for predictive advice and quick review reference;

[0069] S77. Set up a structured output interface log function to record and archive the time, target address, content summary and feedback status of each output result for subsequent performance evaluation and system traceability analysis.

[0070] The beneficial effects of this invention are:

[0071] This invention proposes a feature recognition and negative exclusion method for pathological images by integrating microscopic imaging and deep semantic modeling techniques, which significantly improves the automation level and diagnostic efficiency of pathological image analysis.

[0072] First, this invention introduces a recognition network that integrates a fusion transformer coding module and a gated dynamic receptive field mechanism in the image feature extraction stage, overcoming the problems of fixed receptive fields and insufficient contextual modeling in traditional convolutional structures. This network can adaptively adjust the feature extraction scale according to the image content, effectively capturing organizational features at different spatial levels and improving the model's recognition accuracy in complex structural regions.

[0073] Secondly, this invention establishes a negative exclusion strategy that combines a category-aware decoder with a discriminative boundary separation model, enabling more refined differentiation of sample categories from the perspective of feature space distribution. By introducing a low-dimensional discriminative space and a positive / negative sample embedding mechanism, accurate identification and automatic exclusion of negative samples are achieved, effectively reducing the manual screening work for doctors when faced with a large number of normal images and lowering diagnostic redundancy.

[0074] Furthermore, this invention constructs a confidence-weighted reliability assessment mechanism that comprehensively considers multiple factors such as discrimination boundary distance, intra-class consistency, and model uncertainty, dynamically setting a classification confidence threshold, which significantly improves the ability to identify low-confidence samples. This mechanism not only enhances the system's responsiveness to abnormal or marginal samples but also provides a priority basis for subsequent manual review and intelligent diagnosis, optimizing the allocation efficiency of diagnostic resources.

[0075] Finally, this invention encodes the classification results and negative markers into structured data, which can be seamlessly integrated with hospital diagnostic support systems or remote consultation platforms. This truly achieves end-to-end automation from image acquisition, processing, and recognition to structured output, demonstrating excellent scalability and clinical applicability. Overall, this invention significantly outperforms existing technologies in image recognition accuracy, negative exclusion capability, system robustness, and structured output efficiency, providing an effective technical path for the intelligent development of digital pathology. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0077] Figure 1 This is an overall flowchart of a pathological feature identification and negative exclusion method based on microscopic imaging proposed in this invention;

[0078] Figure 2 This is a schematic diagram of the pathological image recognition network structure proposed in this invention, which integrates transformer coding and gated dynamic receptive field mechanism.

[0079] Figure 3 This is a flowchart of the dynamic threshold screening process for the confidence-weighted evaluation mechanism proposed in this invention. Detailed Implementation

[0080] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0081] refer to Figure 1-3 A method for identifying pathological features and excluding negative results based on microscopic imaging includes the following steps:

[0082] S1. Acquire high-resolution images of pathological tissue sections using a microscopic imaging device, and digitize the high-resolution images to obtain raw microscopic image data;

[0083] S2. Perform image preprocessing on the original microscopic image data, including noise suppression, color normalization, tissue boundary enhancement, and foreground region segmentation, to generate a processed image with clear structure.

[0084] S3. Construct a pathological image recognition network that integrates a transformer coding module and a gated dynamic receptive field mechanism. Take the processed image as input, extract multi-scale spatial semantic features through feature alignment, channel dynamic control and global context modeling, and generate pathological feature vectors.

[0085] S4. Input the pathological feature vector into the classification decoding module, use a multi-layer attention-guided feature fusion structure to perform contextual association modeling of features in different regions, and use a category-aware decoder to generate the classification judgment result of the pathological image.

[0086] S5. Construct a discrimination boundary separation model in the feature compression space. Based on the positive and negative sample embedding distribution, perform negative exclusion judgment on the preliminary classification judgment result, screen samples whose feature performance is close to the negative template and output them as negative results.

[0087] S6. Perform a confidence-weighted evaluation on the negative exclusion results. Combine the uncertainty distribution output by the discrimination boundary separation model with the class boundary distance, set a dynamic threshold, screen out low-confidence samples, and retain high-confidence images for subsequent processing.

[0088] S7. Encode the pathological classification results and negative exclusion labels according to a unified structural format, write them into a structured dataset, and send them synchronously to the diagnostic assistance system for doctors to access or for analysis by the remote consultation system.

[0089] This invention constructs an integrated intelligent pathological image recognition and negative sample exclusion method, covering the entire process from image acquisition, preprocessing, multi-scale feature extraction, attention fusion classification, feature boundary judgment, to reliable screening and structured output. Compared with existing single classification models, this method introduces depth transformer encoding and dynamic receptive field mechanism for collaborative modeling, improving the representation ability of lesion areas and complex tissue morphologies. Simultaneously, through discrimination boundary separation and confidence assessment mechanisms, it effectively solves the problems of inaccurate negative sample exclusion and high misdiagnosis rate in traditional methods, achieving an automated closed loop for high-precision recognition and screening processes, significantly reducing the burden of manual review, and improving the intelligence level and stability of the auxiliary diagnostic system.

[0090] In this embodiment, S1 specifically includes:

[0091] S11. Fix the stained pathological tissue sections onto a glass slide carrier and place them in a fully automated digital microscope platform to prepare for image acquisition.

[0092] S12. Control the fully automatic digital microscope platform to automatically scan the tissue section region by region according to the set scanning magnification and spatial resolution to obtain local field-of-view images;

[0093] S13. Stitch together multiple local view images in spatial order to generate a complete high-resolution slice image;

[0094] S14. Perform a unified format conversion on the high-resolution slice image to generate raw microscopic image data that meets the requirements of subsequent processing.

[0095] By introducing a fully automated digital microscope platform, high-throughput, standardized slice image acquisition and stitching are achieved, and the image format is unified, providing highly consistent, high-quality input data for subsequent model training and deployment. Compared with traditional manual image acquisition methods, this method eliminates problems such as inconsistent image resolution and scanning overlap errors caused by human differences, ensuring the integrity and accuracy of the original image information, effectively supporting the model's robust extraction of minute structural features, and improving the overall system stability.

[0096] In this embodiment, S2 specifically includes:

[0097] S21. Perform image denoising processing on the original microscopic image data, and use edge-preserving filtering to remove background noise and imaging artifacts while preserving tissue structure details.

[0098] S22. Perform color normalization processing on the staining in the original microscopic image, and unify the color distribution of the image based on the target staining template to improve the consistency between images.

[0099] S23. Perform tissue structure edge enhancement operation on the boundary region of the original microscopic image, and enhance the cell and tissue contours through gradient operation to highlight local structural features;

[0100] S24. Use a foreground extraction algorithm to segment the original microscopic image, remove the background region and extract the tissue foreground region to generate a processed image with clear structure and well-defined boundaries.

[0101] This method integrates multiple enhancement and segmentation strategies in the image preprocessing stage, constructing a processing pipeline tailored to the characteristics of stained pathological images, from edge-preserving filtering, color normalization, and structural boundary enhancement to foreground extraction. This processing workflow can significantly reduce interference from non-tissue regions and improve the separation of cellular structures within the image, providing clearer input features for subsequent recognition networks. Compared to traditional methods that only employ color correction or edge extraction, this method can effectively adapt to the differences in images acquired from multiple batches and devices, significantly enhancing the model's transferability in cross-domain tasks.

[0102] In this embodiment, S3 specifically includes:

[0103] S31. Input the processed image obtained in step S2 into the shallow feature extraction module. The shallow feature extraction module consists of multiple standard two-dimensional convolutional layers, batch normalization layers and nonlinear activation functions, which are used to extract basic visual features of the image such as low-level edges, textures and color distribution to generate a primary feature map.

[0104] S32. Input the primary feature map into the gated dynamic receptive field module. The gated dynamic receptive field module consists of multiple parallel convolutional branches. Each branch has a different convolutional kernel size and stride parameter, which is used to extract multi-scale structural features. The gated unit adaptively allocates the weights of each branch according to the channel response value and fuses and outputs a multi-scale feature map.

[0105] S33. Input the multi-scale feature map into the feature enhancement module. The feature enhancement module includes two parts: channel attention mechanism and spatial attention mechanism. It weights and enhances high-response channels to strengthen the spatial expression ability of the lesion area and outputs the enhanced feature map.

[0106] S34. The enhanced feature map is input to the transformer encoding module. The transformer encoding module consists of multiple stacked encoder units. Each encoder unit includes a multi-head self-attention computing layer, a feedforward network and a residual connection structure. At the same time, relative position encoding is introduced to maintain the spatial semantics of the pathological structure, extract the context information between distant structures, and output the context-encoded feature map.

[0107] S35. Construct a cross-scale interactive attention module, align the scales of the encoded feature maps of different depths and establish a cross-attention mapping, calculate the similarity weighting value between scales through the feature matching function, fuse them to generate the contextual relationship between high-expression regions and low-expression regions, and output a cross-scale fused feature map.

[0108] S36. Input the cross-scale fused feature map into the structural semantic separation unit to separate the main path features for structure preservation and the auxiliary path features for semantic enhancement. The main path uses a position attention mechanism to enhance structural stability, and the auxiliary path uses a context enhancement mechanism to strengthen the semantic representation of the lesion region. Then, the features are weighted and combined through the path fusion module.

[0109] S37. Perform a dimension compression mapping operation on the fused feature map, and convert the multi-dimensional image features into a vector representation of a unified dimension through global average pooling and linear transformation to form the final pathological image feature vector and output it.

[0110] This recognition network architecture, which integrates a transformer encoding module and a gated dynamic receptive field mechanism, breaks through the limitations of traditional convolutional neural networks, such as fixed receptive fields and insufficient long-range dependency representation. In the model, the gated branch structure guides multi-scale path selection, adapting to information at different structural scales; the transformer module efficiently models long-distance semantic relationships between tissues through a global attention mechanism, thereby significantly improving the accuracy of lesion recognition. Especially when dealing with pathological images exhibiting strong morphological heterogeneity and blurred regional boundaries, this network structure can better capture high-order features and full-image contextual semantics, greatly enhancing the model's ability to discriminate tissue types and boundaries.

[0111] In this embodiment, S4 specifically includes:

[0112] S41. Receive the pathological image feature vector output in step S3, and expand the feature vector into a two-dimensional spatial feature map through a linear projection layer for subsequent structural reconstruction and classification reasoning.

[0113] S42. Input the two-dimensional spatial feature map into a multi-level classification decoding network. The multi-level classification decoding network consists of multiple stacked attention fusion decoding units. Each attention fusion decoding unit includes an attention fusion module, a convolution restoration module, and a channel fusion module.

[0114] S43. In each attention fusion module, a joint modeling mechanism of channel attention and spatial attention is introduced. The highly correlated feature channels are dynamically enhanced through global pooling and local response adjustment strategies. Spatial location information is used to calculate the attention map of salient regions to achieve spatially guided feature focusing.

[0115] S44. Input the attention-enhanced feature map into the context-aware fusion module. The context-aware fusion module includes a local context aggregation submodule and a non-local relation capture submodule. The local context aggregation submodule calculates the semantic consistency of the neighborhood based on the sliding window strategy, and the non-local relation capture submodule calculates the global semantic correlation between distant regions based on pixel-level feature similarity.

[0116] S45. Input the feature map after context-aware fusion into the category-aware decoder. The category-aware decoder includes a category embedding layer, a feature matching layer and a category response generation layer. The category embedding layer maps the preset pathological tissue type into a high-dimensional embedding vector. The feature matching layer calculates the correlation score between each pixel or region feature and the category embedding. The category response generation layer outputs a preliminary classification probability map based on the correlation score.

[0117] S46. Based on the output of the category-aware decoder, a residual enhancement path is introduced to guide the fusion of deep semantic features and shallow structural features, enhance the discrimination power of classification boundaries, and at the same time preserve the details of tissue morphology to prevent feature information from being overly smoothed.

[0118] S47. Adjust the scale consistency of the preliminary classification probability map and introduce a multi-resolution fusion strategy to upsample, normalize and weight the output features of different spatial levels to form a multi-channel classification map with a unified scale.

[0119] S48. Input the multi-channel classification image into the integrated decision module. This module includes multiple parallel classification heads. Each classification head performs independent discrimination based on different feature branches and generates the final classification decision vector using a weighted method.

[0120] S49. The final classification decision vector is processed by Softmax normalization to output the final classification result of the pathological image, which is used as the input basis data for the subsequent negative exclusion module.

[0121] By constructing a discriminative boundary separation model in a feature compression space, this method achieves the embedding of pathological image samples in a low-dimensional feature space and the modeling of classification boundaries, effectively distinguishing between positive and negative samples. Traditional probability-based classification methods are prone to misidentifying negative samples as positive, especially when feature distribution boundaries are blurred. This method introduces a boundary direction enhancement mechanism and inter-class distance constraints to construct a clear negative template. Through feature matching, discriminative similarity is calculated, significantly improving the accuracy of negative sample identification, reducing the risk of misdiagnosis, and bringing greater safety and value to pathological screening systems.

[0122] In this embodiment, S5 specifically includes:

[0123] S51. Receive the pathological image classification judgment result and its corresponding pathological image feature vector generated in step S4, and input the pathological image feature vector into the feature compression mapping module. The feature compression mapping module includes a fully connected layer, a normalization layer and a dimension reduction transformation layer, which are used to map the original high-dimensional feature vector to a low-dimensional discrimination space.

[0124] S52. Construct positive sample embedding distribution and negative sample embedding distribution in a low-dimensional discriminant space. The positive sample embedding distribution is clustered based on the characteristics of known positive cases, and the negative sample embedding distribution is constructed based on the characteristics of known negative cases to create a boundary template.

[0125] S53. Based on the embedding distribution of the positive and negative samples, a discrimination boundary separation model is constructed. The discrimination boundary separation model includes an inter-class spacing constraint mechanism and a boundary direction enhancement mechanism, which are used to form distinguishable feature region boundaries in a low-dimensional space.

[0126] S54. The feature points of the image to be judged are matched with the negative sample template in the discrimination space. The distance weight function is used to evaluate the closeness of the image to the negative distribution and generate a discrimination similarity score.

[0127] S55. Input the discrimination similarity score into the negative probability scoring module, and generate a multi-factor evaluation vector by combining the category boundary separation strength and discrimination confidence information;

[0128] S56. Input the multi-factor evaluation vector into the negative exclusion decision module. The negative exclusion decision module uses a decision tree strategy to logically combine the decision conditions, identify samples that meet the negative distribution characteristics, and output them as negative results.

[0129] S57. Mark the image samples that are determined to be negative as "excluded" and send them to the downstream module for archiving or recording, so that they will not enter the subsequent in-depth analysis process.

[0130] S58. Image samples that are determined to be non-negative are retained for subsequent confidence assessment or manual review to ensure that the complete diagnostic process for high-risk or uncertain samples is carried out.

[0131] This invention integrates three-dimensional information—boundary proximity, intra-class consistency score, and model prediction uncertainty—through a confidence-weighted evaluation mechanism to comprehensively measure the reliability of sample classification results. Compared to traditional simplified judgment mechanisms based on fixed thresholds or Softmax probability values, this method dynamically adjusts the threshold strategy and constructs a coupled mapping model, thereby achieving refined reliability control for different sample categories and distribution characteristics. This mechanism not only improves the system's response sensitivity to uncertain samples but also ensures the interpretability and risk controllability of the identification results, making it particularly suitable for clinical scenarios with extremely low tolerance for misjudgment.

[0132] In this embodiment, S6 specifically includes:

[0133] S61. Receive the negative exclusion results output in step S5 and their corresponding discrimination similarity scores and multi-factor evaluation vectors, and construct a confidence vector representation for each sample.

[0134] S62. Input the confidence vector into the confidence weighted evaluation module. The confidence weighted evaluation module consists of a boundary proximity evaluation unit, an intra-class consistency scoring unit, and a model uncertainty measurement unit. It calculates the shortest boundary distance between the sample and the negative template, its local distribution consistency score within the class, and the stability index of its prediction output, respectively.

[0135] S63. The three types of confidence scores are weighted and fused, and a single comprehensive confidence score is generated by using a feature weighted aggregation mechanism. The range of the single comprehensive confidence score is normalized to a fixed interval, which serves as the final representation of the sample's credibility.

[0136] S64. Construct a confidence-boundary coupling mapping model to model the relationship between the comprehensive confidence score and the boundary position of the sample in the discrimination space using bivariate modeling, and obtain the uncertainty distribution state of the sample.

[0137] S65. Based on the uncertainty distribution and sample category labels, dynamically adjust the threshold setting strategy. The threshold setting strategy is based on category density estimation and confidence fluctuation range to determine the minimum credible judgment standard for different categories.

[0138] S66. Compare the dynamic threshold with the comprehensive confidence score, mark samples below the threshold as low-confidence samples, and output them to the manual review processing path.

[0139] S67. Mark samples that exceed the dynamic threshold as high-confidence samples and directly enter the subsequent processing stage or output them to the diagnostic report module.

[0140] S68. Record all parameters and scoring results involved in threshold determination, write them to the evaluation log file, and use them as the data source for continuous learning and parameter tuning of the system model.

[0141] S69. Perform confidence-based hierarchical classification on all retained samples and establish three output levels: high confidence, medium confidence, and low confidence, for subsequent path control and processing priority allocation strategies.

[0142] This invention structures and encodes pathological image classification results and negative exclusion markers, enabling efficient integration with diagnostic support systems and achieving standardized output and automated integration of diagnostic data. Through multi-field result encapsulation and standardized transmission protocols, it effectively solves the problems of inconsistent data interfaces and inconvenient result retrieval in traditional pathology systems. This provides a solid data foundation for remote consultations, model decision review, and system collaboration, significantly improving the deployability and information exchange capabilities of intelligent recognition systems in actual clinical workflows.

[0143] In this embodiment, S7 specifically includes:

[0144] S71. Receive the pathological image classification judgment results and negative exclusion results output in steps S4 and S6 respectively, and generate a comprehensive information vector for each image sample, including category label, judgment confidence, negative status mark and evaluation log path.

[0145] S72. Input the comprehensive information vector into the result encoding module. The result encoding module includes a field mapping unit, a format standardization unit, and a structure generation unit. The field mapping unit extracts the required information according to the predefined field template. The format standardization unit unifies the format of each data item to the target encoding rule. The structure generation unit establishes a standard structured record row based on the primary key.

[0146] S73. Write the structured results into the pathological diagnosis database. The pathological diagnosis database adopts a multi-table joint structure. The main table records the unique identifier of the image and the final classification result, and the supplementary tables record the evaluation parameters, negative judgment details and historical identification process metadata.

[0147] S74. The written data is synchronously sent to the diagnostic auxiliary system interface, which adopts a standard data transmission protocol and supports data interaction with hospital information systems, image workstations, manual review terminals or remote collaboration platforms.

[0148] S75. Introduce encryption mechanisms and authentication processes during data transmission to ensure that image recognition results are not tampered with or leaked during transmission.

[0149] S76. Generate an automatic recognition record summary for each classification result and negative exclusion mark, and push it to the doctor's work interface for predictive advice and quick review reference;

[0150] S77. Set up a structured output interface log function to record and archive the time, target address, content summary and feedback status of each output result for subsequent performance evaluation and system traceability analysis.

[0151] By constructing a complete closed-loop mechanism encompassing image acquisition, preprocessing, feature extraction, classification, negative screening, and reliability evaluation, this method highly integrates image processing with diagnostic tasks, supporting intelligent and automated workflow from front-end acquisition to back-end result output. Compared to the traditional workflow's problems of multiple separate modules and high reliance on manual judgment, this method achieves smooth end-to-end data flow and structural optimization, improving diagnostic efficiency while ensuring the traceability and controllability of key judgment steps. It provides a standardized, automated, and highly accurate technical path for building an intelligent pathology platform.

[0152] Example 1:

[0153] To verify the feasibility of this invention in practice, it was applied to the daily work of the pathology department of a tertiary hospital. Doctors there process a large number of microscopic images daily from pathological samples such as breast biopsies, lung punctures, and lymph node sections. These images, after staining, reveal complex and varied tissue structures. Especially in the early screening stage, many images are negative or suspected negative. Manually reviewing each image is not only extremely labor-intensive but also highly susceptible to subjective judgment and fatigue, leading to misdiagnosis or missed diagnoses. To alleviate the pressure on pathologists and improve the efficiency and accuracy of negative screening, the pathology department introduced the microscopic imaging-based pathological feature identification and negative exclusion method proposed in this invention, and systematically upgraded its digital pathology workflow.

[0154] In this application scenario, the doctor first places the stained pathological tissue slide onto a fully automated digital microscope platform and automatically scans the entire slide region by region by setting the scanning magnification and resolution parameters. After scanning, the system seamlessly stitches together hundreds of local images to generate a complete high-resolution slide image, and then converts it to a uniform 25600×15360 pixel TIFF format. These images are then fed into the image preprocessing module, where the system automatically performs noise removal, color normalization, edge enhancement, and tissue region segmentation. Compared to the unprocessed image, the preprocessed image shows a 22.5% improvement in edge sharpness and a nearly 30% reduction in background artifacts, making it more suitable for subsequent feature extraction and analysis.

[0155] After image processing, the system initiates a recognition network that integrates a transformer-encoding module and a gated dynamic receptive field mechanism. After extracting basic features through shallow convolutions, the gating mechanism dynamically selects different receptive field paths to extract multi-scale structural information. This information is then combined with the transformer module to model long-distance semantic relationships between different tissue regions. Experimental data shows that this network structure improves lesion recognition accuracy by approximately 8.7% compared to the traditional ResNet structure. In experiments with breast biopsy images, the model achieved a 94.3% accuracy rate in recognizing ductal carcinoma regions and a 91.6% accuracy rate in recognizing atypical hyperplasia regions.

[0156] During the classification stage, the model further inputs features into a multi-layer attention-guided decoder and outputs preliminary classification results through contextual modeling and category awareness mechanisms. The system constructs a negative template embedding distribution and boundary separation model, using a low-dimensional feature space to exclude negative samples. In nearly a month of application, the system processed 21,856 breast pathology slide images, with negative samples accounting for 62.4%. Through boundary screening, the system successfully excluded 12,294 high-confidence negative images, reducing the workload of manual slide reading by more than half.

[0157] To avoid misclassification of low-confidence negative images, the system introduces a confidence-weighted mechanism. This mechanism performs a weighted fusion of boundary distance, uncertainty score, and intra-class consistency assessment on all excluded samples, and outputs a confidence level based on a dynamic threshold strategy. Verification has shown that this mechanism improves the negative fidelity rate to 96.1% in negative identification samples, effectively preventing false negatives from entering the subsequent reporting stage. For samples with low confidence, the system automatically marks them as requiring "manual review," ensuring that the final diagnosis of high-risk images is made by senior physicians.

[0158] The system ultimately writes the classification results, negative markers, and confidence levels into the hospital's diagnostic database in structured data format, and automatically synchronizes them to the remote consultation platform and auxiliary diagnostic terminals. Doctors can intuitively view the model's predicted results for each slide, region annotation information, confidence scores, and judgment criteria on the electronic pathology platform, achieving a highly efficient collaborative mechanism of "preliminary screening + key verification."

[0159] Overall, the introduction of this invention has reduced the average workload per person in the initial image screening stage of the pathology department by 41.7%, and shortened the overall workflow by an average of about 2.6 hours per day, particularly alleviating diagnostic bottlenecks during holidays and peak screening periods. Furthermore, doctors' consistent feedback on the system's accuracy reached over 92%, with most believing that the system helps reduce repetitive operations, focus on high-risk areas, and improve the scientific rigor and controllability of the work process.

[0160] In summary, this embodiment clearly demonstrates the outstanding value of the present invention in improving the efficiency of negative screening, reducing the false positive rate, and supporting structured diagnostic workflow. It effectively solves the technical problems of low efficiency, large accuracy fluctuations, and concentrated review pressure in the traditional pathological analysis process, and has significant clinical application prospects.

[0161] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying pathological features and excluding negative results based on microscopic imaging, characterized in that, The method comprises the following steps: S1, obtaining a high-resolution image of a pathological tissue section by a microscopic imaging device, and performing digital processing on the high-resolution image to obtain original microscopic image data; S2, performing image preprocessing on the original microscopic image data to generate a processed image with clear structure; S3, constructing a pathological image recognition network that fuses a transformer coding module and a gated dynamic receptive field mechanism, taking the processed image as input, extracting multi-scale spatial semantic features through feature alignment, channel dynamic regulation and global context modeling, and generating a pathological feature vector; S4, inputting the pathological feature vector into a classification decoding module, using a multi-layer attention-guided feature fusion structure to model the context association of different regional features, and using a class-aware decoder to generate a preliminary classification determination result of the pathological image; S5, constructing a discriminant boundary separation model in a feature compression space, based on the embedding distribution of positive and negative samples, performing negative exclusion judgment on the preliminary classification determination result, screening samples with feature performance close to the negative template, and outputting as negative results; S6, performing confidence weighted evaluation on the negative exclusion result, combining the uncertainty distribution and class boundary distance output by the discriminant boundary separation model, and setting a dynamic threshold for high-confidence sample screening; Specifically comprising: receiving the output in step S5 and constructing a confidence vector for each sample, inputting the confidence vector into a confidence weighted evaluation module, the confidence weighted evaluation module being composed of a boundary proximity evaluation unit, an intra-class consistency scoring unit and a model uncertainty measurement unit, respectively calculating the shortest boundary distance of the sample and the negative template, the local distribution consistency score of the sample in the class, and the stability index of the predicted output; weighting and fusing the three types of confidence score results, using a feature weighted aggregation mechanism to generate a single comprehensive confidence score, and normalizing the single comprehensive confidence score value range to a fixed interval; constructing a confidence-boundary coupling mapping model to model the two variables of the single comprehensive confidence score and the boundary position relationship of the sample in the discriminant space, and obtaining the uncertainty distribution state of the sample; According to the uncertainty distribution state and the sample class label, dynamically adjusting the threshold setting strategy, the threshold setting strategy is based on class density estimation and confidence fluctuation range to determine the minimum credible judgment standard under different categories; comparing the dynamic threshold with the single comprehensive confidence score, marking the samples below the threshold as low-confidence samples, and outputting to the artificial review processing path; marking samples higher than the dynamic threshold as high-confidence samples; S7, encoding the pathological classification result and the negative exclusion label according to a unified structure format, writing into a structured data set, and synchronously sending to a diagnosis assistance system.

2. The method of claim 1, wherein, The S1 specifically comprises: S11, fix the pathological tissue section after dyeing treatment on the slide carrier, and place it in the automatic digital microscope platform to prepare for image acquisition; S12, control the automatic digital microscope platform to scan the tissue section region by region according to the set scanning magnification and spatial resolution, and obtain a local field image; S13, splice multiple local field images in the order of spatial position to generate a complete high-resolution section image; S14, perform unified format conversion on the high-resolution section image to generate original microscopic image data.

3. The method of claim 1, wherein, The S2 specifically includes: S21, perform image denoising processing on the original microscopic image data, remove background noise and imaging artifacts by using edge-preserving filtering, and retain tissue structure details; S22, perform color normalization processing on the staining condition in the original microscopic image, and unify the color distribution of the image based on the target staining template; S23, perform tissue structure edge enhancement operation on the boundary region of the original microscopic image, strengthen the cell and tissue contour through gradient operation, and highlight the local structure features; S24, use a foreground extraction algorithm to segment the original microscopic image, remove the background region and extract the tissue foreground region, and generate a processed image with clear structure and clear boundary.

4. The method of claim 1, wherein, The S3 specifically includes: S31, input the processed image obtained in step S2 into a shallow feature extraction module, the shallow feature extraction module is composed of multiple standard two-dimensional convolution layers, multiple normalization layers and nonlinear activation functions, the extracted basic visual features include low-level edges, textures and dye distribution of the image, and a primary feature map is generated; S32, input the primary feature map into a gated dynamic receptive field module, the gated dynamic receptive field module is composed of multiple parallel convolution branches, each branch has different convolution kernel size and stride parameters, and multi-scale structure features are extracted; the gate unit adaptively allocates the weight of each branch according to the channel response value, and fuses and outputs a multi-scale feature map; S33, input the multi-scale feature map into a feature enhancement module, the feature enhancement module includes channel attention mechanism and spatial attention mechanism, weights and enhances high-response channels, strengthens the spatial expression ability of the lesion area, and outputs an enhanced feature map; S34, input the enhanced feature map into a transformer encoder module, the transformer encoder module is composed of multiple stacked encoder units, each encoder unit includes a multi-head self-attention calculation layer, a feedforward network and a residual connection structure, and relative position encoding is introduced at the same time, which extracts the context information between long-distance structures and outputs a context encoding feature map; S35, construct a cross-scale interactive attention module, align the scale of the encoding feature maps at different depths and establish cross-attention mapping, calculate the similarity weighting value between scales through a feature matching function, fuse the context relationship between high expression areas and low expression areas, and output a cross-scale fusion feature map; S36, input the cross-scale fusion feature map into a structure semantic separation unit to separate the main path feature and the auxiliary path feature, use a position attention mechanism for the main path, use a context enhancement mechanism for the auxiliary path, and then perform weighted combination through a path fusion module; S37, perform dimension compression mapping operation on the fused feature map, convert the multi-dimensional image feature into a unified dimension vector representation through global average pooling and linear transformation, and output the final pathological image feature vector.

5. The method of claim 1, wherein, The S4 specifically comprises: S41, receive the pathological image feature vector output in step S3, and expand the feature vector into a two-dimensional spatial feature map through a linear projection layer; S42, input the two-dimensional spatial feature map into a multi-level classification decoding network, the multi-level classification decoding network is composed of a plurality of stacked attention fusion decoding units, each attention fusion decoding unit includes an attention fusion module, a convolution restoration module and a channel fusion module; S43, in each attention fusion module, a joint modeling mechanism of channel attention and spatial attention is introduced, the high correlation feature channel is dynamically enhanced through global pooling and local response adjustment strategy, and the significant region attention map is calculated using spatial position information; S44, input the feature map enhanced by attention into a context perception fusion module, the context perception fusion module includes a local context aggregation sub-module and a non-local relationship capture sub-module, the local context aggregation sub-module calculates the neighborhood semantic consistency based on a sliding window strategy, and the non-local relationship capture sub-module calculates the global semantic correlation between distant regions based on pixel-level feature similarity; S45, input the feature map fused by context perception into a class perception decoder, the class perception decoder includes a class embedding layer, a feature matching layer and a class response generation layer, the class embedding layer maps the preset pathological tissue type into a high-dimensional embedding vector, the feature matching layer calculates the correlation score between each pixel or region feature and the class embedding, and the class response generation layer outputs a preliminary classification probability map according to the correlation score; S46, on the basis of the output of the class perception decoder, a residual enhancement path is introduced to guide the fusion of deep semantic features and shallow structural features, enhance the classification boundary discrimination, and retain the tissue morphology details; S47, perform scale consistency adjustment on the preliminary classification probability map, introduce a multi-resolution fusion strategy, upsample, normalize and weight fuse the output features of different spatial levels to form a multi-channel classification map of uniform scale; S48, input the multi-channel classification map into an integrated decision module, the integrated decision module includes a plurality of parallel classification heads, each classification head performs independent discrimination based on different feature branches, and generates a final classification judgment vector using weight weighting method; S49, perform Softmax normalization processing on the final classification judgment vector to output the final classification result of the pathological image.

6. The method of claim 1, wherein, The S5 specifically comprises: S51, receive the pathological image classification determination result generated in step S4 and its corresponding pathological image feature vector, and input the pathological image feature vector into a feature compression mapping module, the feature compression mapping module comprising a full connection layer, a normalization layer and a dimension reduction transformation layer, mapping the original high-dimensional feature vector into a low-dimensional discriminant space; S52, construct a positive sample embedding distribution and a negative sample embedding distribution in the low-dimensional discriminant space, the positive sample embedding distribution being represented by clustering the features of known positive cases, and the negative sample embedding distribution being constructed by a boundary template according to the features of known negative cases; S53, based on the positive and negative sample embedding distributions, construct a discriminant boundary separation model, the discriminant boundary separation model comprising an inter-class distance constraint mechanism and a boundary direction enhancement mechanism, forming a distinguishable feature region boundary in the low-dimensional space; S54, match the feature points of the image to be determined with the negative sample template in the discriminant space, evaluate the closeness to the negative distribution by using a distance weight function, and generate a discriminant similarity score; S55, input the discriminant similarity score into a negative probability score module, combine the class boundary separation strength and the discriminant confidence information to generate a multi-factor evaluation vector; S56, input the multi-factor evaluation vector into a negative exclusion determination module, the negative exclusion determination module uses a decision tree strategy to logically combine the determination conditions, identifies the samples meeting the negative distribution characteristics, and outputs as negative results; S57, mark the image samples determined as negative as "excluded" state, and send to the downstream module for archiving or recording; S58, keep the image samples determined as non-negative.

7. The method of claim 1, wherein, Before the confidence vector is input into the confidence weight evaluation module, it includes: Receiving the negative exclusion results output by S5, and their corresponding discriminant similarity scores and multi-factor evaluation vectors, and constructing a confidence vector for each sample; The step of marking the samples higher than the dynamic threshold as high-confidence samples further comprises: Recording all parameters participating in the threshold determination and score results, and writing into an evaluation log file; Performing confidence hierarchical classification on all retained samples, and establishing three-level output grades of high-confidence, medium-confidence and low-confidence.

8. The method of claim 1, wherein, The S7 specifically comprises: S71, receiving the pathological image classification determination results and negative exclusion results output by steps S4 and S6 respectively, and generating a comprehensive information vector for each image sample, including a class label, a determination confidence, a negative state mark and an evaluation log path; S72, inputting the comprehensive information vector into a result encoding module, the result encoding module comprising a field mapping unit, a format standardization unit and a structure generation unit, the field mapping unit extracting the required information according to a predefined field template, the format standardization unit unifying the formats of various data to a target coding rule, and the structure generation unit establishing a standard structured record line according to a primary key; S73, writing the structured result into a pathological diagnosis database, the pathological diagnosis database adopting a multi-table joint structure, a main table recording image unique identification and final classification result, and an auxiliary table recording evaluation parameters, negative determination details and historical identification process meta-information; S74, synchronously send the written data to a diagnostic auxiliary system interface, which adopts a standard data transmission protocol; S75, introduce an encryption mechanism and an identity verification process in the data transmission process; S76, generate an automatic identification record summary for each classification result and a negative exclusion mark, and push it to the doctor's work interface; S77, set up a structured output interface log function to record and archive the time, target address, content summary and feedback status of each result output.

Images