Biopsy pathological image intelligent diagnosis classification method

Abstract

The present application relates to the technical field of medical image intelligent diagnosis, and discloses a biopsy pathological image intelligent diagnosis classification method. The method collects digital pathological image data of biopsy samples, extracts morphological feature regions in the image as a basic analysis unit. A multi-scale feature fusion network is constructed, the basic analysis unit is input into a first feature extraction layer to obtain a primary feature atlas, spatial attention weighting is performed through a secondary feature extraction layer, and a fusion feature vector is generated. A dynamic classification threshold pool is established, a feature subspace is divided according to the dimension distribution of the fusion feature vector, and independent classification boundary parameters are allocated to each subspace. A pre-trained pathological knowledge graph is loaded, the fusion feature vector is matched with the node features in the graph in terms of similarity, and knowledge nodes with a matching degree exceeding a preset threshold are selected as auxiliary diagnosis basis. An initial classification label is generated based on the output result of the dynamic classification threshold pool and the auxiliary diagnosis basis, and a final diagnosis classification result is output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image intelligent diagnosis, in particular to a biopsy pathological image intelligent diagnosis classification method. BACKGROUND

[0002] In the field of medical diagnosis, biopsy pathological image diagnosis has always been a key link for disease diagnosis. Traditional biopsy pathological image diagnosis mainly relies on pathologists to directly observe pathological sections under a microscope. Taking gastric cancer diagnosis as an example, China is a high-incidence country of gastric cancer, and about 1 / 5 of all tumor deaths are caused by gastric cancer every year. If early gastric cancer can be detected in time, the 5-year survival rate can exceed 90%, but the early gastric cancer diagnosis rate in China is less than 10%. Before visual analysis of the tissue sample, the pathologist needs to collect the biopsy sample and stain it with hematoxylin and eosin, and then carefully examine the microscopic images of the biopsy tissue sample under different magnifications to identify patterns, textures and various morphological characteristics to make a diagnosis.

[0003] With the continuous development of medical technology and the increasing demand for disease diagnosis accuracy and efficiency, computer-aided diagnosis (CAD) technology has emerged. The main purpose of CAD is to help doctors diagnose diseases more quickly and accurately by using computer science and technology. It inputs medical images into a computer and uses computer programs to analyze the images to identify disease characteristics, thereby assisting doctors in diagnosis.

[0004] CAD technology is widely used in the medical field, covering orthopedics, pneumology, breast surgery, cardiology and other fields. In cancer diagnosis, CAD can be used for global detection of cancer, lesion positioning and segmentation, and benign and malignant judgment. For example, in lung cancer screening, CAD technology can assist doctors in more accurately identifying lung nodules and improving the detection rate of early lung cancer; in breast cancer diagnosis, it can also help doctors find potential lesions in breast images. Early CAD technology mainly relies on expert experience and manual analysis, using some simple image processing methods and pattern recognition techniques, such as morphological feature-based classification and threshold segmentation-based methods, but these methods have limited accuracy and are difficult to meet the complex and varied diagnostic needs of clinics. SUMMARY

[0005] The purpose of the present application is to provide a biopsy pathological image intelligent diagnosis classification method to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a biopsy pathological image intelligent diagnosis classification method, which comprises:

[0007] Collecting digital pathological image data of biopsy samples, extracting morphological feature regions in the image as basic analysis units;

[0008] A multi-scale feature fusion network is constructed, and the basic analysis unit is input into a first feature extraction layer to obtain a primary feature map, and the primary feature map is subjected to spatial attention weighting through a secondary feature extraction layer to generate a fusion feature vector;

[0009] A dynamic classification threshold pool is established, and a feature subspace is divided according to the dimension distribution of the fusion feature vector, and an independent classification boundary parameter is allocated to each feature subspace;

[0010] A pre-trained pathology knowledge graph is loaded, and the fusion feature vector is subjected to similarity matching with the node features in the pathology knowledge graph, and the knowledge nodes with a matching degree exceeding a preset threshold are selected as auxiliary diagnosis basis;

[0011] An initial classification label is generated based on the output result of the dynamic classification threshold pool and the auxiliary diagnosis basis, and the initial classification label is corrected through a cross-modal consistency verification module to output a final diagnosis classification result.

[0012] Preferably, the extraction process of the morphological feature region comprises:

[0013] An adaptive threshold segmentation algorithm is used to divide the digital pathology image into regions, and the nuclear-cytoplasmic ratio and texture complexity index of each region are calculated;

[0014] Regions with a nuclear-cytoplasmic ratio greater than a first critical value and a texture complexity index less than a second critical value are marked as effective morphological feature regions;

[0015] Edge gradient detection is performed on adjacent effective morphological feature regions, and continuous regions with a gradient difference less than a third critical value are merged.

[0016] Preferably, the construction process of the multi-scale feature fusion network comprises:

[0017] Parallel convolution kernel groups are set in the first feature extraction layer, and each convolution kernel group contains different sizes of hole convolution kernels;

[0018] The channel attention weight of the secondary feature extraction layer is point multiplied with the spatial attention weight to generate a feature enhancement coefficient;

[0019] Each channel of the primary feature map is subjected to nonlinear scaling through the feature enhancement coefficient to generate a fusion feature vector.

[0020] Preferably, the operation process of the dynamic classification threshold pool comprises:

[0021] Principal component analysis is performed on the fusion feature vector, and principal component axes with a variance contribution rate exceeding a fourth critical value are selected;

[0022] Divide the feature space into hypercube subspaces along the principal component axis direction, and count the classification accuracy of historical samples in each subspace;

[0023] According to the classification accuracy, dynamically adjust the classification boundary parameters of the subspace, so that the boundary tolerance of the high accuracy subspace is smaller than that of the low accuracy subspace.

[0024] Preferably, the matching process of the pathological knowledge graph comprises:

[0025] Project the knowledge node features into the same vector space as the fusion feature vector, and calculate the cosine similarity matrix;

[0026] Select the top N knowledge nodes corresponding to the maximum value in the similarity matrix to construct a temporary reasoning path;

[0027] Verify the logical consistency of the auxiliary diagnosis basis through the association relationship of the nodes in the temporary reasoning path.

[0028] Preferably, the operation process of the cross-modal consistency verification module comprises:

[0029] Map the initial classification label to the pathological report text feature space, and extract the text keyword vector;

[0030] Calculate the modal alignment loss value of the text keyword vector and the fusion feature vector;

[0031] When the modal alignment loss value exceeds the fifth critical value, correct the confidence weight of the initial classification label according to the text keyword vector.

[0032] Preferably, the method further comprises a quality control step for the digital pathology image:

[0033] Detect the focal length offset and the staining uniformity index during the image acquisition stage;

[0034] When the focal length offset exceeds the sixth critical value or the staining uniformity index is lower than the seventh critical value, trigger the image reacquisition protocol;

[0035] Standardize the light compensation and color gamut calibration of the image passing the quality control.

[0036] Preferably, the training process of the multi-scale feature fusion network comprises:

[0037] Construct an incremental training data set containing malignant lesion region annotations;

[0038] Calculate the feature distribution dispersion of each class after each training period;

[0039] When the dispersion of a specific class exceeds the eighth critical value, add an adversarial generated sample of the class to the training data set.

[0040] Preferably, the method further comprises a step of generating the explainability of the classification result:

[0041] Ranking the contribution of each dimension in the fusion feature vector to the classification result;

[0042] Mapping the top K dimensions in terms of contribution to the microstructure features in the pathology knowledge graph;

[0043] Generating a diagnosis basis report containing the contribution ranking and microstructure features.

[0044] Preferably, the output process of the final diagnosis classification result comprises:

[0045] Mapping the corrected classification label to the clinical diagnosis code system;

[0046] Generating a structured diagnosis code conforming to the international disease classification standard according to the mapping result;

[0047] Writing the structured diagnosis code into the pathology database after term standardization verification.

[0048] Compared with the prior art, the present application has the following advantages:

[0049] The multi-scale feature fusion network is a highlight of the present method. In traditional pathological image analysis, single-scale feature extraction often cannot fully capture the information in the image. However, the present method can fully obtain the features of the image at different scales by inputting the basic analysis unit into the first feature extraction layer to obtain the primary feature atlas, and then using the secondary feature extraction layer to perform spatial attention weighting on the primary feature atlas to generate a fusion feature vector.

[0050] The spatial attention weighting mechanism is like a pair of "smart glasses" for pathological image analysis, which can focus on the key areas in the image and highlight the features that are most important for diagnosis. Taking the identification of tumor cells as an example, traditional methods may be disturbed by the surrounding normal tissues, making it difficult to accurately determine the morphology and features of tumor cells. However, in the present method, the spatial attention weighting mechanism can automatically enhance the feature expression of the tumor cell region and weaken the interference of normal tissues, making the features of tumor cells more prominent, thereby making the subsequent analysis and diagnosis more accurate. Compared with traditional feature extraction methods, the present method can more comprehensively and accurately reflect the essential features of pathological images, providing a more solid foundation for subsequent diagnosis.

[0051] The establishment of the dynamic classification threshold pool is another innovation of the method. It divides the feature subspace according to the dimension distribution of the fusion feature vector and assigns independent classification boundary parameters to each feature subspace. In actual pathological image diagnosis, the features of different types of pathological images differ greatly, and there may be subtle differences in the same type of pathological image in different patients. The traditional fixed classification threshold method is difficult to adapt to such complex and variable conditions, and is prone to misdiagnosis or missed diagnosis.

[0052] The dynamic classification threshold pool is like an "intelligent classifier" that can flexibly adjust the classification standard according to the specific features of the pathological image. For some common pathological images with obvious features, it can use a more stringent classification threshold to ensure the accuracy of the diagnosis; while for some atypical and complex pathological images, it can automatically relax the classification threshold to avoid missed diagnosis. When diagnosing rare disease pathological images, since their features are very different from common diseases, the traditional method may not be able to accurately judge due to the fixed classification threshold. The dynamic classification threshold pool can assign appropriate classification boundary parameters according to the unique feature subspace of the rare disease pathological image, thereby achieving accurate classification and greatly improving the classification ability of complex pathological images. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The working principle diagram of the biopsy pathological image intelligent diagnosis classification method is described.

[0054] Figure 2 The flowchart for morphological feature region extraction is described.

[0055] Figure 3 The flowchart for multi-scale feature fusion network construction is described. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Please refer to Figure 1The digital pathology image data of the biopsy sample is collected by a high-resolution scanning device, and the image data is stored in a full-slice digital format. Morphological feature regions in the image are extracted as basic analysis units, which refer to significant regions containing cell morphology and tissue structure. A multi-scale feature fusion network is constructed, and the basic analysis units are input into the first feature extraction layer of the multi-scale feature fusion network. The first feature extraction layer outputs a primary feature atlas. The secondary feature extraction layer performs spatial attention weighting processing on the primary feature atlas. The spatial attention weighting is based on the importance of the spatial position of the feature atlas to assign a weight coefficient, and generates a fusion feature vector. A dynamic classification threshold pool is established. The dynamic classification threshold pool divides feature subspaces according to the dimensional distribution characteristics of the fusion feature vector. Each feature subspace corresponds to an independent classification boundary parameter. A pre-trained pathology knowledge graph is loaded. The pathology knowledge graph contains pathological concept nodes and relationship edges. The fusion feature vector and the node features in the pathology knowledge graph are matched for similarity calculation. Knowledge nodes with a matching degree exceeding a preset threshold are selected as auxiliary diagnosis basis. The output result of the dynamic classification threshold pool is combined with the auxiliary diagnosis basis to generate an initial classification label. The initial classification label is input into a cross-modal consistency verification module. The cross-modal consistency verification module logically corrects the initial classification label, and outputs a final diagnosis classification result.

[0058] Embodiment 1: see Figure 2The extraction process of morphological feature regions adopts an adaptive threshold segmentation algorithm to divide the digital pathology image into regions. The adaptive threshold segmentation algorithm dynamically calculates the segmentation threshold based on the local gray level characteristics of the image, and analyzes the gray level distribution characteristics in the neighborhood of each pixel to adapt to the light and dark changes of different regions of the image. This local adaptive method can effectively deal with the brightness fluctuation problem caused by uneven staining and slice thickness differences. The entire digital pathology image is divided into multiple candidate regions with relatively uniform characteristics. The boundary determination of the candidate region depends on the coordination of the local threshold and the global gray level distribution, avoiding the generation of too many fragmented small regions. Each candidate region needs to meet the minimum area constraint and shape continuity condition. The nucleus-cytoplasm ratio and texture complexity index of each candidate region are calculated by a special image analysis algorithm. The calculation of the nucleus-cytoplasm ratio requires accurate segmentation of the cell nucleus region and the cytoplasm region. This process uses a semantic segmentation network based on deep learning to complete the fine division of cell structures. The semantic segmentation network uses an encoder-decoder architecture. The encoder part uses a pre-trained ResNet network to extract multi-level features, and the decoder part gradually restores the spatial details through deconvolution and skip connection. The network outputs a probability map of each pixel belonging to the nucleus, cytoplasm or background. The binary segmentation result is obtained by probability threshold processing. The texture complexity index is calculated by a plurality of statistics of the gray level co-occurrence matrix, including contrast, correlation and entropy value parameters. These parameters describe the texture characteristics of the image from different angles. The contrast reflects the clarity of the image, the correlation represents the directionality of the texture, and the entropy value reflects the randomness of the texture. The weighted combination of the three parameters constitutes the comprehensive measurement of the texture complexity.

[0059] The region with the nuclear-cytoplasmic ratio greater than the first critical value and the texture complexity index less than the second critical value is marked as an effective morphological feature region. The first critical value is set by referring to the professional consensus of pathologists on the morphological changes of diseased cells and the distribution range of normal and abnormal nuclear-cytoplasmic ratios based on statistical analysis of a large number of clinical cases. The second critical value is determined according to the difference in texture features between normal and diseased tissues. By comparing the distribution characteristics of texture parameters in benign and malignant lesion regions, the threshold point with the best discrimination ability is selected. The optimization process of the critical value uses the receiver operating characteristic curve analysis to maximize the balance point of sensitivity and specificity. Edge gradient detection is performed on adjacent effective morphological feature regions. Sobel operator is used to calculate the gradient amplitude and direction of the region boundary pixels, thereby quantifying the morphological continuity between regions. Sobel operator includes horizontal and vertical convolution kernels to detect gradient changes in the x and y axes of the image. The gradient amplitude calculates the Euclidean norm of the two directional derivatives, and the gradient direction is calculated by the inverse tangent function. Continuous regions with a gradient difference less than a third critical value are merged. The merging operation is realized by the closing operation in morphological processing. The size of the structure element of the closing operation is configured according to the typical scale of the tissue structure. The selection of the structure element considers the size and arrangement of cells. The diameter of the circular structure element is usually set to 10-20 pixels, which adapts to the cell size under different magnifications.

[0060] The connected regions formed after region merging are used as the basic analysis unit, which needs to meet the minimum area threshold and the morphological regularity requirement to exclude regions with small area or abnormal shape. The minimum area threshold is determined according to the image resolution and clinical significance, usually set to more than 1000 pixels to avoid small regions with limited diagnostic value. The morphological regularity is evaluated by calculating the circularity and Euler number of the region. The circularity measures the degree of closeness to a circle, and the Euler number describes the topological structure of the region. The adaptive threshold segmentation algorithm is implemented using a sliding window mechanism. The size of the sliding window is adaptively adjusted according to the image resolution and cell density. High-resolution images use larger window sizes, and low-resolution images use smaller window sizes. The window size range is usually between 128x128 pixels and 512x512 pixels. The cell density is estimated by the number of cell nuclei per unit area. High cell density areas use smaller windows to obtain more detailed segmentation results. The threshold value in each window is calculated using the maximum inter-class variance method to maximize the separation of foreground and background. The maximum inter-class variance method iteratively searches for the optimal threshold to maximize the variance between the foreground and background classes, achieving optimal segmentation results.

[0061] The segmented region boundary is smoothed to eliminate the interference of jagged edges on subsequent feature extraction. The smoothing process uses a Gaussian filter combined with morphological operations. The kernel size of the Gaussian filter is adjusted according to the roughness of the edge, and the morphological opening operation removes small burrs, and the closing operation fills the concave part. After the boundary is smoothed, an edge tracking algorithm is used to extract the closed contour line to ensure that each region has a complete and continuous boundary representation. The calculation of the nucleus-cytoplasm ratio depends on the accurate segmentation of the nucleus and cytoplasm. The nucleus segmentation uses a deep learning model based on the U-Net architecture. The deep learning model is trained on a large number of labeled cell images and can accurately identify the nucleus boundary under different staining conditions. The encoder part of the U-Net model is composed of multiple convolutional layers and pooling layers, which gradually extract high-level semantic features. The decoder part restores the spatial information through upsampling and jump connection, and realizes accurate pixel-level classification. The training data includes images of hematoxylin-eosin staining, immunohistochemical staining and other staining methods, which enhances the generalization ability of the model. The division of the cytoplasmic region is determined by the area between the expanded contour of the nucleus and the cell membrane boundary. The expansion of the nucleus uses morphological dilation operation, and the dilation radius is set according to the cell type and empirical value, usually 3-5 pixels. The calculation of the texture complexity index is based on the multi-scale gray level co-occurrence matrix. The gray level co-occurrence matrix is calculated in four directions (0°, 45°, 90°, 135°), and the feature values at different distances are extracted. Multi-scale analysis uses three different distance parameters (1, 3, 5 pixels) to capture texture features from micro to macro. The gray level co-occurrence matrix at each scale calculates four statistics: contrast, correlation, energy, and homogeneity. These statistics describe the texture features from different angles. Contrast measures the degree of gray level change, correlation reflects the linear dependence of gray level, energy represents the uniformity of texture, and homogeneity embodies the local consistency of gray level distribution.

[0062] The integration of texture features adopts principal component analysis method to reduce feature dimension while preserving the most discriminative texture information. Principal component analysis projects multiple texture features to a new feature space, and selects the principal components with cumulative contribution rate exceeding 85% as the final feature representation. After feature dimension reduction, clustering analysis is used to verify the texture features' ability to distinguish tissue types, and the feature selection strategy is optimized. The determination of the first critical value and the second critical value is optimized by grid search combined with cross-validation to maximize the classification performance of the effective morphological feature region on the training data set. Grid search systematically tests different critical value combinations within a predefined parameter range, and cross-validation uses a five-fold cross-validation method to ensure the stability of the evaluation results. The performance evaluation uses F1 score as the main indicator, balancing the influence of precision and recall. In the process of edge gradient detection, the Sobel operator is convolved in the horizontal and vertical directions respectively, and the gradient amplitude is calculated by the square root of the sum of the squares of the two direction results. Before convolution, the image is preprocessed, including Gaussian smoothing to remove noise and avoid noise interference in gradient calculation. The gradient direction is used to determine the direction of the edge, which is calculated by the angle of the gradient vector, and the angle value is normalized to the range of 0-360 degrees.

[0063] The merging decision of adjacent regions comprehensively considers the similarity of gradient amplitude and direction. The similarity of gradient amplitude is measured by Euclidean distance, and the similarity of gradient direction is calculated by angle difference. The merging condition of two regions needs to meet the amplitude similarity and direction similarity thresholds at the same time to avoid false merging of regions with obvious texture feature differences. The third critical value is set based on the statistical distribution of the edge intensity between regions, and a compromise value is selected to balance over-merging and under-segmentation. The statistical distribution analysis uses kernel density estimation method, and the critical value is set at the inflection point of the distribution curve. The morphological closing operation uses a circular structural element with a radius proportional to the average size of the cells. The closing operation can fill small holes inside the region and smooth the edges. The radius of the structural element is determined by analyzing the area and perimeter relationship of the cell region to ensure that it can effectively connect adjacent cell regions without damaging the original morphology. After closing operation, a hole filling operation is performed to ensure that there are no blank points inside the region. The screening criteria of the basic analysis unit include the comprehensive evaluation of geometric features and texture features. Geometric features include area, perimeter, circularity, and other parameters. The circularity calculation formula is 4π times the area divided by the square of the perimeter, which measures the compactness of the region. Texture features analyze the uniformity and consistency inside the analysis unit. Uniformity is measured by the variance of the gray scale distribution, and consistency is evaluated by the histogram distribution of the local binary pattern. The comprehensive evaluation uses linear discriminant analysis to project geometric features and texture features to a classification space, and selects according to the projection score.

[0064] After each basic analysis unit is generated, its position coordinates, morphological feature descriptors, and quality evaluation scores are attached. The position coordinates record the image coordinates and bounding box information of the region center. The morphological feature descriptors include affine invariant features such as Hu moments and Zernike moments, which ensure the invariance of the features to rotation and scaling. The quality evaluation scores are calculated based on the contrast, integrity, and consistency of the region, and low-quality regions are marked as unreliable samples. These structured information provides a standardized input data format for the subsequent multi-scale feature fusion network, ensuring data consistency and comparability. Quality monitoring indicators include segmentation success rate, feature stability, and other parameters, and abnormal situations trigger reprocessing or manual intervention mechanisms. The automated conversion system establishes error detection and correction mechanisms, continuously optimizes the processing flow through log recording and performance statistics. The system supports batch processing and high-concurrency computing, meeting the needs of large-scale clinical applications, with a processing speed of dozens of high-resolution pathology images per minute.

[0065] Example 2: see Figure 3 The construction process of the multi-scale feature fusion network sets parallel convolution kernel groups in the first feature extraction layer, each convolution kernel group containing different sizes of dilated convolution kernels. The dilation rates of the dilated convolution kernels are carefully configured to adapt to the feature extraction needs of multi-scale biological structures in pathology images. The dilated convolution kernel with a dilation rate of 1 corresponds to the basic cell local feature perception, whose receptive field covers a single cell nucleus and its adjacent cytoplasmic region, and can capture microscopic features such as chromatin distribution and nuclear membrane morphology. The dilated convolution kernel with a dilation rate of 3 expands the receptive field to the cell cluster level, perceiving the spatial arrangement relationship between multiple cells, the connection state between cells, and the architectural features of local tissues. The dilated convolution kernel with a dilation rate of 6 corresponds to larger-scale tissue region features, which can identify macroscopic tissue structure patterns such as glandular structures, interstitial distribution, and vascular networks. This multi-scale parallel design enables the network to simultaneously process pathology features from micro to macro, providing a comprehensive information foundation for subsequent precise diagnosis.

[0066] The base analysis unit inputs to the parallel convolution kernel group in the standardized three-channel image format, and each convolution kernel group independently extracts features and outputs a 256-dimensional feature map. The input image is preprocessed before convolution operation, including normalizing the pixel value to the [0, 1] interval, and applying a color consistency correction algorithm to eliminate the influence of staining differences. Different sizes of the hole convolution kernel use a parallel processing architecture, and the three convolution paths share the same input tensor but use different convolution parameters to ensure the independence of feature extraction at different scales. Each convolution kernel group internally contains two consecutive convolution layers, the first layer uses 3x3 basic convolution for feature transformation, and the second layer applies a specified dilation rate of the hole convolution to expand the receptive field. After the convolution operation, the ReLU activation function is used to introduce a nonlinear transformation, and the batch normalization layer is used to stabilize the training process. The generation of the primary feature map is realized by splicing the output feature maps of the three convolution kernel groups in the channel dimension, forming a 768-dimensional fused feature representation. The feature splicing operation preserves the feature information extracted at different scales, allowing the subsequent network layers to learn the relevance between different scale features. The spliced feature map is compressed and reorganized in the channel dimension through 1x1 convolution, reducing the computational complexity while enhancing the feature expression ability. The spatial resolution of the primary feature map is maintained through the step size setting in the initial convolution layer, ensuring that the key spatial information is not lost.

[0067] The secondary feature extraction layer integrates channel attention mechanism and spatial attention mechanism to form a dual attention network structure. The calculation of channel attention weight uses global average pooling to compress the spatial dimension, reducing the two-dimensional feature map of each channel to a single scalar value, which represents the global distribution characteristics of the channel features. The pooled channel descriptor is subjected to nonlinear transformation through two fully connected layers, the first fully connected layer reduces the dimension to 1 / 16 of the original channel number, and the second fully connected layer restores the original channel number. The Sigmoid activation function maps the output value to the [0, 1] interval to generate the weight coefficient of each channel, and the size of the weight coefficient reflects the importance of the channel feature to the current task. The calculation of spatial attention weight is based on the spatial position response of the primary feature map, which is compressed into a single-channel spatial saliency map through 1x1 convolution. The parameters of the 1x1 convolution kernel are learned through training, which can automatically identify the discriminative spatial regions in the feature map. The Softmax function normalizes the spatial saliency map so that the sum of the weights of all spatial positions is 1, highlighting important regions while suppressing irrelevant background interference. The spatial attention weight map has the same spatial size as the input feature map, and the weight value at each position reflects the contribution of each position to feature expression.

[0068] The channel attention weight and the spatial attention weight are point multiplied to realize the effective fusion of the two attention mechanisms. The point multiplication operation multiplies the elements to generate a feature enhancement coefficient matrix, which contains attention information of both channel and spatial dimensions. The dimension of the feature enhancement coefficient matrix completely corresponds to the primary feature map, and each spatial position and channel has an independent enhancement coefficient, forming a fine-grained feature selection mechanism. The application of the enhancement coefficient is realized through the broadcast mechanism to ensure the accuracy of the dimension matching. The channels of the primary feature map are nonlinearly scaled with the feature enhancement coefficient. The nonlinear scaling is realized by element-by-element multiplication, that is, each element of the feature map is multiplied by its corresponding enhancement coefficient. This scaling operation can adaptively enhance useful features and suppress noise features, improving the quality of feature expression. The scaled feature map is normalized by LayerNormalization, which calculates the mean and variance along the feature channel dimension to reduce the internal covariate shift problem. The normalized feature map is introduced through the GeLU activation function to introduce a smoother nonlinear transformation, enhancing the expression ability of the network. The generation of the fusion feature vector is compressed into a 1024-dimensional vector representation through global average pooling, which averages all spatial positions of each feature channel to obtain the global statistics at the channel level. The pooled feature vector is transformed in dimension through a fully connected layer, converting a 768-dimensional input into a 1024-dimensional output to increase the expression ability of the features. The weight matrix of the fully connected layer is initialized orthogonally to ensure training stability, and the bias term is initialized to zero. The output feature vector is L2 normalized to make its norm uniform to 1, facilitating subsequent similarity calculation and classification decision.

[0069] The training process of the multi-scale feature fusion network constructs an incremental training dataset containing malignant lesion region annotations. The dataset construction follows strict medical image annotation standards. The incremental training dataset initially contains one hundred thousand annotated images, each annotated by three pathologists independently. The annotation results are subjected to consistency testing, and only annotations agreed upon by at least two experts are adopted. The annotated data includes pixel-level lesion region segmentation masks and image-level diagnosis labels, providing multi-level supervision information. The class distribution of the dataset is balanced through oversampling and undersampling techniques to adjust the number of samples in each class, avoiding classifier bias towards the majority class. The feature distribution dispersion of each class is calculated after each training period. The Mahalanobis distance is used to measure the dispersion of samples of the same class in the feature space. The Mahalanobis distance considers the feature covariance matrix, accurately reflecting the degree of aggregation of intra-class samples. The dispersion is calculated based on the feature vectors of all samples in the training set, and the centroid and scatter matrix of each class of samples are calculated separately. The dispersion threshold is dynamically adjusted according to the number of samples and the difficulty of classification. Classes with fewer samples allow for greater dispersion, while classes with more samples require tighter feature distribution.

[0070] When the dispersion of a specific category exceeds the eighth critical value, adversarial generated samples of that category are added to the training dataset. These adversarial generated samples are synthesized using a conditional generative adversarial network (GAN). The generator network adopts a U-Net architecture, taking category labels and random noise as input, and generates synthetic images through an encoder-decoder structure. The discriminator network uses a convolutional neural network structure to distinguish between real and generated images. During adversarial training, the generator and discriminator are optimized alternately, until the generator can produce synthetic images that are highly similar to the distribution of real samples. The proportion of adversarial generated samples added is dynamically adjusted according to the dispersion value; the higher the dispersion, the larger the proportion added, but the maximum proportion does not exceed 20% of the total training set to avoid synthetic samples dominating the training process. The optimizer used during training is the AdamW algorithm, which introduces a weight decay regularization term on top of the Adam optimizer to prevent overfitting. The initial learning rate is set to 0.001, and a cosine annealing strategy is used to dynamically adjust the learning rate. The learning rate decays according to a cosine function with the number of training epochs, reaching one percent of the initial value at the end of training. Cosine annealing allows the learning rate to be periodically restarted, helping the model escape local optima. The loss function is designed as a weighted sum of the focal loss and the center loss, with the focal loss's focusing parameter set to 2 to balance the gradient contributions of easily classified and hard-classified samples. The weight coefficients of the center loss are determined through cross-validation to ensure that the magnitudes of the two losses are matched. The batch size is set to 32, and the number of training epochs is set to 200. After each training epoch, the model performance is evaluated on an independent validation set. The validation set is randomly partitioned from the training data, accounting for 20% of the total samples. Model performance is evaluated using weighted F1 score and AUC value as the main metrics. The weighted F1 score considers class imbalance, and the AUC value measures the model's overall discriminative ability. The convergence of the training loss and validation loss is monitored simultaneously during training. An early stopping mechanism is triggered when the validation loss stops decreasing for several consecutive epochs.

[0071] The initialization of network weights adopts the He normal distribution initialization method, which is particularly suitable for ReLU series of activation functions, and can maintain the stability of gradients in the forward and backward propagation processes. The bias term of the convolutional layer is initialized to zero to avoid introducing unnecessary bias. The fully connected layer in the attention mechanism uses Xavier uniform initialization to ensure that the distribution of activation values remains stable throughout the network. This differentiated initialization strategy adapts to the characteristics of different network components. During training, various data augmentation techniques are applied, including random horizontal flipping, random rotation, color jittering, and elastic deformation. The probability of random horizontal flipping is set to 0.5, the angle range of random rotation is [-30°, 30°], and color jittering includes random adjustments of brightness, contrast, and saturation. Elastic deformation simulates the natural deformation of tissue sections, enhancing the model's invariance to geometric transformations. Data augmentation operations are performed in parallel on the GPU, reducing the preprocessing overhead during training. Gradient clipping technique is applied during training, with a gradient norm threshold of 1.0. Before each parameter update, the L2 norm of the gradient is calculated, and if it exceeds the threshold, the gradient is scaled proportionally. Gradient clipping prevents gradient explosion, especially when training deep networks to maintain numerical stability. The early stopping mechanism is triggered based on the continuous ten rounds of validation set performance without improvement, and the model parameters with the best validation performance are saved.

[0072] The inference stage of the multi-scale feature fusion network uses full-precision calculation mode to ensure that the calculation accuracy meets the requirements of medical diagnosis. The input image is preprocessed and converted to a floating-point tensor format, and the preprocessing steps are consistent with the training stage, including color normalization and size standardization. During the forward propagation process of the network, the intermediate feature maps are stored using half-precision floating-point numbers, reducing memory usage while maintaining sufficient numerical accuracy. The output layer uses full-precision calculation to ensure the accuracy of the final results. The computational graph structure of the network uses static graph optimization, which analyzes and reconstructs the computational graph before inference. Operator fusion technology combines multiple consecutive operations into a single kernel function, reducing memory access overhead. Common fusion modes include convolution and normalization layer fusion, activation function and subsequent operation fusion. Memory reuse technology pre-allocates a pool of video memory, which is reused during the inference process, reducing the overhead of dynamic memory allocation. These optimization measures significantly improve the inference speed, meeting the real-time requirements of clinical practice. The multi-scale feature fusion network also integrates online learning capabilities, supporting the learning of new case data after deployment. The online learning process uses a small batch update strategy, using a small number of new samples for incremental training each time. The elastic weight consolidation technique is applied during model updating to protect learned knowledge from being forgotten. The online learning module includes strict quality control mechanisms, and new samples must be confirmed by multiple experts before being used for model updating to ensure the reliability of the learning process. This design enables the system to continuously improve and adapt to changes in pathological diagnosis practice.

[0073] Embodiment 3: Operation process of dynamic classification threshold pool Principal component analysis is performed on the fusion feature vector, which projects the high-dimensional feature vector into a low-dimensional space and retains the main variation direction of the data. The fusion feature vector is taken as a row vector of the input matrix, and the eigenvalues and eigenvectors of the covariance matrix are calculated. The principal component axes with a variance contribution rate exceeding a fourth critical value are selected, and the fourth critical value is set to 85%, to ensure that the feature space after dimension reduction retains sufficient information. The number of principal component axes is automatically determined according to the cumulative contribution rate of the eigenvalues, and the minimum number of principal axes that meet the cumulative contribution rate greater than the fourth critical value is adopted. The projected feature vector has orthogonality and unit norm, and the feature dimension is reduced to between 20% and 30% of the original dimension. The feature space is divided into hypercube subspaces along the principal component axis direction, and the hypercube subspaces are divided using the equal frequency binning method, with each dimension divided into eight intervals, so that each interval contains approximately the same number of training samples. The numbering of the hypercube subspaces uses a multi-dimensional indexing mechanism, and each subspace has a unique identifier. The classification accuracy of historical samples in each subspace is calculated, and the calculation of the classification accuracy is based on the last 1,000 samples passing through the subspace. The proportion of correctly classified samples is taken as the accuracy indicator of the subspace. The storage of historical samples uses a sliding window mechanism, and the oldest sample is removed when a new sample enters, to maintain the timeliness of the statistics.

[0074] The classification boundary parameters of the subspace are dynamically adjusted according to the classification accuracy, including the offset of the decision boundary and the confidence threshold. The boundary tolerance of the high-accuracy subspace is set to a small value, and the boundary tolerance is inversely proportional to the classification accuracy. The adjustment process uses an exponential weighted moving average method, and the new accuracy indicator updates the boundary parameters with a weight of 0.1. The frequency of dynamic adjustment is set to once every 100 new samples to avoid frequent adjustments that cause system instability.

[0075] The decision function of the dynamic classification threshold pool uses the following formula:

[0076]

[0077] Wherein: represents the classification result, is a sign function, is the number of support vectors, is a Lagrange multiplier, is a kernel function, is a support vector, is a feature vector of the sample to be classified, is an adjusted bias term. The kernel function is selected as a radial basis function, and the scale parameter is automatically adjusted according to the density of the feature space.

[0078] The matching process of the pathology knowledge graph projects the knowledge node features into the same vector space as the fusion feature vector. The projection operation uses a linear transformation matrix, which is learned through canonical correlation analysis. The knowledge node features come from structured knowledge in pathology textbooks, expert consensus, and clinical guidelines. A cosine similarity matrix is calculated, with rows corresponding to fusion feature vectors and columns corresponding to knowledge node features. Each element represents the directional similarity between two vectors. The top N knowledge nodes corresponding to the maximum value in the similarity matrix are selected to construct a temporary reasoning path. The value of N is set to five to balance the comprehensiveness and efficiency of reasoning. The construction of the temporary reasoning path uses a breadth-first search algorithm with a search depth limit of three relationships. The association between knowledge nodes includes parent-child relationships, parallel relationships, and causal relationships. The association between nodes in the temporary reasoning path verifies the logical consistency of the diagnostic basis. Logical consistency checks are based on description logic reasoning rules to check for contradictory relationship connections in the path. The update mechanism of the pathology knowledge graph is set to synchronize the latest medical research progress every month. New pathological concepts and relationships are added to the graph after expert review. The feature vectors of knowledge nodes are trained using deep knowledge representation learning techniques, integrating multi-modal information from text descriptions and image features. The similarity calculation uses an approximate nearest neighbor search algorithm with a hierarchical navigable small-world graph index to speed up the retrieval process. The scoring function of the temporary reasoning path considers path length, node authority, and relationship strength factors to select the optimal reasoning path to support diagnostic decision-making. The dynamic classification threshold pool cooperates with the pathology knowledge graph using an asynchronous communication mechanism. Feature space partitioning and knowledge matching are executed in parallel. When the classification accuracy of a subspace consistently falls below the threshold, the knowledge graph enhancement verification process is triggered, and the classification boundary is corrected using domain knowledge. The system maintains a feedback loop mechanism, and misclassified samples are confirmed by experts to update the statistics of the dynamic classification threshold pool and the weight parameters of the pathology knowledge graph.

[0079] The operation process of the cross-modal consistency verification module maps the initial classification label to the pathology report text feature space. The mapping operation uses a label embedding model based on an attention mechanism. The initial classification label is input into the label embedding model in one-hot encoding form, and the label embedding model outputs a 128-dimensional text-oriented feature vector. The pathology report text feature space is constructed through clinical natural language processing technology, and the semantic representation of the pathology description text is extracted using a medical pre-trained language model. The extraction of the text keyword vector uses a term weighting algorithm to calculate the TF-IDF weight of each medical term in the pathology report, and selects the top ten terms with the highest weight to form the keyword vector. The modal alignment loss value of the text keyword vector and the fusion feature vector is calculated using a bilinear matching network, which learns the correspondence between the two modal features. The calculation of the modal alignment loss value is based on the weighted combination of the cosine distance and the Euclidean distance of the feature vectors, and the weight coefficient is optimized through the training process. When the modal alignment loss value exceeds the fifth critical value, the confidence weight correction mechanism of the initial classification label is triggered, and the correction mechanism uses a gated recurrent unit network to model the time series dependency relationship. The adjustment of the confidence weight is based on the semantic consistency score of the text keyword vector and the fusion feature vector, and the scoring function considers the semantic relevance and context matching degree of the terms. The quality control steps of the digital pathology image are monitored in real time during the image acquisition stage, including the focal length offset and the staining uniformity index. The detection of the focal length offset is achieved by analyzing the high-frequency component power spectrum of the image. The calculation of the staining uniformity index is based on the variance analysis of the image color channel, which evaluates the consistency level of hematoxylin-eosin staining. When the focal length offset exceeds the sixth critical value or the staining uniformity index is lower than the seventh critical value, the image acquisition system automatically triggers the reacquisition protocol, which includes adjusting the microscope focal length and re-preparing the staining sample.

[0080] The images that pass the quality control are subjected to standardized illumination compensation and color gamut calibration. The illumination compensation algorithm uses a homomorphic filtering technique to separate the illumination component and the reflection component. The color gamut calibration uses the reference color values of the standard color card to establish a color transformation matrix, which corrects the color deviation caused by differences in staining conditions. Refer to Table 1 for the specific settings of the quality control parameters.

[0081] Table 1: Image quality control parameter setting table

[0082]

[0083] The starting conditions of the image reacquisition protocol include the comprehensive judgment of hardware state monitoring and image quality evaluation results, and the hardware state monitoring covers the cleanliness of the microscope objective, the intensity of the light source and the performance of the camera sensor. The reacquisition times are limited to three times, and the number of times exceeding the threshold value triggers the manual intervention process, and the technical personnel check the equipment state and sample preparation conditions. The intensity parameter of the standardized light compensation is adjusted adaptively according to the histogram characteristics of the image brightness distribution, avoiding excessive compensation leading to detail loss. The reference standard of color gamut calibration uses the digital pathology image color standard issued by the International Society of Pathology to ensure the color consistency of different batches of images. The text feature extraction process of the cross-modal consistency verification module includes three steps of report analysis, term standardization and vectorized representation. The report analysis uses a rule engine to identify key information segments in the pathology report. The term standardization maps the medical terms described in free text to the standard concepts of the Unified Medical Language System, eliminating the influence of expression differences. The vectorized representation uses a deep learning language model to generate contextualized embeddings, capturing the semantic nuances of terms. The threshold value of the modal alignment loss value is set based on ROC curve analysis on the validation set, selecting the critical point that optimizes the classification performance.

[0084] The iterative optimization process of the confidence weight correction algorithm uses a reinforcement learning framework to model the weight adjustment decision as a Markov decision process. The reward function considers the improvement of classification accuracy and the calibration degree of model confidence, and learns the optimal correction strategy through a policy gradient method. The quality control database of digital pathology images records detailed parameters of each quality detection, and establishes a correlation model between quality indicators and diagnostic accuracy. The dynamic adjustment mechanism of quality control parameters regularly updates the critical value based on the learning results of historical data, to adapt to the effects of equipment aging and reagent batch changes. The calibration and maintenance plan of the image acquisition system is based on the trend analysis of quality control data, and the preventive maintenance period is determined according to the equipment usage frequency and quality index drift speed. The reference color card of color gamut calibration is detected for optical properties every month to ensure the accuracy of the color reference value. The working log of the cross-modal consistency verification module and the quality control system is integrated into the audit tracking system, recording detailed information and adjustment results at each decision point to support traceability analysis of the diagnosis process.

[0085] The contribution ranking of each dimension in the fusion feature vector to the classification result is calculated by the integrated gradient algorithm, which integrates the gradient value along the path from the baseline to the actual feature vector. Each dimension of the fusion feature vector corresponds to the high-level abstract features of a specific feature map in the multi-scale feature fusion network, such as nuclear morphological abnormalities, cell arrangement disorder, or interstitial fibrosis patterns. The top K dimensions in the contribution ranking are selected as the key decision features, and the value of K is dynamically adjusted according to the required depth of interpretability, with a conventional diagnostic setting of ten to fifteen key features. The semantic analysis of the key decision features is achieved by querying the feature dictionary in the pathology knowledge graph, which contains thousands of pre-defined microscopic structure feature descriptors.

[0086] The top K dimensions in the contribution ranking are mapped to the microscopic structure features in the pathology knowledge graph, and the mapping process uses a cross-modal alignment model based on attention mechanism. The microscopic structure features include cell atypia degree, mitotic figure count, and glandular structure integrity, among other morphological indicators, each of which has standardized description text and typical image examples. The mapping model calculates the semantic similarity between the feature dimensions and the microscopic structure features, and selects the microscopic structure feature with the highest similarity as the pathological explanation of the model feature dimension. For example, a high-contribution feature dimension may be mapped to the microscopic structure feature of "significant cell nucleus polymorphism", and another dimension may correspond to the histological change of "back-to-back glandular infiltration". A diagnosis basis report containing contribution ranking and microscopic structure features is generated, and the structured natural language generation technology is used to combine the standard report template with dynamic content. The report starts with the diagnostic conclusion and confidence score, and the main part details the pathological significance of each key feature and its impact weight on the diagnostic decision in descending order of contribution. The report ends with the relevant image region index supporting the diagnosis, which facilitates the review of key morphological basis by pathologists. The diagnosis basis report is output in HTML and PDF formats simultaneously, supporting interactive viewing of feature positioning information.

[0087] The output process of the final diagnosis classification result maps the revised classification labels to a clinical diagnosis code system, which adopts the coding rules of the International Classification of Diseases, Tenth Revision (ICD-10) and the International Classification of Diseases for Oncology, Third Edition (ICD-O-3). The mapping table contains thousands of entries, covering various pathological diagnoses from benign lesions to malignant tumors. For example, intraductal papilloma of the breast with atypical hyperplasia is mapped to ICD-10 code D05.9 and ICD-O-3 code 8500 / 2. The mapping process uses a rule-based reasoning engine to handle synonyms and hierarchical relationships of diagnosis terms. According to the mapping result, structured diagnosis codes conforming to the international disease classification standard are generated, which contain fields such as anatomical site, morphological type, behavior code, and differentiation level. The code generator verifies the consistency of the diagnosis term with the coding rules and automatically fills in the required code fields. For example, a diagnosis of gastric adenocarcinoma will generate an anatomical site code of C16.9, a morphological code of 8140 / 3, and a behavior code of M-81403. The structured diagnosis code is checked for term standardization, and the rule library contains tens of thousands of term standardization rules to ensure the accuracy of the code.

[0088] After the term standardization check, the structured diagnosis code is written to the pathology database, which uses a relational database management system to store diagnosis records. The database table structure includes modules such as patient basic information, specimen information, diagnosis content, and quality indicators. The write operation follows the database transaction characteristics of atomicity, consistency, isolation, and durability to ensure data integrity. The database index strategy is optimized for common query conditions to speed up the retrieval and statistical analysis of diagnosis records. The database access interface uses role-based permission control to protect the security of patient privacy data. The visualization display of the diagnosis report integrates a digital pathology image viewer, which highlights the image areas corresponding to key features. Pathologists can click on the feature descriptions in the report to directly locate the relevant areas in the image, enabling the linkage viewing of text descriptions and visual evidence. The report system records audit logs for each access to meet the traceability requirements of medical quality management. The structured diagnosis code is synchronously uploaded to the hospital information system to participate in the generation of clinical pathway management and disease statistical reports.

[0089] It should be noted that, in the present document, relational terms are used only to establishe a relationship of one entity or action to another entity or action, and do not necessarily imply a specific order or need of one entity or action to the other. Moreover, the terms "comprises", "comprising", or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0090] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A biopsy pathological image intelligent diagnosis classification method, characterized in that, The method comprises the following steps: Collecting digital pathology image data of biopsy samples, extracting morphological feature regions in the image as basic analysis units; Constructing a multi-scale feature fusion network, inputting the basic analysis units into the first feature extraction layer to obtain primary feature maps, and performing spatial attention weighting on the primary feature maps through the secondary feature extraction layer to generate a fusion feature vector; Establishing a dynamic classification threshold pool, dividing feature subspaces according to the dimension distribution of the fusion feature vector, and assigning independent classification boundary parameters to each feature subspace; Loading a pre-trained pathology knowledge graph, performing similarity matching between the fusion feature vector and the node features in the pathology knowledge graph, and selecting knowledge nodes with a matching degree exceeding a preset threshold as auxiliary diagnostic evidence; Generating an initial classification label based on the output results of the dynamic classification threshold pool and the auxiliary diagnostic evidence, correcting the initial classification label through a cross-modal consistency verification module, and outputting a final diagnostic classification result; The extraction process of the morphological feature region comprises: Using an adaptive threshold segmentation algorithm to divide the digital pathology image into regions, calculating the nucleus-to-cytoplasm ratio and texture complexity index of each region; Marking the region with a nucleus-to-cytoplasm ratio greater than a first critical value and a texture complexity index less than a second critical value as an effective morphological feature region; Performing edge gradient detection on adjacent effective morphological feature regions, and merging continuous regions with a gradient difference less than a third critical value; The operation process of the dynamic classification threshold pool comprises: Performing principal component analysis on the fusion feature vector, and selecting principal component axes with a variance contribution rate exceeding a fourth critical value; Dividing the feature space into hypercube subspaces along the principal component axis direction, and counting the classification accuracy of historical samples in each subspace; Dynamically adjusting the classification boundary parameters of the subspaces according to the classification accuracy, so that the boundary tolerance of the high-accuracy subspace is less than that of the low-accuracy subspace. 2.The biopsy pathological image intelligent diagnosis and classification method according to claim 1, characterized in that, The construction process of the multi-scale feature fusion network comprises: Setting parallel convolution kernel groups in the first feature extraction layer, each convolution kernel group containing different sizes of hole convolution kernels; Point-multiplying the channel attention weight and the spatial attention weight of the secondary feature extraction layer to generate a feature enhancement coefficient; Performing nonlinear scaling on each channel of the primary feature map through the feature enhancement coefficient to generate a fusion feature vector. 3.The intelligent diagnosis and classification method for biopsy pathological images according to claim 2, characterized in that, The matching process of the pathology knowledge graph comprises: Projecting the knowledge node features into the same vector space as the fusion feature vector, and calculating a cosine similarity matrix; Selecting the top N knowledge nodes corresponding to the maximum value in the similarity matrix to construct a temporary reasoning path; Verifying the logical consistency of the auxiliary diagnostic evidence through the association relationship of the nodes in the temporary reasoning path. 4.The method of claim 3, wherein, The operation process of the cross-modal consistency verification module comprises: Mapping the initial classification label to the pathology report text feature space, and extracting a text keyword vector; Calculating the modal alignment loss value of the text keyword vector and the fusion feature vector; When the modal alignment loss value exceeds a fifth critical value, correcting the confidence weight of the initial classification label according to the text keyword vector.

5. The intelligent diagnostic classification method for biopsy pathology images according to claim 4, characterized in that, Further comprising a quality control step for the digital pathology image: Detecting focal offset and staining uniformity indicators during image acquisition; When the focal length offset exceeds the sixth critical value or the dyeing uniformity index is lower than the seventh critical value, an image reacquisition protocol is triggered; The images passing the quality control are subjected to standardization light compensation and color gamut calibration.

6. The intelligent diagnostic classification method for biopsy pathology images according to claim 5, characterized in that, The training process of the multi-scale feature fusion network comprises: An incremental training data set containing malignant lesion region annotations is constructed; The feature distribution dispersion of each category is calculated after each training cycle; When the dispersion of a specific category exceeds the eighth critical value, an adversarial generated sample of the category is added to the training data set.

7. The intelligent diagnostic classification method for biopsy pathology images according to claim 6, characterized in that, An explainability generation step of the classification result is further included: The contribution degree ranking of each dimension in the fusion feature vector to the classification result is recorded; The top K dimensions in the contribution degree ranking are mapped to the microscopic structure features in the pathology knowledge graph; A diagnosis basis report containing the contribution degree ranking and the microscopic structure features is generated. 8.The biopsy pathological image intelligent diagnosis and classification method of claim 7, characterized in that, The output process of the final diagnostic classification result comprises: The revised classification label is mapped to a clinical diagnosis code system; A structured diagnosis code conforming to the international disease classification standard is generated according to the mapping result; The structured diagnosis code is subjected to term standardization verification and written into a pathology database.

Images