A model quantification method and system based on pathological image classification

Abstract

The application provides a model quantification method and system based on pathological image classification, comprising: obtaining an original pathological image, inputting the original pathological image into an anomaly detection model to generate an anomaly heat map, wherein the anomaly detection model comprises a compression anomaly detection submodule and a prediction anomaly detection submodule; performing feature extraction on the original pathological image based on the anomaly heat map to obtain an anomaly region feature dataset; verifying and analyzing the anomaly region feature dataset by using a multi-agent game mechanism to obtain an anomaly analysis dataset; obtaining actual feedback data of the anomaly analysis dataset; and performing feedback optimization based on the actual feedback data to improve the long-term practicability and stability in the auxiliary image classification and archiving tasks.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and more specifically, to a model quantification method and system based on pathological image classification. Background Technology

[0002] Currently, the analysis of digital pathology whole-slide images is playing an increasingly important role in assisting medical research, pathology teaching, and biobank management. Traditional analysis methods mostly rely on manual interpretation by pathologists, which is inefficient and easily affected by subjective factors, or they employ automated analysis methods based on deep learning.

[0003] However, existing traditional methods often have certain limitations. First, when faced with complex and varied pathological abnormality patterns, especially rare or novel patterns, traditional methods have limited generalization ability and are prone to producing high false positive rates. Second, traditional methods are prone to insufficient attention to key areas in images or over-analysis of simple areas. Summary of the Invention

[0004] To address the shortcomings of existing technologies, embodiments of the present invention provide a model quantization method and system based on pathological image classification to solve at least one of the aforementioned technical problems in the prior art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a model quantization method based on pathological image classification, comprising: The original pathological image is acquired and input into the anomaly detection model to generate an anomaly heatmap. The anomaly detection model includes a compressed anomaly detection submodule and a predicted anomaly detection submodule. Based on the aforementioned abnormal heatmap, feature extraction is performed on the original pathological image to obtain an abnormal region feature dataset. A multi-agent game mechanism is used to verify and analyze the feature dataset of abnormal regions, and an anomaly analysis dataset is obtained. Obtain the actual feedback data from the anomaly analysis dataset, and perform feedback optimization based on the actual feedback data.

[0006] In one possible implementation of the first aspect, the anomaly detection model includes: A convolutional neural network is used as the shared feature extraction backbone network of the anomaly detection model. Based on the shared feature extraction backbone network, features are extracted from the original pathological image to obtain the original image feature vector. A convolutional neural network architecture is used to construct a compressed anomaly detection submodule; A self-supervised learning strategy is adopted to train the compressed anomaly detection submodule to predict the conditional probability distribution of pixel blocks in the original pathological image given the feature vector of the original image. The loss function uses negative log-likelihood loss, and the training aims to minimize the negative log probability of the true pixel value under the predicted conditional probability distribution. The predictive anomaly detection submodule is constructed using a U-Net encoder-decoder architecture; A self-supervised learning strategy is adopted, and pre-training is performed using a joint loss function based on L1 loss and structural similarity loss.

[0007] In one possible implementation of the first aspect, the original pathological image is acquired and input into an anomaly detection model to generate an anomaly heatmap, including: The anomaly detection model uses a multi-scale sliding window strategy to extract features from the original pathological images and obtain the feature vector of the original images. The compressed anomaly detection submodule maps the original image feature vector to the conditional probability distribution of each pixel block in the original pathological image; The original pathological image is encoded and compressed based on the conditional probability distribution, and the encoding length of the compressed original pathological image is used as the first compression length. The original pathological image is encoded and compressed using a lossless compression algorithm, and the encoded length of the compressed original pathological image is used as the second compression length. The image compression ratio is obtained based on the first compression length and the second compression length, and compression anomaly detection is performed based on the image compression ratio.

[0008] In one possible implementation of the first aspect, the method further includes: The prediction anomaly detection submodule performs image prediction based on the feature vector of the original image to generate a predicted pathological image; A difference analysis was performed between the predicted pathological image and the original pathological image to obtain the absolute error between the predicted pathological image and the original pathological image. Obtain the structural similarity index between the predicted pathological image and the original pathological image, and construct a prediction difference map based on the structural similarity index and the absolute error; Anomaly detection is performed based on the predicted difference map, and an anomaly heatmap is constructed based on the predicted difference map and the image compression rate.

[0009] In one possible implementation of the first aspect, feature extraction is performed on the original pathological image based on the abnormal heatmap to obtain an abnormal region feature dataset, including: A lightweight encoder-decoder network based on the information bottleneck principle is used to extract deep features from the high-confidence anomaly regions marked in the anomaly heatmap, and the reconstruction error and feature deviation are obtained. The reconstruction error is expressed as the normalized average absolute difference between the input image block and the decoder output block, which is used to reflect the abnormal intensity of the image at the pixel level. The feature deviation is represented by the Mahalanobis distance between the bottleneck feature corresponding to the high-confidence abnormal region and the pre-stored normal prototype center, which is used to characterize the degree of deviation of the current region feature from the normal pattern in the feature space. The reconstruction error, feature deviation, and features contained in the corresponding regions of the anomaly heatmap are output as the anomaly region feature dataset.

[0010] In one possible implementation of the first aspect, a multi-agent game mechanism is used to perform verification analysis on the anomaly region feature dataset to obtain an anomaly analysis dataset, including: The abnormal region feature dataset is injected into the multi-agent resource pool for resource matching, and a candidate agent set is generated for each abnormal region. Obtain the regional feature vectors corresponding to each abnormal region. Each candidate agent in the candidate agent set makes a game bid based on the regional feature vectors to generate a game bid dataset. The VCG mechanism is used to perform multi-objective optimization and decision-making on the game bidding dataset, generating corresponding anomaly analysis decisions. Based on the anomaly analysis decision, the abnormal region is analyzed to obtain an anomaly analysis dataset.

[0011] In one possible implementation of the first aspect, obtaining actual feedback data from the anomaly analysis dataset and performing feedback optimization based on the actual feedback data includes: Interactive image explanation information is generated based on the anomaly analysis dataset, and the interactive image explanation information is transmitted to the interactive terminal, while the actual feedback data submitted by the interactive terminal is collected synchronously. The actual feedback data includes at least anomaly analysis correctness indicators and anomaly analysis correction indicators; Incremental learning and parameter optimization are performed on each sub-module within the anomaly detection model based on the actual feedback data.

[0012] Secondly, embodiments of the present invention also provide a model quantization system based on pathological image classification, comprising: The data acquisition module is used to acquire the original pathological image and input the original pathological image into the anomaly detection model to generate an anomaly heatmap. An anomaly detection model is used to extract features from the original pathological image and to detect anomalies in the original pathological image by combining the compressed anomaly detection submodule and the predicted anomaly detection submodule contained in the model. The feature extraction module is used to extract features from the original pathological image by combining the abnormal heatmap to obtain an abnormal region feature dataset. The verification and analysis module is used to perform verification and analysis on the abnormal region feature dataset according to the multi-agent game mechanism to obtain the abnormal analysis dataset. The feedback optimization module is used to obtain the actual feedback data of the anomaly analysis dataset and perform feedback optimization based on the actual feedback data.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The original pathological images are acquired and input into the anomaly detection model to generate an anomaly heatmap. This step uses a dual-path parallel processing and multi-scale sliding window strategy to quickly screen the entire original pathological slice image while maintaining low computational overhead, reducing unnecessary in-depth analysis of significantly normal areas. At the same time, the generated anomaly heatmap can provide a preliminary reference for the location of anomalies, providing a basis for subsequent analysis.

[0014] Based on the aforementioned abnormal heatmap, feature extraction is performed on the original pathological image to obtain an abnormal region feature dataset. This step involves deep feature extraction and refined quantization of candidate regions, which helps improve the generalization ability for complex and ever-changing pathological abnormality patterns. At the same time, it further locates the key regions of the original pathological image, avoiding the waste of resources on the analysis of non-critical regions, thereby improving the efficiency of resource utilization.

[0015] A multi-agent game mechanism is used to verify and analyze the feature dataset of anomaly regions to obtain anomaly analysis dataset. This step introduces an auction mechanism to tilt verification resources toward the most uncertain or complex regions, which helps to make efficient use of current computing resources. Obtain actual feedback data from the anomaly analysis dataset, perform feedback optimization based on the actual feedback data, and ultimately improve the long-term practicality and stability in assisted image classification and archiving tasks. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a model quantization method based on pathological image classification according to the present invention; Figure 2 This is a schematic diagram of an anomaly detection model in a model quantization method based on pathological image classification according to the present invention; Figure 3This is a schematic diagram of a model quantization system based on pathological image classification according to the present invention. Detailed Implementation

[0017] To make the technical solution of the present invention clearer and its technical advantages more apparent, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of the present invention.

[0018] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the steps of a model quantization method based on pathological image classification according to the present invention. Figure 2 This is a schematic diagram of an anomaly detection model in a model quantization method based on pathological image classification according to the present invention. The following is a detailed introduction to this model quantization method based on pathological image classification.

[0020] Step S1: Obtain the original pathological image and input it into the anomaly detection model to generate an anomaly heatmap.

[0021] In this embodiment, step S1 includes: Step S1-1: Construct and train the anomaly detection model.

[0022] Specifically, the anomaly detection model includes a compressed anomaly detection submodule and a predicted anomaly detection submodule. A convolutional neural network is used as the shared feature extraction backbone network of the anomaly detection model. The shared feature extraction backbone network consists of three convolutional layers, each using a small 3x3 convolutional kernel.

[0023] Specifically, the first convolutional layer contains 32 filters with a stride of 2, used for initial downsampling and basic feature extraction; the second convolutional layer contains 64 filters with a stride of 1, used to capture medium-granularity features; the third convolutional layer contains 128 filters with a stride of 2, used to output high-level semantic features, which are finally compressed into a 128-dimensional feature vector by a global average pooling layer.

[0024] It should be noted that a lightweight channel attention mechanism is introduced between the second and third convolutional layers. This mechanism obtains the channel dimension statistics through global average pooling, and then generates the weights of each channel through two fully connected layers, enabling the network to autonomously emphasize information-rich feature channels. Each convolutional layer is followed by batch normalization and ReLU activation function to accelerate training convergence and enhance nonlinear expressive power. The weights of all layers are quantized to 4 bits to reduce computation and storage overhead.

[0025] Furthermore, a convolutional neural network architecture is used to construct a compressed anomaly detection submodule. A self-supervised learning strategy is adopted to train the compressed anomaly detection submodule to predict the conditional probability distribution of pixel blocks in the original pathological image given the feature vector of the original image. The loss function is negative log-likelihood loss, and the training aims to minimize the negative log probability of the true pixel value under the predicted conditional probability distribution.

[0026] In one possible embodiment, the compressed anomaly detection submodule comprises two fully connected layers. The first layer uses the LeakyReLU activation function to map the 128-dimensional feature vector to 64 dimensions. The second layer is used to further map the 64-dimensional vector to 32 dimensions. Finally, a linear layer outputs 256×256×256 parameters, which correspond to a 256×256 image patch, with each pixel having a probability distribution of 256 possible values.

[0027] The training process of the compressed anomaly detection submodule adopts a self-supervised learning approach. Specifically, given a small patch from a normal pathological image as input, the compressed anomaly detection submodule is required to predict the conditional probability distribution of any value between 0 and 255 for the target pixel based on the known context pixels to the left and above the target pixel. For example, in a uniformly stained cytoplasmic region, it should be able to predict pixel values ​​close to the context with high confidence, or at the edge of the cell nucleus, it should predict a color mutation pattern. Then, the negative log probability of the true pixel value under this distribution is calculated, and the training objective is to minimize the negative log likelihood loss of the true pixel value under the predicted distribution.

[0028] It should be noted that, since the pixel distribution of pathological images is extremely complex and not a simple Gaussian distribution, a conditional normalization flow is introduced, and an invertible neural network is deployed inside the module. During the training process, this neural network learns how to transform a simple basic distribution, such as a uniform distribution or a standard Gaussian distribution, into a complex real pixel value distribution corresponding to the current context through a series of reversible and learnable transformations. This enables the compressed anomaly detection submodule to capture the statistical characteristics of normal tissue, from uniform cytoplasm to complex chromatin texture.

[0029] Furthermore, a U-Net encoder-decoder architecture is used to construct a prediction anomaly detection submodule, which employs a self-supervised learning strategy and is pre-trained using a joint loss function based on L1 loss and structural similarity loss.

[0030] In one possible embodiment, the core task of the anomaly detection prediction submodule is to predict the current pixel value based on contextual information and detect anomalies by comparing the difference between the predicted value and the true value. This submodule is constructed using a U-Net-style encoder-decoder architecture. Specifically, the network structure of the anomaly detection prediction submodule includes an encoder and a decoder. The encoder consists of four downsampled convolutional layers, each using a 3×3 convolution with a stride of 2, and the number of channels in each convolutional layer is set to 64, 128, 256, and 512, respectively. The decoder contains four upsampled transposed convolutional layers symmetrically relative to the encoder, with the number of channels set to 256, 128, 64, and 3, respectively. The encoder and decoder are fused through skip connections to preserve spatial detail information, and all layers use 2-bit weights and 4-bit activation values.

[0031] The training strategy also employs self-supervised learning, with the training objective being to minimize the difference between the predicted and real images. The loss function combines L1 loss and structural similarity loss, i.e. in For total losses, The preset balance parameter is usually set to 0.5. For L1 loss, The loss function is based on structural similarity; and during training, a training mechanism based on conditional generative adversarial networks is introduced to improve the prediction accuracy of sub-modules.

[0032] It should be noted that the compressed anomaly detection submodule and the predicted anomaly detection submodule are not trained independently, but are trained collaboratively using a multi-task learning framework. Specifically, the parameter updates of the shared feature extraction backbone network are simultaneously affected by the gradients of both submodules, and gradient normalization and conflict resolution algorithms are used to balance the parameter update directions from different tasks. During training, an alternating optimization strategy is adopted, that is, in each round of training, the backbone network is fixed first, and the parameters of the two submodules are updated, and then the submodules are fixed and the backbone network is updated, to ensure that the two tasks develop in a balanced manner and avoid a single path dominating the training process. Furthermore, during training, knowledge distillation, label smoothing, and random weight averaging methods are used to further improve the model's generalization ability and robustness.

[0033] Step S1-2: Analyze the original pathological images based on the anomaly detection model to generate an anomaly heatmap.

[0034] Specifically, the anomaly detection model uses a multi-scale sliding window strategy to extract features from the original pathological image, obtain the original image feature vector, and the compressed anomaly detection submodule maps the original image feature vector to the conditional probability distribution of each pixel block in the original pathological image.

[0035] The original pathological image is encoded and compressed based on the conditional probability distribution. The encoded length of the compressed original pathological image is used as the first compression length. The original pathological image is encoded and compressed using a lossless compression algorithm. The encoded length of the compressed original pathological image is used as the second compression length. The image compression ratio is obtained based on the first compression length and the second compression length. Compression anomaly detection is performed based on the image compression ratio.

[0036] Furthermore, the anomaly detection prediction submodule performs image prediction based on the feature vector of the original image to generate a predicted pathological image. It then performs a difference analysis between the predicted pathological image and the original pathological image to obtain the absolute error between the two images and obtains the structural similarity index between them. Finally, it constructs a prediction difference map based on the structural similarity index and the absolute error.

[0037] Anomaly detection is performed based on the predicted difference map, and an anomaly heatmap is constructed based on the predicted difference map and the image compression rate.

[0038] In one possible embodiment, a multi-scale sliding window processing strategy, from coarse to fine, is used to extract features from the original pathological image and construct a multi-level image pyramid. Specifically, firstly, the original pathological image is downsampled at a coarse level, such as by 1 / 16 downsampling. That is, during this sampling stage, a relatively large sliding window is used to quickly scan the original pathological image with a large step size. For example, the original pathological image with an original resolution of 80,000×70,000 pixels is quickly sampled with a window size of 512×512 pixels and a step size of 256 pixels to generate a low-resolution global overview image.

[0039] It should be noted that the sliding window step size in this stage is not fixed. Instead, it is used to analyze the complexity of the image content within each window in real time by calculating texture features such as the entropy value of the local gray-level co-occurrence matrix. For regions with simple and homogeneous textures, such as large areas of fat or blank backgrounds, the step size will be automatically increased to quickly skip over the region. For regions with complex textures and dense structures, such as glandular clusters, a denser scan is used, i.e., the step size is reduced, to scan the region more precisely. Based on this preliminary scan, a low-resolution anomaly probability sketch is generated, and candidate abnormal regions in the original pathological image that need to be analyzed in detail at high resolution are identified accordingly.

[0040] Subsequently, refined analysis is performed only in the regions marked as candidate anomalies. This involves dense scanning with a smaller window size and step size compared to the initial coarse-grained downsampling to capture subtle anomalies. For example, for a liver tissue slice, approximately 400 windows are processed at low resolution to generate a low-resolution global overview. Based on this global overview, approximately 70% of the stroma and background regions are quickly excluded, and three candidate anomaly regions exhibiting local texture disorder and uneven staining are marked. For these three anomaly regions, approximately 1200 windows are processed at the original resolution to extract the anomaly features of these regions. Finally, a raw image feature vector is generated by combining the corresponding features within the multi-level image pyramid.

[0041] The compressed anomaly detection submodule receives the original image feature vector extracted from the shared feature extraction backbone network. It uses a conditional normalization flow mechanism to map the original image feature vector into a conditional probability distribution. Specifically, the conditional normalization flow consists of a series of reversible affine coupling layers. Each layer uses the features of the previous step or the original context features as conditions to perform a progressive transformation on a simple prior distribution such as a multivariate Gaussian distribution. Each coupling layer splits the input vector into two parts, performs scaling and translation transformations on only one part based on neural network parameterization, and leaves the other part unchanged and recombines it with the transformed part. The recombination requires that the overall transformation is reversible and that the Jacobian determinant is easy to compute. After several of the above reversible transformations, a set of parameters is output, such as the set of parameters of the mean, variance, and mixture weights of the Gaussian mixture model. This set of parameters defines the conditional probability distribution that each pixel value in the image patch should follow under the current feature context.

[0042] Based on this conditional probability distribution, a custom 8-bit specific-point arithmetic encoder is used to perform lossless compression on the original pathological image. This arithmetic encoder can encode the pixel sequence of the entire image block into a unique fractional interval. The length of the resulting compressed bitstream is used as the first compression length, denoted as . Specifically, for the pixel to be encoded, the encoder looks up the cumulative probability interval corresponding to the true value of the pixel in its conditional probability distribution. The encoder maintains a dynamic current range. Its initial value is set to [0, 1). When encoding each pixel, it is narrowed proportionally according to the probability range of that pixel, that is... , As the number of encoded pixels increases, the interval As the range narrows, the number of bits required for its binary representation increases. After processing all pixels, the encoder outputs the shortest binary bitstream that uniquely represents any number within the final interval; its length is the compressed size. Simultaneously, a general compression algorithm such as LZW is used to compress the same image patch, resulting in a second compressed length, denoted as . And define the image compression rate as When R is significantly greater than 1, it indicates that the customized 8-bit specific point arithmetic encoder shows a significant advantage over the general compression algorithm, meaning that the statistical characteristics of the image patch highly conform to the patterns learned by the model from a large amount of normal data, and it is normal. When R is close to 1, it means that the customized 8-bit specific point arithmetic encoder failed to find effective statistical patterns, and its compression efficiency is no different from that of the general algorithm, suggesting that there may be anomalies in this area. When R is significantly less than 1, it means that using the prior knowledge of the normal model for encoding is less efficient than directly using the general compression algorithm, which means that the current model has encountered extreme samples, training data bias, or model defects. Since R<1 is a very strong statistical anomaly signal, the prediction error of the difference map in the anomaly detection submodule will also show a very high value, because the visual pattern cannot be accurately predicted. At this time, the two indicators will resonate and jointly mark the area as highly suspicious.

[0043] The anomaly detection prediction submodule constructs a prediction difference map based on the feature vectors of the original image. Specifically, firstly, based on the shared feature vectors of the original image, a lightweight decoding network attempts to reconstruct the input image, resulting in a reconstructed image. Subsequently, the pixel-level absolute error between the original image and the reconstructed image is calculated at the pixel level. Next, the structural similarity index is calculated within a local neighborhood, for example, within a window of size 11×11 pixels. ;according to in The prediction error value reflects the deviation between the visual content of the image patch and the best prediction made by the submodule based on the context; β is a normalized preset adjustable parameter whose value is obtained from historical data analysis, and a prediction difference map is constructed based on the obtained prediction error value.

[0044] It should be noted that image compression ratio and prediction error are inherently complementary. Image compression ratio is highly sensitive to statistical anomalies in pixel value distribution, such as abnormal distributions of color intensity, while prediction error is more sensitive to anomalies in spatial structure patterns, such as abrupt changes in edge direction and texture patterns. For example, a region with only local color deepening but no change in texture may not show significant prediction error, but its pixel value distribution will deviate from the normal statistical model, leading to a significant decrease in image compression ratio. Conversely, a region with complex texture distortion but normal average color will have an abnormally high prediction error. An adaptive weighted fusion of image compression ratio and prediction error is used to generate a comprehensive anomaly score. Based on this comprehensive anomaly score, an anomaly heatmap is constructed. Specifically, firstly, a rapid meta-feature analysis is performed on the current image patch. For example, the main coloring regions are identified through its color histogram, and its texture complexity is analyzed through local binary pattern analysis to obtain the corresponding meta-features. Then, based on these meta-features, weights are dynamically assigned to the image patch. , and These are used respectively for weighted image compression ratio and prediction error, i.e. ,in This is represented as a comprehensive anomaly score. For example, for a densely packed region of glandular structures with regular, cavity-like textures, structural disorder is a more typical anomaly marker for this type of region, thus assigning a higher weight to the prediction error. In this case, it can be expressed as... The ratio is set to 3:7; finally, a value is calculated for each analyzed image patch in the entire image. The values ​​are then mapped to each pixel using bilinear interpolation, ultimately generating a continuous, pixel-level anomaly heatmap.

[0045] It should be noted that, in order to extract discrete candidate regions from the anomaly heatmap for further in-depth analysis, a dynamic threshold segmentation algorithm based on local statistical characteristics is used for anomaly region localization. Specifically, for each pixel position in the anomaly heatmap, a local window is defined centered on each pixel, such as 500×500 pixels, and the mean of all comprehensive anomaly scores within this window is calculated. Sum and standard deviation And set the dynamic threshold of the window Set as ,in This is an adjustable sensitivity parameter, derived through analysis of historical data; it only applies when the pixel's overall anomaly score... Larger than its corresponding local window A pixel is only marked as an anomaly when it reaches a certain value. After anomaly marking, a binarized anomaly mask is obtained. Morphological closing operations are used to connect adjacent small anomaly points and fill small holes to form connected and complete connected regions. Each connected region is marked as an independent object, and its features such as the bounding rectangle boundary, pixel area, and average anomaly intensity are recorded. All connected regions are sorted according to their average comprehensive anomaly score, and the region with the highest priority will be sent to the next stage, that is, the top N connected regions with the highest average comprehensive anomaly scores are selected as candidate anomaly regions.

[0046] Step S2: Based on the abnormal heatmap, feature extraction is performed on the original pathological image to obtain an abnormal region feature dataset.

[0047] Specifically, a lightweight encoder-decoder network based on the information bottleneck principle is used to perform deep feature extraction on the high-confidence abnormal regions marked in the abnormal heatmap, obtain the reconstruction error and feature deviation, and output the reconstruction error, feature deviation, and the features contained in the corresponding regions in the abnormal heatmap as the abnormal region feature dataset.

[0048] It should be noted that the lightweight encoder consists of four convolutional layers, each using a 3x3 convolutional kernel, group normalization, and the LeakyReLU activation function, with its weights and activation values ​​quantized to 4 bits and 8 bits, respectively. The purpose of this encoder is to compress a high-dimensional pixel space image into a low-dimensional bottleneck feature vector, that is, to forcibly discard redundant information in the original pathological image that is not related to distinguishing between normal and abnormal, such as changes in illumination and slight differences in staining, and retain only the most discriminative semantic features. For example, for a glandular tissue region marked as "suspicious structural disorder", the encoder may ignore the color intensity of the secretions in the glandular lumen and focus on capturing key morphological information such as the degree of distortion of the glandular contour and whether the polarity of the cell nucleus is lost, and encode this information into a 32-dimensional digital vector.

[0049] The decoder attempts to reconstruct the original image patch from this bottleneck vector. The decoder has a symmetrical structure to the encoder, upsampling through transposed convolutional layers. By jointly training the decoder and encoder, it is ensured that the bottleneck vector carries sufficient information to reconstruct the original image. Specifically, a pre-defined ternary loss function is used for training. This loss function includes an L1 reconstruction loss to ensure the network can achieve high-quality reconstruction of normal regions; and a feature clustering loss based on contrastive learning. During training, all bottleneck features from normal samples are attracted to a learnable "prototype network" and clustered in a compact cluster in the feature space; while any abnormal samples... Features are pushed away from the center of the normal cluster by a marginal loss; KL divergence regularization of bottleneck features is used to constrain their feature distribution to be close to a simple prior distribution. By combining with a preset training dataset for training, the lightweight encoder-decoder network is trained to be able to easily reconstruct normal regions, and its features are close to the center of the normal cluster; for abnormal regions, reconstruction is difficult, and its features are far away from the center of the normal cluster. The preset training dataset covers a variety of tissue types such as breast, prostate, and lung, and whole-section pathological images such as H&E staining schemes. It comes from public datasets and historical data samples, such as the public dataset of The Cancer Genome Atlas and anonymized data samples provided by partners.

[0050] Understandably, the reconstruction error is represented as the normalized average absolute difference between the input image block and the decoder output block, which is used to reflect the intensity of the image anomaly at the pixel level; the feature deviation is represented as the Mahalanobis distance between the bottleneck feature corresponding to the high-confidence anomaly region and the pre-stored normal prototype center, which is used to characterize the degree of deviation of the current region features from the normal pattern in the feature space.

[0051] In one possible embodiment, firstly, each marked candidate anomalous region in the anomalous heatmap is fed into a lightweight encoder-decoder network. Based on the trained lightweight encoder-decoder network, the reconstruction error and feature offset of the marked high-confidence anomalous regions in the anomalous heatmap are calculated. Specifically, the reconstruction error is represented as the normalized mean absolute error between the input image patch and the decoder output image patch. This error value is used to reflect the incomprehensibility of the region at the pixel level. For example, for a region that is only slightly stained but structurally intact, the encoder may still be able to capture its structural information, and the decoder can also roughly reconstruct it, so the reconstruction error is low. However, for a region where the cell nucleus size, shape, and chromatin distribution are significantly anomalous, the decoder will have difficulty reconstructing this complex and disordered pattern from its limited bottleneck vector, resulting in a higher reconstruction error. The feature offset is represented by calculating the Mahalanobis distance between the bottleneck feature vector of the region and the "normal feature cluster center" learned during the training phase. This offset directly quantifies the deviation distance of the region from the normal prototype in the abstract feature space. It should be noted that this step also trains a lightweight gradient boosting decision tree. This decision tree is used to determine whether a region needs to initiate the game reasoning process in step S3. Specifically, during the training of this gradient boosting decision tree, the training sample set required comes from a large number of candidate anomaly regions accumulated in historical data after processing in steps S1 and S2. The features of each sample are the image compression rate, prediction error, reconstruction error, and feature deviation corresponding to that region. These four indicators are fused into a four-dimensional anomaly description vector, and the training sample set is then processed... Labeling involves marking samples requiring high-priority verification as positive samples and samples that can be safely filtered as negative samples. By running this rule on a large amount of historical data, a four-dimensional vector training set with pseudo-labels is constructed. This training set is the training sample set. For example, regions with a comprehensive anomaly score higher than a high global threshold and whose feature deviation is significantly greater than the corresponding preset threshold are considered as strong anomaly candidates and given high-priority positive labels; regions that are marked due to minor reasons such as slight unevenness in staining, have extremely low feature deviation, and good reconstruction quality are marked as filterable negative labels.

[0052] Based on the training sample set, a lightweight gradient boosting decision tree is initialized. It takes the four-dimensional anomaly description vectors contained in the training sample set as input and the corresponding automated pseudo-labels as the learning objective. The training objective is set to minimize the cross-entropy loss between the predicted label and the pseudo-label. During training, the maximum depth and number of leaf nodes are limited to constrain the model complexity of the decision tree. Finally, the trained gradient boosting decision tree can perform a series of simple threshold comparisons based on the classification boundary learned during training and defined by historical data consensus. For example, if the feature deviation > threshold A1 and the reconstruction error > threshold A2, it is classified as positive, requiring further analysis. This outputs a binary verification priority flag, which determines whether further operations are needed.

[0053] The four-dimensional anomaly description vector ({image compression rate, prediction error, reconstruction error, feature deviation}), verification priority flag, spatial coordinates, and original image feature vector corresponding to the region requiring subsequent operations are output as the anomaly region feature dataset.

[0054] Step S3: Use a multi-agent game mechanism to verify and analyze the abnormal region feature dataset to obtain an anomaly analysis dataset.

[0055] Specifically, the abnormal region feature dataset is injected into a multi-agent resource pool for resource matching. A candidate agent set is generated for each abnormal region, and the region feature vector corresponding to each abnormal region is obtained. Each candidate agent in the candidate agent set makes a game bid based on the region feature vector, generating a game bid dataset. The VCG mechanism is used to perform multi-objective optimization and decision-making on the game bid dataset to generate corresponding anomaly analysis decisions. The abnormal region is analyzed based on the anomaly analysis decisions to obtain an anomaly analysis dataset.

[0056] Understandably, each agent in the multi-agent resource pool is represented as a lightweight spiking neural network optimized for a specific type of pathological abnormality pattern. During training, each agent is trained using a large number of samples of a specific type of abnormality. Specifically, firstly, a large number of image samples of specific types of abnormalities labeled by pathology experts are used for independent supervised training. For example, samples corresponding to a certain cell arrangement disorder pattern, a certain chromatin distribution characteristic, or a certain microstructural distortion are used to train the agent, thereby making the agent more sensitive and accurate to the response pattern of this type of abnormality, with the goal of minimizing the classification loss of that specific type of abnormality. The target is trained, and at this stage, the projection of its final classification layer weights onto the feature space is defined as a domain-specific encoding vector, which is used to represent the feature subspace sensitive to its decision boundary. Next, the parameters of each agent in the multi-agent resource pool are fine-tuned using actual historical data. Then, during the actual deployment phase, the verification results and feedback information of the agents in actual operation are continuously collected. When an agent succeeds consecutively on a specific task, contrastive learning is used to fine-tune the agent's domain-specific encoding vector toward the feature direction corresponding to the successful cases. Conversely, agents that fail consecutively are eliminated using a tournament mechanism, and new agents are trained based on the latest data to replace them.

[0057] Each agent's attributes include at least one domain-specific encoding vector, which represents the anomalous feature space it excels in; a set of historical verification accuracies, which is a set of historical verification accuracies for different processing task types, for example, the historical verification accuracy for processing cell arrangement topological anomalies is 85%; basic resource cost, which is an estimate of the number of computational operations required to perform a standard verification; and the current state, such as idle, busy, or low-power sleep.

[0058] In one possible embodiment, when a high-priority candidate anomaly region enters the verification queue of the multi-agent resource pool, a fast screening and auction process is initiated. This process is used to quickly evaluate the potential matching degree between each candidate anomaly region and each agent in the resource pool, and to select the most promising candidate combination. Specifically, firstly, the anomaly region feature dataset corresponding to each candidate anomaly region is broadcast to the multi-agent resource pool. Each agent calculates the cosine similarity between the feature offset dimension in the four-dimensional anomaly description vector of each candidate anomaly region and its professional domain encoding vector, and uses the cosine similarity as the feature matching degree score. Next, since each verifier agent caches the spatial locations and task summaries of the top N regions recently processed in memory, the spatial distance between the spatial locations of the top N regions and the spatial locations in the current broadcast information is calculated. If the spatial distance is within a preset spatial proximity range, it indicates that the two regions are spatially close, that is, the agent may have contextual information related to the current region to be verified.

[0059] Next, the context fingerprints of two spatially adjacent regions are extracted, and the Hamming distance between their corresponding context fingerprints is used as a context similarity index. If the context similarity is greater than a preset threshold, it is determined that the agent has similar context information to the current broadcast information, and the normalized value of the context similarity index is used as the context similarity score. Then, the feature matching score and the context similarity score are weighted and fused together with a preset weight ratio to generate a comprehensive matching score. The preset weight ratio is obtained from historical data analysis. Finally, the agents with the highest comprehensive matching scores are selected as a candidate agent set.

[0060] It should be noted that the context fingerprint comprises two parts: texture features and color distribution features. Texture features are defined as follows: for any anomaly region to be analyzed, a square neighborhood larger than the anomaly region itself is extracted, centered on its coordinates. Within this neighborhood, a Local Binary Pattern (LBP) operator is used to generate an LBP response map. A texture feature vector is obtained based on this LBP response map. For example, using an LBP operator with a radius of 2 pixels and 8 sampling points, an LBP response map is generated. A 256-dimensional histogram of this response map is calculated, and non-uniform quantization and smoothing are applied to the histogram, ultimately compressing it to obtain a 64-dimensional texture feature vector. Color distribution features are represented by using color... Multiple color channels are separated from the original pathological image. The first, second, and third moments of each channel are calculated. A color distribution feature vector is generated based on the first, second, and third moments. For example, the square neighborhood of the abnormal region to be analyzed is uniformly divided into a 2x2 grid. The H (hue) and E (eosin) color channels are separated from the H&E staining image corresponding to the abnormal region. Three values ​​of the first, second, and third moments are generated for each channel, for a total of six values ​​for the two channels. These six color moments are calculated in each grid, resulting in a 24-dimensional color distribution feature vector. Finally, the texture feature vector and the color distribution feature vector are concatenated to generate the corresponding context texture feature vector.

[0061] After the rapid screening auction process concludes, an auction phase is initiated. Specifically, each agent within the candidate agent set generates a task uncertainty score based on the regional average comprehensive anomaly score of the candidate anomaly region, i.e., task uncertainty score = 1 - regional average comprehensive anomaly score. Simultaneously, the historical verification accuracy rate of the task type corresponding to the current candidate anomaly region is acquired. The task uncertainty score and historical verification accuracy rate are then fused using a weighting ratio derived from historical data analysis to generate a verification resource control factor. Each agent adjusts its own basic resource cost based on the verification resource control factor to generate a resource bid, i.e., resource bid = (1 + verification resource control factor) * basic resource cost. Under the premise that the total cost does not exceed the system's current available computing budget, the auctioneer selects the agent with the highest resource bid for each candidate anomaly region to generate a verification agent set.

[0062] The algorithm verifies the prediction error of each agent within the agent set that acquires candidate anomaly regions within the anomaly heatmap. Based on the prediction error within this region, it performs analysis and reasoning to generate an anomaly analysis dataset. Specifically, it acquires the prediction error map of the corresponding region within the anomaly heatmap. This error map corresponds to the original image patch in spatial location and contains multiple channels at each point. Each channel represents the degree of error predicted by a specific visual feature model, such as edges in a specific direction or textures at a specific frequency. The error map is then encoded into an error pulse sequence. That is, the larger the error value, the higher the frequency of pulse firing by the corresponding neuron on the time axis, thus ensuring that the spatial distribution information of the error is also carried at the precise time point of pulse firing. The encoded error pulse sequence is then injected into the spiking neural network contained within the agent.

[0063] The spiking neural network consists of a primary feature detection layer, a pattern integration layer, and a high-level decision layer, each containing corresponding spiking neurons. The primary feature detection layer, as the first layer of neurons, receives the input error pulse sequence. When it receives a pulse input that conforms to a preset spatiotemporal pattern and reaches a preset threshold, it generates a corresponding output pulse. For example, a neuron is specifically designed to detect error pulses with progressively increasing intensity that appear at three consecutive spatial locations in the horizontal direction on the nuclear membrane edge channel. Only when the above conditions are met simultaneously will an output pulse be generated. This output pulse means that a continuous nuclear membrane edge error with a rightward increasing intensity characteristic has been detected near a certain location. This corresponds to an image feature of a nuclear membrane contour line that gradually darkens or becomes clearer in a pathological image. This feature is one of the local features corresponding to irregularities or folds in the nuclear membrane.

[0064] The pattern integration layer is used to receive pulses from multiple primary feature detection neurons. For example, a neuron may specifically detect the pattern of "a closed loop formed by multiple short edges connected end to end in space". This pattern corresponds to a complete but irregular cell nucleus outline. The integrating neuron will only generate the corresponding pulse when the pulses from primary neurons representing different edge segments arrive sequentially within a specific time window.

[0065] The higher-level decision layer is used to receive pulse inputs from neurons that integrate various patterns. For example, a decision neuron may integrate multiple integration patterns such as "the existence of an irregular closed loop", "the region within the loop has diffusion error pulses in the chromatin texture channel", and "the size of the loop exceeds the normal range". These input pulses are weighted and summed through synapses that follow the STDP rule. When the weighted sum exceeds a preset threshold, the decision neuron generates a pulse cluster. Based on the spatiotemporal characteristics of the pulse cluster and the corresponding mapping function in the preset decoding table, the final verification conclusion is generated.

[0066] It should be noted that the spatiotemporal characteristics of the pulse cluster include at least the average pulse frequency, the pulse synchronicity index, and the cluster duration. The average pulse frequency represents the number of pulses fired by a single spiking neuron per unit simulated time, reflecting the intensity of the neuron's response to the input stimulus. The pulse synchronicity index measures the degree of coordination in the firing of pulses by two or more spiking neurons in the time dimension. The value of this index is the reciprocal of the standard deviation of the time difference between pulses. A higher index value indicates a stronger temporal consistency of activity among neurons. The cluster duration represents the simulated time length from the first firing of a neuron or a specific group of neurons to its last firing, characterizing the time span of a specific neural activity pattern.

[0067] For example, an agent is used to verify the features of the nuclear contour channel. The spatiotemporal features of its output pulse clusters are {average pulse frequency: 2250Hz, pulse synchronization index: 0.85, cluster duration: 22ms}. Assuming that one of the mapping functions matched by the agent is contour irregularity score = 0.3*(average pulse frequency / 2500) + 0.5*(1-pulse synchronization index) + 0.2*(cluster duration / 30), the corresponding contour irregularity score is 0.754. Then, the partial verification conclusion generated by the agent is {agent ID, target pattern: nuclear membrane irregularity, quantization parameters: [contour irregularity score: 0.754, ...], triggering neuron clusters: ...}.

[0068] Finally, the abnormal heatmap, abnormal region feature dataset, and verification conclusions are output as an anomaly analysis dataset.

[0069] It should be noted that after each agent completes pre-training, it uses a calibration dataset to construct its initial decoding table. Specifically, firstly, the trained spiking neural network processes a set of samples whose abnormality type and degree are known, recording the output pulse sequence patterns generated by its decision-layer neurons under various inputs, such as pulse frequency, timing, and combination of firing neurons. Next, cluster analysis is performed on the pulse sequence patterns to generate multiple pulse pattern clusters, and recurring, stable pulse pattern clusters are mapped to specific pathological morphological descriptions to construct the initial decoding table. Subsequently, this initial decoding table will be updated online based on actual feedback data after system deployment.

[0070] Step S4: Obtain the actual feedback data of the anomaly analysis dataset, and perform feedback optimization based on the actual feedback data.

[0071] Specifically, interactive image explanation information is generated based on the anomaly analysis dataset, and the interactive image explanation information is transmitted to the interactive terminal. The actual feedback data submitted by the interactive terminal is collected synchronously. The actual feedback data includes at least anomaly analysis correctness indicators and anomaly analysis correction indicators. The anomaly analysis correctness indicators are represented by a binary judgment of "confirmation" or "rejection" by the administrator for each area of ​​interest marked by the system. The anomaly analysis correction indicators are represented by a combination of parameters by the administrator to adjust the numerical values ​​of the quantitative description provided by the system or to fine-tune the boundaries of the areas.

[0072] For example, the administrator confirmed the marking of candidate abnormal regions A and B, but rejected the marking of region C because region C was judged to be a production artifact. At the same time, the administrator manually corrected the "contour irregularity score" of region A from 0.754 reported by the system to 0.65 because the system mistakenly included some glandular fold shadows in the contour. These operations were automatically recorded as a feedback tuple with a timestamp: {region ID, operator ID, action: correction, indicator: contour irregularity score, value before correction: 0.754, value after correction: 0.65}.

[0073] Based on actual feedback data, incremental learning and parameter optimization are performed on each sub-module within the anomaly detection model. For example, if an administrator repeatedly rejects regions marked as abnormal due to "tissue tearing artifacts," image patches of these rejected regions are collected and used as normal samples to compress the anomaly detection sub-module during the incremental learning phase. Through quantized perception training, the parameters of the sub-module are fine-tuned so that it learns that "pixel value mutations at tear edges" are also a predictable contextual pattern, thereby reducing overreactions to such artifacts in the future. If an administrator identifies a region with disordered glandular structure and corrects its score from 0.7 to 0.9, the image patch of this region will participate in the training of the anomaly detection sub-module with higher weights during incremental learning, forcing it to attempt to further refine the reconstruction of this complex disordered structure.

[0074] Figure 3 This is a schematic diagram of a model quantization system based on pathological image classification according to the present invention.

[0075] Specifically, a model quantization system based on pathological image classification includes: The data acquisition module is used to acquire the original pathological images and input them into the anomaly detection model to generate an anomaly heatmap.

[0076] An anomaly detection model is used to extract features from the original pathological image and to detect anomalies in the original pathological image by combining the compressed anomaly detection submodule and the predicted anomaly detection submodule contained in the model.

[0077] The feature extraction module is used to extract features from the original pathological image by combining the abnormal heatmap, and obtain an abnormal region feature dataset.

[0078] The verification and analysis module is used to perform verification and analysis on the feature dataset of abnormal regions based on the multi-agent game mechanism to obtain the abnormal analysis dataset.

[0079] The feedback optimization module is used to obtain the actual feedback data of the anomaly analysis dataset and perform feedback optimization based on the actual feedback data.

[0080] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0081] The above embodiments can be implemented, in whole or in part, through software, hardware (such as circuits), firmware, or any other combination thereof.

[0082] When implemented using software, the above embodiments can be implemented in whole or in part as a computer program product, which includes one or more computer instructions or computer programs; when the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part.

[0083] It is understood that the computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device; the computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via transmission methods such as infrared, wireless, or microwave; the computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0084] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0085] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0086] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A model quantization method based on pathological image classification, characterized in that, It includes the following steps: The original pathological image is acquired and input into the anomaly detection model to generate an anomaly heatmap. The anomaly detection model includes a compressed anomaly detection submodule and a predicted anomaly detection submodule. Based on the aforementioned abnormal heatmap, feature extraction is performed on the original pathological image to obtain an abnormal region feature dataset. A multi-agent game mechanism is used to verify and analyze the feature dataset of abnormal regions, and an anomaly analysis dataset is obtained. Obtain the actual feedback data from the anomaly analysis dataset, and perform feedback optimization based on the actual feedback data.

2. The model quantization method based on pathological image classification according to claim 1, characterized in that, Anomaly detection models include: A convolutional neural network is used as the shared feature extraction backbone network of the anomaly detection model. Based on the shared feature extraction backbone network, features are extracted from the original pathological image to obtain the original image feature vector. A convolutional neural network architecture is used to construct a compressed anomaly detection submodule; A self-supervised learning strategy is adopted to train the compressed anomaly detection submodule to predict the conditional probability distribution of pixel blocks in the original pathological image given the feature vector of the original image. The loss function uses negative log-likelihood loss, and the training aims to minimize the negative log probability of the true pixel value under the predicted conditional probability distribution. The predictive anomaly detection submodule is constructed using a U-Net encoder-decoder architecture; A self-supervised learning strategy is adopted, and pre-training is performed using a joint loss function based on L1 loss and structural similarity loss.

3. The model quantization method based on pathological image classification according to claim 1, characterized in that, Obtain the original pathological images and input them into the anomaly detection model to generate an anomaly heatmap, including: The anomaly detection model uses a multi-scale sliding window strategy to extract features from the original pathological images and obtain the feature vector of the original images. The compressed anomaly detection submodule maps the original image feature vector to the conditional probability distribution of each pixel block in the original pathological image; The original pathological image is encoded and compressed based on the conditional probability distribution, and the encoding length of the compressed original pathological image is used as the first compression length. The original pathological image is encoded and compressed using a lossless compression algorithm, and the encoded length of the compressed original pathological image is used as the second compression length. The image compression ratio is obtained based on the first compression length and the second compression length, and compression anomaly detection is performed based on the image compression ratio.

4. The model quantization method based on pathological image classification according to claim 3, characterized in that, The method further includes: The prediction anomaly detection submodule performs image prediction based on the feature vector of the original image to generate a predicted pathological image; A difference analysis was performed between the predicted pathological image and the original pathological image to obtain the absolute error between the predicted pathological image and the original pathological image. Obtain the structural similarity index between the predicted pathological image and the original pathological image, and construct a prediction difference map based on the structural similarity index and the absolute error; Anomaly detection is performed based on the predicted difference map, and an anomaly heatmap is constructed based on the predicted difference map and the image compression rate.

5. The model quantization method based on pathological image classification according to claim 1, characterized in that, Based on the aforementioned abnormal heatmap, feature extraction is performed on the original pathological image to obtain an abnormal region feature dataset, including: A lightweight encoder-decoder network based on the information bottleneck principle is used to extract deep features from the high-confidence anomaly regions marked in the anomaly heatmap, and the reconstruction error and feature deviation are obtained. The reconstruction error is expressed as the normalized average absolute difference between the input image block and the decoder output block, which is used to reflect the abnormal intensity of the image at the pixel level. The feature deviation is represented by the Mahalanobis distance between the bottleneck feature corresponding to the high-confidence abnormal region and the pre-stored normal prototype center, which is used to characterize the degree of deviation of the current region feature from the normal pattern in the feature space. The reconstruction error, feature deviation, and features contained in the corresponding regions of the anomaly heatmap are output as the anomaly region feature dataset.

6. The model quantization method based on pathological image classification according to claim 1, characterized in that, A multi-agent game mechanism is used to validate and analyze the anomaly region feature dataset, obtaining an anomaly analysis dataset, including: The abnormal region feature dataset is injected into the multi-agent resource pool for resource matching, and a candidate agent set is generated for each abnormal region. Obtain the regional feature vectors corresponding to each abnormal region. Each candidate agent in the candidate agent set makes a game bid based on the regional feature vectors to generate a game bid dataset. The VCG mechanism is used to perform multi-objective optimization and decision-making on the game bidding dataset, generating corresponding anomaly analysis decisions. Based on the anomaly analysis decision, the abnormal region is analyzed to obtain an anomaly analysis dataset.

7. The model quantization method based on pathological image classification according to claim 1, characterized in that, Obtain actual feedback data from the anomaly analysis dataset, and perform feedback optimization based on the actual feedback data, including: Interactive image explanation information is generated based on the anomaly analysis dataset, and the interactive image explanation information is transmitted to the interactive terminal, while the actual feedback data submitted by the interactive terminal is collected synchronously. The actual feedback data includes at least anomaly analysis correctness indicators and anomaly analysis correction indicators; Incremental learning and parameter optimization are performed on each sub-module within the anomaly detection model based on the actual feedback data.

8. A model quantization system based on pathological image classification, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire the original pathological image and input the original pathological image into the anomaly detection model to generate an anomaly heatmap. An anomaly detection model is used to extract features from the original pathological image and to detect anomalies in the original pathological image by combining the compressed anomaly detection submodule and the predicted anomaly detection submodule contained in the model. The feature extraction module is used to extract features from the original pathological image by combining the abnormal heatmap to obtain an abnormal region feature dataset. The verification and analysis module is used to perform verification and analysis on the abnormal region feature dataset according to the multi-agent game mechanism to obtain the abnormal analysis dataset. The feedback optimization module is used to obtain the actual feedback data of the anomaly analysis dataset and perform feedback optimization based on the actual feedback data.

Images