Brain glioma full-section pathological image intelligent analysis method based on double-path analysis framework

Abstract

A brain glioma full-slice pathological image intelligent analysis method based on a double-path analysis framework specifically comprises the following steps: S1, preprocessing full-slice images, dividing each full-slice image into different foreground tissue areas and background areas by using a channel threshold segmentation algorithm, and performing image segmentation on the foreground tissue areas and the background areas; cutting the foreground tissue region into a plurality of image blocks with fixed sizes, and extracting a feature vector of each image block to form a feature cluster; s2, establishing a dynamic sampling mechanism, performing dynamic multi-instance learning on the feature cluster of each full-slice image by using a two-stage attention network, and finally obtaining slice-level feature representation of the full-slice image; and S3, quantifying tumor microenvironment spatial relationship information to obtain a feature representation of a spatial relationship quantization graph, and splicing the feature representation with the slice-level feature representation in the S2 to obtain a final prediction result. According to the method, the prediction efficiency and robustness are improved, and the medical decision speed and quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to an intelligent analysis method for whole-slice pathological images of brain gliomas based on a dual-path analysis framework. Background Art

[0002] Gliomas are the most common primary intracranial tumors in adults, with different aggressiveness and prognosis. Among them, isocitrate dehydrogenase (IDH) mutations in diffuse gliomas are the main genetic events that define adult diffuse gliomas, while telomerase reverse transcriptase (TERT) mutations are more common in glioblastomas and oligodendrogliomas, indicating that the tumor behavior is more aggressive. However, these molecular classifications are currently mainly obtained through immunohistochemical molecular sequencing, which is not only labor-intensive and time-consuming. Therefore, if the molecular stratification of gliomas can be identified in a timely and accurate manner based on these biomarkers, this is crucial for providing more precise and targeted treatments.

[0003] Whole-slice images, especially hematoxylin and eosin (HE)-stained slices, not only provide information about the tumor microenvironment, but also directly reflect the transformation of cell nuclei and tissues, and are the gold standard for histological diagnosis of brain gliomas. With the help of artificial intelligence technology, this microscopic information can be quickly mined and used for accurate diagnosis of tumors. However, there are still some challenges in the whole-slice image analysis and modeling process, especially how to effectively extract key tumor microenvironment features from ultra-high resolution images.

[0004] The size of whole-slice images is usually in the tens of thousands of pixels, and due to memory limitations, traditional deep learning methods cannot be directly trained end-to-end. To this end, a patch-based multi-instance learning method is proposed, which allows the processing of whole-slice images through weakly supervised training of the network, even with only slice-level labels. Among them, attention-based multi-instance learning can better highlight the role of key instances by further assigning weights to each instance in the bag. However, directly applying attention-based fusion to thousands of image patches will reduce efficiency, and a large amount of redundant information will hinder the extraction of key local information. In addition, many other instance learning methods ignore the spatial relationship between tumor microenvironment information and image patches in whole-slice images. This omission may lead to information loss, and this lost information often presents spatial local correlation in cells, tissues or lesions in images, which is crucial in pathologists' diagnosis.

[0005] Therefore, with higher requirements for tumor microenvironment characteristics, the present invention aims to provide an intelligent analysis method for whole-slice pathological images of gliomas based on a dual-pathway analysis framework, which can effectively classify and identify whole-slice pathological images while providing high performance, and the entire processing process is automatically completed by a computer without the need for human intervention.

[0006] It should be understood that the above statements merely provide background technology related to the present invention and do not necessarily constitute prior art. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent analysis method for whole-slice pathological images of gliomas based on a dual-path analysis framework, which can effectively classify and identify whole-slice pathological images while providing high performance, and the entire processing process is automatically completed by a computer without the need for human intervention.

[0008] In order to achieve the above-mentioned purpose, the present invention provides an intelligent analysis method for whole-slice pathological images of glioma based on a dual-path analysis framework, which specifically comprises the following steps: S1, preprocessing of whole-slice images, using a channel threshold segmentation algorithm to divide each whole-slice image into different foreground tissue areas and background areas, and cutting the foreground tissue area into a number of image blocks of fixed size, and then extracting the feature vector of each image block to form a feature cluster; S2, establishing a dynamic sampling mechanism, using a two-stage attention network to perform dynamic multi-instance learning on the feature clustering of each whole-slice image, and finally obtaining a slice-level feature representation of the whole-slice image; S3, quantifying the spatial relationship information of the tumor microenvironment, obtaining a feature representation of the spatial relationship quantization map, and splicing it with the slice-level feature representation in S2 to obtain the final prediction result.

[0009] Furthermore, in the S1, dividing each whole slice image into different foreground tissue areas and background areas specifically includes the following steps: K1, reading the whole slice image, converting it to the HSV color space and performing median filtering; K2, binarizing the image at a set saturation channel threshold; K3, performing a morphological closing operation on the binarized image in K2.

[0010] Furthermore, the saturation channel threshold ranges from 10 to 30.

[0011] Furthermore, the dynamic sampling mechanism includes center sampling and dynamic random sampling.

[0012] Furthermore, the two-stage attention network uses two parallel fully connected layers L and U to obtain the underlying spatial features, and uses a softmax function and a mapping layer M to obtain the attention score.

[0013]

[0014] In the formula, a i is the attention score of the i-th image block feature in the full slice image, · is the dot product, tanh and sigm are tanh and sigmoid functions respectively, and They are the image block features of different inputs.

[0015] Furthermore, in the first stage of the attention network, two dynamic sampling mechanisms are used to extract K image block features of two sizes of 768 dimensions respectively, and the two 768-dimensional features are mapped to 512 dimensions through the first stage of the attention network projection layer. The attention scores of the two features mapped to 512 dimensions are calculated respectively, and the initial 768-dimensional image block features are sorted according to the attention scores. The image block features with higher scores extracted by the two sampling mechanisms are selected and spliced ​​to obtain a 768-dimensional splicing-level feature representation, completing the first stage of the attention network.

[0016] Furthermore, in the second-stage attention network, the dimension of the splicing-level feature representation is compressed to 512 dimensions again using the second-stage attention network projection layer, and the attention score of the corresponding i-th image block feature is calculated, thereby obtaining the slice-level feature representation and completing the second-stage attention network calculation, as shown in the following formula:

[0017]

[0018] In the formula, h slide is the slice-level feature representation, s j is the attention score, It is the concatenation-level feature representation.

[0019] Furthermore, the S3 specifically includes the following steps: Y1. Perform center sampling on the M full-slice images in the training set, sample K center image blocks from each slice, obtain M×K center image blocks, and perform K-means clustering on all the center image blocks again to divide them into G groups; Y2. Extract the center image blocks in each group again by the center sampling method to obtain G representative image blocks; Y3. Calculate the similarity distance between all image blocks in each full-slice image and these G representative image blocks, and assign the representative image blocks with the closest similarity to the corresponding group; Y4. Calculate the proportion of image blocks in each group for each full-slice image in the training set to obtain G Ratio values, the ratio values ​​of all training set images are combined into an M×G matrix, and a T test is performed using M×1 labels to obtain G T statistics, and the T statistics of the G image block groups are ranked to obtain a new group value; Y5, by assigning RGB color values ​​to the group values ​​of the image blocks, the mapped image blocks are placed back to the corresponding positions of the full medical image to obtain a spatial relationship quantization map, and the spatial relationship quantization map is encoded using the residual neural network ResNet18 to extract the potential features of the spatial relationship quantization map; Y6, the slice-level feature representation is spliced ​​with the feature representation of the spatial relationship quantization map, and sent to the classifier for slice-level prediction to obtain the final prediction result.

[0020] Furthermore, the similarity distance is calculated using the cosine similarity distance formula:

[0021]

[0022] Where A and B represent the feature vectors of the image blocks.

[0023] In summary, the present invention provides an intelligent analysis method for whole-slice pathological images of gliomas based on a dual-path analysis framework. The dynamic multi-instance learning path based on the two-stage attention mechanism effectively selects and concentrates on analyzing image information in key areas through the two-stage attention mechanism, minimizes the interference of non-critical image blocks, and improves the model's ability to extract key features of the tumor microenvironment; secondly, the spatial relationship information quantification path uses the spatial information quantification method to enhance the model's ability to parse the spatial relationship between images, so that the model can better understand and characterize the complexity of the tumor microenvironment; finally, the entire pathological image analysis process is automated, greatly reducing the need for manual intervention, improving the efficiency and robustness of the prediction, making the present invention more convenient and quick in clinical applications, and helping to improve the speed and quality of medical decision-making. The verification on public data sets reflects the significant advantages of the present invention in improving the accuracy of glioma molecular typing and prognosis prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1It is a flow chart of a specific method of the present invention;

[0025] Figure 2 It is a schematic diagram of the Kaplan-Meier survival curve of the prognosis prediction result of the present invention. DETAILED DESCRIPTION

[0026] The following is combined with Figures 1-2 , the present invention is further explained by describing a preferred specific embodiment in detail.

[0027] It should be noted that the drawings are in a very simplified form and use non-precise proportions. They are only used to conveniently and clearly assist in explaining the embodiments of the present invention, and are not used to limit the conditions for the implementation of the present invention. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention.

[0028] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only the elements explicitly listed, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0029] like Figure 1 As shown, the present invention provides an intelligent analysis method for whole-slice pathological images of glioma based on a dual-path analysis framework, which specifically comprises the following steps:

[0030] S1. Preprocessing of the whole slice image: using the channel threshold segmentation algorithm, each whole slice image is divided into different foreground tissue areas and background areas, and the foreground tissue area is cut into a number of image blocks of fixed size, and then the feature vector of each image block is extracted to form feature clustering;

[0031] S2. Establish a dynamic sampling mechanism to perform dynamic multi-instance learning on the feature clustering of each full-slice image using a two-stage attention network, and finally obtain the slice-level feature representation of the full-slice image;

[0032] S3. Quantify the spatial relationship information of the tumor microenvironment to obtain the feature representation of the spatial relationship quantization map, and concatenate it with the slice-level feature representation in S2 to obtain the final prediction result.

[0033] Specifically, in S1, firstly, the whole slice image is read, converted to HSV color space and median filtered; secondly, the image is binarized at the set saturation channel threshold, and the binarized image is subjected to morphological closing operation, so that each whole slice image can be divided into a foreground tissue area and a background area. The saturation channel threshold ranges from 10 to 30, and in a preferred embodiment of the present invention, the saturation channel threshold is 20.

[0034] Furthermore, the foreground tissue area is cut into non-overlapping image blocks of a fixed size of 256×256 pixels by a sliding window, and then a pre-trained encoder is used to extract the features of the non-overlapping image blocks to obtain the corresponding feature vectors, and then a clustering algorithm is used to divide the feature vectors into K feature clusters, where i=1, 2, …, K.

[0035] Among them, the features in the i-th feature cluster can be expressed as It can be named as the feature cluster of the ith feature, where d represents the number of image block features contained in the ith feature cluster, and 768 represents the feature dimension. Therefore, each full slice image can be represented as F = {C1, C2, …, C K}, which can also be named as feature clustering pool F. Specifically, the original dimension of an image block is 256×256×3. After being encoded by the encoder (i.e., extracting features from the image block), it becomes a 768-dimensional feature vector. The pre-trained encoder in the prior art is used here, which is based on the Transformer network structure, so the output dimension is 768.

[0036] Furthermore, in S2, there are two dynamic sampling mechanisms, including center sampling and dynamic random sampling; the center sampling refers to extracting the image block features at the center of each category cluster in the feature clustering pool F, representing the overall features of the cluster; the dynamic random sampling refers to randomly selecting an image block feature from each feature cluster during each network iteration.

[0037] Among them, the S2 contains two stages. In the first stage of the attention network, for each full slice image, two dynamic sampling mechanisms are used to extract K image block features of size 768, namely: Among them, F cs It is represented by the feature obtained by center sampling, F drs Represents the features obtained by dynamic random sampling; and in order to further improve the ability of features to express image block information, through the projection layer The above 768-dimensional feature F cs ,F drs Mapped to 512 dimensions, as shown below:

[0038]

[0039] In the formula, layernorm is the normalization function and relu is the activation function; thus, the low-dimensional feature is obtained as well as in, It is represented as the feature obtained by center sampling is mapped from 768-dimensional feature to 512-dimensional feature; It is represented as the feature obtained by dynamic random sampling from 768-dimensional features to 512-dimensional features; W p Represented as the projection layer structure of the first stage attention network.

[0040] Then, and The feature vector is input into a two-stage attention network that uses two parallel fully connected layers and To obtain the underlying spatial features, and use the softmax function and mapping layer To obtain the attention score a i ; This score represents the influence of each image block on the final full-slice image prediction; where the attention score of the i-th image block feature in the full-slice image is composed of the input image block feature Determine as follows:

[0041]

[0042] Where · is the dot product, tanh and sign are the tanh function and sigmoid function respectively.

[0043] Then, according to the obtained attention score, the input image block feature F cs and F drs By sorting, we can get the top t image blocks with the highest scores, that is, and The concatenated feature representation obtained by concatenating these two features is At this point, the first stage of the attention network is completed, and the second stage of the attention network begins, and the splicing level feature representation is As the input of the attention network in the second stage.

[0044] In the second stage of the attention network, similarly, a projection layer is used The splicing-level feature representation The feature dimension of is compressed to 512 dimensions, where W q Represented as the projection layer structure of the second-stage attention network, we can get And calculate the attention score of the corresponding i-th image block feature Finally, the attention score of each image block can be weighted and aggregated to obtain the slice-level feature representation h slide , so far the second stage of attention network calculation is completed, as shown in the following formula:

[0045]

[0046] Furthermore, in S3, in order to mine tumor microenvironment information, firstly, all the whole slice images in the training set (a total of M images) are subjected to center sampling, and K center image blocks are sampled from each slice, so that M×K center image blocks can be obtained. All these center image blocks are subjected to K-means clustering again to divide them into G groups; then, the center image blocks are extracted from each group again by the center sampling method to obtain G representative image blocks; the cosine similarity distance between all image blocks in each whole slice image and these G representative image blocks is calculated using the following formula, where A and B represent the feature vectors of the image blocks, and each image block is classified into the corresponding group based on the representative image block with the closest similarity; the formula is:

[0047]

[0048] Next, for each full-slice image in the training set, the proportion of image blocks in each group is calculated to obtain G proportion values. The proportion values ​​of all training set images form an M×G matrix, and a T test is performed using M×1 labels to obtain G T statistics. Among them, the larger the absolute value of the T statistic, the more significant the difference between the full-slice images of different categories in this group of image blocks, that is, the stronger the ability to distinguish the type of image blocks; by ranking the T statistics of the G image block groups, a new group value is obtained.

[0049] Subsequently, by assigning RGB color values ​​to the group values ​​of the image blocks, specifically, mapping smaller group values ​​(such as group 1) to bluer image blocks in the RGB image, and larger group values ​​(such as group 256) to redder image blocks, the mapped image blocks are placed back to the corresponding positions of the full slice image to obtain the spatial relationship quantization map. The color of the quantization map represents the correlation between the image block and the classification result, and the residual neural network ResNet18 is used to encode the spatial relationship quantization map to extract the potential features h of the spatial relationship quantization map. spatial .

[0050] Finally, the slice-level feature representation h obtained in S2 is slide The feature representation h of the spatial relationship quantization map obtained in step S3 spatial The slices are spliced ​​and sent to the classifier for slice-level prediction to obtain the final prediction result.

[0051] It can be understood that, in a preferred embodiment of the present invention, the total image M=345, the sampled image blocks K=100, and the number of groups G=256.

[0052] Specifically, in a preferred embodiment of the present invention, Adam optimizer is used for optimization during training, and the learning rate is set to 1e-4, the weight decay is 5e-5, and dropout is set to 0.5 to reduce overfitting, and 100 iterations of training are performed using cross entropy loss as a loss function. The data set is then divided into a training set, a validation set, and a test set at a ratio of 4:1:1, the model is trained using the training set, and its performance is evaluated using the test set. Wherein, on all data sets, 5 times of cross validation are used. And in order to evaluate and compare classification performance, the present invention uses widely accepted evaluation indicators such as accuracy (Accuracy, ACC), F1 score (F1score), receiver operating characteristic (Receiver Operating Characteristic, ROC) curve area (Area UnderCurve, AUC) and C index. It is also deployed using Python and PyTorch frameworks on the network. Wherein, training and reasoning processes are performed on a workstation equipped with Intel Xeon Silver 4114 (20 cores 2.20GHz) CPU and NVIDIAGeForce TITAN V GPU.

[0053] Experimental data analysis:

[0054] Table 1 shows the comparison results of the method of the present invention and four prior art methods in molecular classification prediction and prognosis prediction of brain glioma. The present invention adopts two innovative technologies: dynamic multi-instance learning based on two-stage attention mechanism and quantification of tumor microenvironment spatial relationship information. The present invention uses 87 cases of validation set data and 90 cases of independent test set data. The images used are whole-section HE-stained images of brain glioma.

[0055] Compared with the image analysis methods used in Method 1, Method 2, Method 3 and Method 4, the present invention significantly improves the analysis ability of whole-slice images. The method of the present invention performs well in predicting key molecular markers such as IDH mutation, TERT mutation and 1p / 19q co-deletion. The prediction accuracy is improved by 2% to 5% compared with the prior art, the area under the ROC curve is improved by 3% to 6%, and the F1 score index is improved by 4% to 8%. All three evaluation indicators are significantly better than any of Method 1, Method 2, Method 3 and Method 4.

[0056] Table 2 shows the results of the method of the present invention in prognosis prediction. The C index obtained by inference is 5% to 8% higher than the C index obtained by WHO grade and IDH mutation status. In particular, combining the prediction results of the method of the present invention with the two clinical information of WHO grade and IDH mutation status can further improve the ability to predict prognosis, and the C index can be improved by about 1%.

[0057] Table 1. Prediction results of IDH, TERT, and 1p / 19q molecular typing by different methods

[0058]

[0059] Table 2 Clinical information and C index for prognosis prediction by the method of the present invention

[0060]

[0061] Figure 2 It is a schematic diagram of the Kaplan-Meier survival curve of the prognostic prediction results of the present invention, which shows the stratification quality between the low-risk and high-risk patient groups predicted by the model obtained by the present invention for the data set of this study, wherein the patients are divided into high-risk and low-risk categories according to the risk scores generated by the model, and the Log-rank test is used to evaluate the statistical differences between the two cohorts, and the p value is less than 0.0001, indicating a highly significant statistical difference.

[0062] In summary, the present invention provides an intelligent analysis method for whole-slice pathological images of gliomas based on a dual-path analysis framework. First, the dynamic multi-instance learning path based on the two-stage attention mechanism effectively selects and concentrates on analyzing image information in key areas through the two-stage attention mechanism, minimizes the interference of non-critical image blocks, and improves the model's ability to extract key features of the tumor microenvironment; secondly, the spatial relationship information quantification path uses the spatial information quantification method to enhance the model's ability to parse the spatial relationship between images, so that the model can better understand and characterize the complexity of the tumor microenvironment; finally, the entire pathological image analysis process is automated, greatly reducing the need for manual intervention, and improving the efficiency and robustness of the prediction, making the present invention more convenient and quick in clinical applications, helping to improve the speed and quality of medical decision-making, and verification on public data sets reflects the significant advantages of the present invention in improving the accuracy of glioma molecular typing and prognosis prediction.

[0063] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. An intelligent analysis method for whole-slice pathological images of glioma based on a dual-path analysis framework, characterized in that: The specific steps include: S1. Preprocessing of the whole slice image: using the channel threshold segmentation algorithm, each whole slice image is divided into different foreground tissue areas and background areas, and the foreground tissue area is cut into a number of image blocks of fixed size, and then the feature vector of each image block is extracted to form feature clustering; S2. Establish a dynamic sampling mechanism to perform dynamic multi-instance learning on the feature clustering of each full-slice image using a two-stage attention network, and finally obtain the slice-level feature representation of the full-slice image; S3. Quantify the spatial relationship information of the tumor microenvironment to obtain the feature representation of the spatial relationship quantization map, and concatenate it with the slice-level feature representation in S2 to obtain the final prediction result.

2. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: In the above S1, dividing each whole-slice image into different foreground tissue regions and background regions specifically comprises the following steps: K1, read the full slice image, convert it to HSV color space and perform median filtering; K2, binarize the image at the set saturation channel threshold; K3, performing a morphological closing operation on the binarized image in K2.

3. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 2, characterized in that: The saturation channel threshold ranges from 10 to 30.

4. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: The dynamic sampling mechanism includes center sampling and dynamic random sampling.

5. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: The two-stage attention network uses two parallel fully connected layers L and U to obtain the underlying spatial features, and uses a softmax function and a mapping layer M to obtain the attention score. In the formula, a i is the attention score of the i-th image block feature in the full slice image, · is the dot product, tanh and sigm are tanh and sigmoid functions respectively, and They are respectively the image block features of different inputs.

6. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: In the first stage of the attention network, two dynamic sampling mechanisms are used to extract K image block features of two sizes of 768 dimensions respectively, and the two 768-dimensional features are mapped to 512 dimensions through the first stage attention network projection layer. The attention scores of the two features mapped to 512 dimensions are calculated respectively, and the initial 768-dimensional image block features are sorted according to the attention scores. The image block features with higher scores extracted by the two sampling mechanisms are selected and spliced ​​to obtain a 768-dimensional splicing-level feature representation, completing the first stage of the attention network.

7. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 6, characterized in that: In the second-stage attention network, the dimension of the splicing-level feature representation is compressed to 512 dimensions again using the second-stage attention network projection layer, and the attention score of the corresponding i-th image block feature is calculated, thereby obtaining the slice-level feature representation and completing the second-stage attention network calculation, as shown in the following formula: In the formula, h slide is the slice-level feature representation, s j is the attention score, It is the image block level feature representation.

8. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: The S3 specifically includes the following steps: Y1. Perform center sampling on the M full slice images in the training set, sample K center image blocks from each slice, obtain M×K center image blocks, and perform K-means clustering on all center image blocks again to divide them into G groups; Y2. Extract the central image block in each group again by the central sampling method to obtain G representative image blocks; Y3. Calculate the similarity distances between all image blocks in each full-slice image and the G representative image blocks, and assign the representative image blocks with the closest similarity to corresponding groups; Y4. Calculate the proportion of image blocks in each group for each full-slice image in the training set to obtain G proportion values, organize the proportion values ​​of all training set images into an M×G matrix, and perform a T test using M×1 labels to obtain G T statistics, and rank the T statistics of the G image block groups to obtain a new group value; Y5. Assign RGB color values ​​to the group values ​​of the image blocks, put the mapped image blocks back to the corresponding positions of the full slice image to obtain the spatial relationship quantization map, and use the residual neural network ResNet18 to encode the spatial relationship quantization map to extract the potential features of the spatial relationship quantization map; Y6. Concatenate the slice-level feature representation with the feature representation of the spatial relationship quantization map, and send them into a classifier for slice-level prediction to obtain a final prediction result.

9. The method for intelligent analysis of whole-slice pathological images of glioma based on a dual-path analysis framework as claimed in claim 1, characterized in that: The similarity distance calculation uses the cosine similarity distance formula: Where A and B represent the feature vectors of the image blocks.

Images