WSI image continuous retrieval method based on distance consistency playback

Abstract

The invention discloses a WSI image continuous retrieval method based on distance consistency playback. The method comprises the following steps: acquiring a data stream E; performing feature extraction on the WSI image data sets at different moments to obtain feature block sets H at different moments; in combination with a WSI encoder, feature blocks at each moment in the feature block set H are processed, full-slice representations at all moments are obtained, and WSI encoder parameter updating and local buffer area construction are carried out; and constructing an index structure based on the full slice representation at all moments, and performing WSI image continuous retrieval in combination with the WSI encoder after parameter updating. According to the method, the continuous retrieval of the pathological image is realized, the disastrous forgetting risk is reduced, the consistency of returning the old task to the queue before and after learning the new task is ensured, and the pathological image retrieval efficiency and accuracy are further improved.

Description

Technical Field

[0001] The present invention relates to the field of digital image processing, and more particularly to a WSI image continuous retrieval method based on distance consistency replay. Background Art

[0002] With the rapid development of computational pathology and artificial intelligence, content-based histopathology image retrieval (CBHIR) technology has become an important tool in pathological diagnosis and research. CBHIR can return whole slide images (WSI) with similar content to the query pathological image from an established database.

[0003] However, although the current CBHIR technology performs well on static databases, it forgets previously learned knowledge when learning new data, a phenomenon known as catastrophic forgetting in artificial intelligence systems. Catastrophic forgetting manifests itself as inconsistencies in the return queue of old tasks before and after learning new tasks. This is quite serious for image retrieval systems in medical scenarios. The inconsistency of these results over time may seriously damage the reliability of the system, which inevitably limits the practical application of CBHIR systems in clinical data increment scenarios.

[0004] Currently, continual learning has been proposed to mitigate catastrophic forgetting when continuously learning in non-static data streams. It can be roughly divided into three categories: replay methods, regularization-based methods, and parameter isolation methods. However, there are few studies on continuous learning retrieval tasks in natural scenes, and most of the research focuses on tasks such as image classification and semantic segmentation. However, image retrieval tasks are more fine-grained than classification tasks, and directly applying the continuous learning method of classification tasks to retrieval tasks cannot achieve satisfactory results.

[0005] In addition, in the field of digital pathology, there is a continuous learning method for pathological image classification problems, which establishes a buffer zone for the image area and replays the data through the "separation-recombination" method, while achieving data enhancement to a certain extent. However, this method cannot be applied to the retrieval task of pathological images, because segmenting pathological images into small tissue areas cannot obtain fine features at the full slice level.

[0006] Therefore, how to design a WSI image continuous retrieval method based on distance consistency replay to reduce the risk of catastrophic forgetting and ensure the consistency of the return queue of old tasks before and after learning new tasks is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a WSI image continuous retrieval method based on distance consistency replay, which processes multiple pathological image data sets at different times by continuously acquiring data streams, and uses a reservoir random sampling algorithm to effectively reduce the burden of high-dimensional data processing, thereby improving the accuracy, stability and efficiency of WSI image retrieval.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A WSI image continuous retrieval method based on distance consistency replay, comprising:

[0010] S1. Obtain data stream E; where E={E1,……,E t}, is the data set at time t, x i is the whole slice image, y i is the category label of the full slice image, N t is the number of full-slice image-label pairs;

[0011] S2. Extract features from the WSI image datasets at different times to obtain feature block sets H at different times; where H = {H1, ..., H t}, H t is the feature block at time t;

[0012] S3, combined with the WSI encoder, the feature blocks at each moment in the feature block set H are processed respectively to obtain the full slice representation at all moments, and the WSI encoder parameters are updated and the local buffer is constructed;

[0013] S4. Based on the full slice representation at all times, an index structure is constructed, and combined with the WSI encoder with updated parameters, continuous retrieval of WSI images is performed.

[0014] Furthermore, in S3, the feature blocks at the initial moment are processed, including:

[0015] Based on the feature block set H1, the batch size is set, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the initial moment;

[0016] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0017]

[0018] Where f1 is the full slice encoder at the initial moment.

[0019] Furthermore, the local buffer is constructed, including:

[0020] Perform reservoir random sampling on the feature block set H1 to obtain samples And store it in the local buffer M2;

[0021] The sample Input the full slice encoder to get samples Whole-slice characterization:

[0022]

[0023] Based on sample Full slice characterization, computational samples The distance matrix D1 between each sample in is stored in the local buffer M2;

[0024] Among them, the calculation sample The distance matrix D1 between each sample in is:

[0025]

[0026] in, and are the i-th and j-th representations in C1, respectively, and d(·) is the Euclidean distance.

[0027] Furthermore, the WSI encoder parameters are updated, including:

[0028] Calculate the compound loss function value:

[0029]

[0030] in, To represent the corresponding true category label, To represent the corresponding predicted category label, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function;

[0031] The pairwise loss function:

[0032]

[0033] in, and They are The mth sample and the nth sample in , is the true label, S is the total number of samples, d h is the feature dimension, y mn is the correlation between the mth sample and the nth sample, y mn =1 indicates correlation, y mn=0 means no correlation.

[0034] Furthermore, in S3, the feature blocks at the intermediate time are processed, including:

[0035] Based on the feature block set H e and local buffer M e The replay samples in are used to set the batch size, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the intermediate time;

[0036] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0037]

[0038] In the formula, f e is the full slice encoder at time e, e∈(1,t).

[0039] Furthermore, the local buffer is constructed, including:

[0040] For the feature block set H at time e e and local buffer M e Replay samples in the reservoir and perform random sampling to obtain samples and

[0041] The sample and Merge to obtain samples And store it in the local buffer M e+1 ;

[0042] The sample Input the full slice encoder to get samples Whole-slice characterization:

[0043]

[0044] Based on sample Full slice characterization, computational samples The distance matrix D between each sample in e And store it in the local buffer M e+1 ;

[0045] Among them, the calculation sample The distance matrix D between each sample in e for:

[0046]

[0047] in, and C e The i-th and j-th representations in , d(·) Euclidean distance.

[0048] Furthermore, the WSI encoder parameters are updated, including:

[0049] Calculate the compound loss function value:

[0050]

[0051] Among them, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor;

[0052] The distance consistency loss function:

[0053]

[0054] Among them, D e-1 is the local buffer M e Replay matrix.

[0055] Furthermore, in S3, the feature blocks at the final moment are processed, including:

[0056] Based on the feature block set H t and local buffer M t The replay samples in are used to set the batch size, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the final moment;

[0057] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0058]

[0059] In the formula, f t is the full slice encoder at the final moment.

[0060] Furthermore, the WSI encoder parameters are updated, including:

[0061] Calculate the compound loss function value:

[0062]

[0063] Among them, L p (·) is the pairwise loss function, L CE(·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor;

[0064] The distance consistency loss function:

[0065]

[0066] Among them, D t-1 is the local buffer M t Replay matrix.

[0067] Furthermore, the S4 further includes:

[0068] The Spearman rank correlation coefficient SRC or Kendall rank correlation coefficient KRC were used to construct statistical indicators for consistency evaluation of WSI image continuous retrieval;

[0069] Calculate the complete SRC or KRC value:

[0070]

[0071] in, is the average return queue consistency of all tasks, n is the total number of tasks, is the average return queue consistency of task g after all subsequent tasks, p g,k It means that after task k is completed, the SRC or KRC value of the test set retrieval return queue of task g is compared with the SRC or KRC value just after task g is completed.

[0072] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0073] Beneficial effects:

[0074] 1. Process multiple pathological image datasets at different times by continuously acquiring data streams.

[0075] It allows the system to dynamically adapt to data characteristics that change over time and to maximize learning efficiency and effectiveness at different stages by adjusting processing strategies.

[0076] 2. The encoder parameters are continuously updated at each moment, and the loss function is constructed differently at different moments, which enables the model to continuously evolve to cope with emerging data patterns. The customized loss function ensures that the model can be optimized for specific tasks at each stage, thereby improving the overall accuracy and generalization ability.

[0077] 3. By replaying historical samples, the problem of "catastrophic forgetting" can be alleviated, that is, the model will not forget the characteristics of old data while learning new data, and maintain long-term memory. Ensure that good performance and stability can be maintained even when the data distribution changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0079] Figure 1 A flow chart of a WSI image continuous retrieval method based on distance consistency replay provided by an embodiment of the present invention;

[0080] Figure 2 A schematic diagram of specific steps of a WSI image continuous retrieval method based on distance consistency replay provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] like Figure 1 As shown, this embodiment provides a WSI image continuous retrieval method based on distance consistency replay, including:

[0083] S1. Obtain data stream E; where E={E1,……,E t}, is the data set at time t, x i is the whole slice image, y i is the category label of the full slice image, N t is the number of full-slice image-label pairs;

[0084] S2. Extract features from the WSI image datasets at different times to obtain feature block sets H at different times; where H = {H1, ..., H t}, H t is the feature block at time t;

[0085] S3, combined with the WSI encoder, the feature blocks at each moment in the feature block set H are processed respectively to obtain the full slice representation at all moments, and the WSI encoder parameters are updated and the local buffer is constructed;

[0086] S4. Based on the full slice representation at all times, an index structure is constructed, and combined with the WSI encoder with updated parameters, continuous retrieval of WSI images is performed.

[0087] This method achieves continuous retrieval of pathological images, reduces the risk of catastrophic forgetting, ensures the consistency of the old tasks returned to the queue before and after learning new tasks, and further improves the efficiency and accuracy of pathological image retrieval.

[0088] like Figure 2 As shown, the above steps and related features are further described in detail below;

[0089] In this embodiment S1, new data is continuously acquired from various sources, such as pathological images from hospitals, laboratories or scientific research institutions. These images are pre-processed to form a data stream containing multiple time points.

[0090] The dataset E at each moment t It contains all the full-slice images at that moment and their corresponding category labels, which may be the types of various diseases or other related pathological features. The data stream reflects the latest developments and research trends in the field of pathology, covering various types of lesions and pathogens, and providing comprehensive information support for pathologists.

[0091] In this embodiment S2, multiple links such as image preprocessing, feature selection, feature quantization, feature fusion and feature storage are involved, aiming to efficiently extract key information of the image.

[0092] First, the accuracy of feature extraction is improved by performing preprocessing operations such as scaling, denoising and color correction on the original WSI image data at different times. Specifically, scaling: using methods such as bicubic interpolation to reduce large-size WSI images to a suitable size, reducing the computational burden while maintaining sufficient resolution. Denoising: using methods such as Gaussian filtering and median filtering to remove noise in the image, such as artifacts during the scanning process and mechanical errors of the scanner, to improve the accuracy of feature extraction. Color correction: using color equalization, white balance and other technologies to correct color deviations caused by different scanning conditions, ensuring that all images are compared in the same color space.

[0093] Next, a pre-trained convolutional neural network (CNN) is used to automatically learn and extract key features in the image, such as texture, shape, and color. In this embodiment, different network layers are selected as feature extraction layers. If you are concerned about low-level features, such as texture and edges, you can choose a shallower network layer; if you are concerned about high-level features, such as object recognition, you can choose a deeper network layer.

[0094] Subsequently, the extracted features are converted into numerical vectors, and the computational complexity is reduced and the retrieval efficiency is improved through operations such as feature dimensionality reduction and normalization. In feature dimensionality reduction, principal component analysis (PCA) and linear discriminant analysis (LDA) methods are used to compress feature dimensions and reduce computational overhead. In normalization, the features are standardized to meet certain statistical assumptions, such as zero mean and unit variance.

[0095] At the same time, if there are multiple levels of feature extraction results, the features at different levels are fused to obtain a more comprehensive and accurate feature representation. Finally, the extracted features are saved to form a feature block set H for subsequent retrieval and comparison operations.

[0096] In this embodiment S3, in combination with the WSI encoder, the feature blocks at each moment in the feature block set H are processed respectively to obtain the full slice representation at all moments, and the WSI encoder parameters are updated and the local buffer is constructed;

[0097] The following further describes the above S3 in detail from the initial moment, the middle moment and the final moment, so as to clarify the difference and connection between the processing of the feature blocks at different moments by the technical solution of this embodiment;

[0098] Process the feature blocks at the initial moment, including:

[0099] Based on the feature block set H1, the batch size is set, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the initial moment;

[0100] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0101]

[0102] Where f1 is the full slice encoder at the initial moment.

[0103] Local buffer construction, including:

[0104] Perform reservoir random sampling on the feature block set H1 to obtain samples And store it in the local buffer M2;

[0105] The sample Input the full slice encoder to get samples Whole-slice characterization:

[0106]

[0107] Based on sample Full slice characterization, calculation of samples The distance matrix D1 between each sample in is stored in the local buffer M2;

[0108] Among them, the calculation sample The distance matrix D1 between each sample in is:

[0109]

[0110] in, and are the i-th and j-th representations in C1, respectively, and d(·) is the Euclidean distance.

[0111] WSI encoder parameter updates, including:

[0112] Calculate the compound loss function value:

[0113]

[0114] in, To represent the corresponding true category label, To represent the corresponding predicted category label, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function;

[0115] The pairwise loss function:

[0116]

[0117] in, and They are The mth sample and the nth sample in , is the true label, S is the total number of samples, d h is the feature dimension, y mn is the correlation between the mth sample and the nth sample, y mn =1 indicates correlation, y mn =0 means no correlation.

[0118] Process the feature blocks at the intermediate moment, including:

[0119] Based on the feature block set He and local buffer M e The replay samples in are used to set the batch size, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the intermediate time;

[0120] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0121]

[0122] In the formula, f e is the full slice encoder at time e, e∈(1,t).

[0123] Local buffer construction, including:

[0124] For the feature block set H at time e e and local buffer M e Replay samples in the reservoir and perform random sampling to obtain samples and

[0125] The sample and Merge to obtain samples And store it in the local buffer M e+1 ;

[0126] The sample Input the full slice encoder to get samples Whole-slice characterization:

[0127]

[0128] Based on sample Full slice characterization, computational samples The distance matrix D between each sample in e And store it in the local buffer M e+1 ;

[0129] Among them, the calculation sample The distance matrix D between each sample in e for:

[0130]

[0131] in, and C e The i-th and j-th representations in , d(·) Euclidean distance.

[0132] WSI encoder parameter updates, including:

[0133] Calculate the compound loss function value:

[0134]

[0135] Among them, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor;

[0136] The distance consistency loss function:

[0137]

[0138] Among them, D e-1 is the local buffer M e Replay matrix.

[0139] Process the feature blocks at the final moment, including:

[0140] Based on the feature block set H t and local buffer M t The replay samples in the , set the batch size, input each batch of data into the full slice encoder, and obtain the full slice representation at the final moment;

[0141] Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as:

[0142]

[0143] In the formula, f t is the full slice encoder at the final moment.

[0144] WSI encoder parameter updates, including:

[0145] Calculate the compound loss function value:

[0146]

[0147] Among them, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor;

[0148] The distance consistency loss function:

[0149]

[0150] Among them, D t-1 is the local buffer M t Replay matrix.

[0151] In the above steps, at the initial moment, the local buffer is constructed by randomly sampling the reservoir to obtain samples and store them in the local buffer. At the intermediate and final moments, in addition to random sampling of the reservoir, the feature block set and the replay samples in the local buffer are merged to obtain samples. This differentiated local buffer construction method can flexibly adjust the selection and combination of samples according to the needs of different moments, enhance the diversity of samples, and help improve the retrieval effect.

[0152] And at the initial, middle and final moments, the WSI encoder parameters are continuously updated to adapt to the task requirements at different moments.

[0153] Especially at the initial moment, the pairwise loss function and the cross entropy loss function are used to optimize the WSI encoder. In the middle and final moments, in addition to the above two loss functions, the distance consistency loss function is also introduced to further improve the performance of the WSI encoder. Through the comprehensive use of the pairwise loss function, the cross entropy loss function and the distance consistency loss function, the parameters of the WSI encoder are continuously adjusted to enable it to better cope with the task requirements at different times, and the robustness and generalization ability of the system are improved.

[0154] In S4 of this embodiment, an index structure is constructed based on the full slice representation at all times, and continuous retrieval of WSI images is performed in combination with the WSI encoder after parameter update.

[0155] In order to measure the consistency of the full-slice encoder in retrieving the return queue on the old task after learning the new task, the Spearman rank correlation coefficient (SRC) or Kendall rank correlation coefficient (KRC) is used to construct statistical indicators. The specific details are as follows:

[0156] The Spearman rank correlation coefficient (SRC) is a nonparametric statistical method based on ranking, which is used to measure the correlation between two variables. Its calculation formula is as follows:

[0157]

[0158] Among them, d i is the ranking difference of the i-th pair of data, and n is the number of samples. The value of SRC ranges between -1 and 1. The closer the value is to 1, the stronger the positive correlation between the two variables is, the closer it is to -1, the stronger the negative correlation is, and the closer it is to 0, the no correlation is.

[0159] Kendall rank correlation coefficient (KRC) is a non-parametric statistical method based on ranking, which is used to measure the correlation between two variables. Its calculation formula is as follows:

[0160]

[0161] Among them, C is the number of consistent pairs, D is the number of inconsistent pairs, and n is the number of samples. The value range of KRC is also between -1 and 1. The closer the value is to 1, the stronger the positive correlation between the two variables is, the closer it is to -1, the stronger the negative correlation is, and the closer it is to 0, the no correlation is.

[0162] When used specifically, calculate the complete SRC or KRC value:

[0163]

[0164] in, is the average return queue consistency of all tasks, n is the total number of tasks, is the average return queue consistency of task g after all subsequent tasks, p g,k It means that after task k is completed, the SRC or KRC value of the test set retrieval return queue of task g is compared with the SRC or KRC value just after task g is completed.

[0165] In addition, this embodiment provides the results of comparing the technical solution of the present application with other technical solutions based on indicators such as average precision mean mAP, K recall rate, and K precision.

[0166] Table 1 compares the results of fully supervised training and fine-tuning training:

[0167] Table 1

[0168]

[0169] Tables 2-4 show the results of comparing other continuous learning methods:

[0170] Table 2 shows the comparison results of the buffer for 5 pathological images:

[0171] Table 2

[0172]

[0173] Table 3 shows the comparison results of the buffer with 10 pathological images:

[0174] Table 3

[0175]

[0176]

[0177] Table 4 shows the comparison results of the buffer with 15 pathological images:

[0178] Table 4

[0179]

[0180] This method achieves continuous retrieval of pathological pathology images. Table 1 shows the results of this method with full supervision training and fine-tuning training, which represent the upper and lower limits of the performance of the continuous learning method, respectively. The results show that this method preliminarily achieves continuous retrieval of pathological pathology images;

[0181] Tables 2-4 compare this method with various other types of continuous learning methods. The results show that this method achieves better continuous learning performance than other methods. At the same time, the SRC and KRC indicators in the above tables measure the ability of different continuous learning methods to maintain the consistency of past task queues, which shows that our method achieves better queue consistency.

[0182] Based on the above embodiments, the image continuous retrieval method in the technical solution of this application realizes the effective management and retrieval of a large number of pathological image data streams by integrating reservoir random sampling, full slice encoding and distributed computing technologies. It uses the distance consistency replay strategy to maintain efficient maintenance of historical data memory while continuously accepting new data sets, ensuring the stability and adaptability of the retrieval system.

[0183] In specific operations, the system extracts features from WSI images at different times and adopts a reservoir random sampling strategy, combines full-slice coding to generate high-order representations, and then integrates new and old data features through a merging strategy to form a unified representation set. Further generate and store distance matrices to lay the foundation for building an efficient index structure. Ultimately, the system is able to continuously perform image retrieval, which not only significantly improves retrieval efficiency and accuracy, but also successfully solves the problem of knowledge forgetting in continuous learning, ensuring the consistency between retrieval results and old knowledge when introducing new data, and achieving a balance between stability and plasticity.

[0184] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0185] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A WSI image continuous retrieval method based on distance consistency replay, characterized in that: include: S1. Obtain data stream E; where E={E1,……,E t }, is the data set at time t, x i is the whole slice image, y i is the category label of the full slice image, N t is the number of full-slice image-label pairs; S2. Extract features from the WSI image datasets at different times to obtain feature block sets H at different times; where H = {H1, ..., H t }, H t is the feature block at time t; S3, combined with the WSI encoder, the feature blocks at each moment in the feature block set H are processed respectively to obtain the full slice representation at all moments, and the WSI encoder parameters are updated and the local buffer is constructed; S4. Based on the full slice representation at all times, an index structure is constructed, and combined with the WSI encoder with updated parameters, continuous retrieval of WSI images is performed.

2. The WSI image continuous retrieval method based on distance consistency replay according to claim 1 is characterized in that: In S3, the feature blocks at the initial moment are processed, including: Based on the feature block set H1, the batch size is set, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the initial moment; Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as: Where f1 is the full slice encoder at the initial moment.

3. The WSI image continuous retrieval method based on distance consistency replay according to claim 2 is characterized in that: Local buffer construction, including: Perform reservoir random sampling on the feature block set H1 to obtain samples And store it in the local buffer M2; The sample Input the full slice encoder to get samples Whole-slice characterization: Based on sample Full slice characterization, computational samples The distance matrix D1 between each sample in is stored in the local buffer M2; Among them, the calculation sample The distance matrix D1 between each sample in is: in, and are the i-th and j-th representations in C1, respectively, and d(·) is the Euclidean distance.

4. The WSI image continuous retrieval method based on distance consistency replay according to claim 2 is characterized in that: WSI encoder parameter updates, including: Calculate the compound loss function value: in, To represent the corresponding true category label, To represent the corresponding predicted category label, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function; The pairwise loss function: in, and They are The mth sample and the nth sample in , is the true label, S is the total number of samples, d h is the feature dimension, y mn is the correlation between the mth sample and the nth sample, y mn =1 indicates correlation, y mn =0 means no correlation.

5. The WSI image continuous retrieval method based on distance consistency replay according to claim 1, characterized in that: In S3, the feature blocks at the intermediate time are processed, including: Based on the feature block set H e and local buffer M e The replay samples in are used to set the batch size, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the intermediate time; Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as: In the formula, f e is the full slice encoder at time e, e∈(1,t).

6. The WSI image continuous retrieval method based on distance consistency replay according to claim 5, characterized in that: Local buffer construction, including: For the feature block set H at time e e and local buffer M e Replay samples in the reservoir and perform random sampling to obtain samples and The sample and Merge to obtain samples And store it in the local buffer M e+1 ; The sample Input the full slice encoder to get samples Whole-slice characterization: Based on sample Full slice characterization, computational samples The distance matrix D between each sample in e And store it in the local buffer M e+1 ; Among them, the calculation sample The distance matrix D between each sample in e for: in, and C e The i-th and j-th representations in , d(·) Euclidean distance.

7. The WSI image continuous retrieval method based on distance consistency replay according to claim 5, characterized in that: WSI encoder parameter updates, including: Calculate the compound loss function value: Among them, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor; The distance consistency loss function: Among them, D e-1 is the local buffer M e Replay matrix.

8. The WSI image continuous retrieval method based on distance consistency replay according to claim 1, characterized in that: In S3, the feature blocks at the final moment are processed, including: Based on the feature block set H t and local buffer M t The replay samples in are used to set the batch size, and the data of each batch is input into the full-slice encoder to obtain the full-slice representation at the final moment; Among them, the data of batch b Input the full slice encoder to get data The full slice representation of is expressed as: In the formula, f t is the full slice encoder at the final moment.

9. A method for continuous retrieval of pathological images based on distance consistency replay according to claim 8, characterized in that: WSI encoder parameter updates, including: Calculate the compound loss function value: Among them, L p (·) is the pairwise loss function, L CE (·) is the cross entropy loss function, L DC (·) is the distance consistency loss function, α is the weight factor; The distance consistency loss function: Among them, D t-1 is the local buffer M t Replay matrix.

10. The WSI image continuous retrieval method based on distance consistency replay according to claim 1, characterized in that: The S4 further includes: The Spearman rank correlation coefficient SRC or Kendall rank correlation coefficient KRC were used to construct statistical indicators for consistency evaluation of WSI image continuous retrieval; Calculate the complete SRC or KRC value: in, is the average return queue consistency of all tasks, n is the total number of tasks, is the average return queue consistency of task g after all subsequent tasks, p g,k It means that after task k is completed, the SRC or KRC value of the test set retrieval return queue of task g is compared with the SRC or KRC value just after task g is completed.

Images