Digital pathological section diagnosis method based on link algorithm

Abstract

The invention relates to the technical field of disease diagnosis, in particular to a digital pathological section diagnosis method based on a link algorithm, and the method comprises the following steps: creating a tissue; creating a link algorithm; marking the digital pathological section data; checking the digital pathological section data; corresponding parameters are set for a link algorithm, and training is carried out; if the training effect of the link algorithm meets the requirement, cascading a new module, and if the training effect of the link algorithm does not meet the requirement, continuing to enrich the data set until the requirement is met; and after the whole link algorithm is trained, exporting a trained model, and reasoning the digital pathological section by using the model. According to the method, a link algorithm is created, so that operators influencing the reasoning speed and memory occupation are removed while the precision is kept; by creating a link algorithm, a user can train a data set in a highly self-defined manner in combination with various downstream tasks so as to generate a model meeting own requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of disease diagnosis, and in particular to a diagnosis method for digital pathological sections based on a link algorithm. Background Art

[0002] Pathology, a crucial tool for disease diagnosis, faces a severe shortage of pathology professionals in my country. With the continued rise in cancer incidence in recent years, leveraging artificial intelligence (AI) to help doctors achieve accurate diagnoses and improve both efficiency and quality has become a key research focus and a hot topic in the field of pathology worldwide.

[0003] Current diagnostic methods have the following problems: 1. The model is too simple, resulting in insufficient generalization of learned features and inability to accurately locate certain pathological tissue features. The model is too complex, and its architecture and memory access method are not friendly to hardware deployment. 2. In practical applications, operations such as model training are typically run in a containerized manner on compute nodes, and these containers occupy the local storage space of the compute nodes. Since a compute node is typically shared by multiple users, when the local storage space of a node is heavily occupied or even exhausted by containers, all user services on the node may experience anomalies or even become completely unavailable. 3. The actual number of pixels in pathological images can reach hundreds of thousands or even higher, which can lead to the following problems: a. High storage usage. A single pathology image can sometimes take up several GB. b. Slow training and reasoning speed; c. Previewing or viewing slices is slow; 4. Pathological tissue annotation is difficult. To enable the model to accurately segment or identify the corresponding structural tissue, corresponding rectangular, circular, polygonal and other annotation boxes have emerged. However, in order to correctly frame the area, these boxes require the annotator to repeatedly adjust and modify the same area. 5. Data-driven models conflict with patient privacy; 6. When using models to diagnose pathological sections, tasks such as classification, target detection, and segmentation are often dynamically combined and cascaded in the workflow. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: in order to solve the technical problem that the existing model is too simple and affects the accuracy of pathological diagnosis results, the present invention provides a diagnostic method for digital pathological sections based on a link algorithm. By improving the diagnostic method, the accuracy of pathological diagnosis results can be improved.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a diagnostic method for digital pathological sections based on a link algorithm, comprising the following steps: S1. Create tissue and upload digital pathology slide datasets; S2, create link algorithm; S3, using annotation tools to annotate the digital pathology slide data in S1; S4, review the digital pathology slide data that have been annotated in S3; S5, setting corresponding parameters for the link algorithm created by S2, and training the link algorithm using the digital pathology slide data reviewed and annotated by S4; S6. If the training effect of the link algorithm in S5 meets the requirements, repeat S2 as needed and cascade new modules. If the training effect of the link algorithm in S5 does not meet the requirements, repeat S3 and S4 to supplement the data and continue to enrich the data set until the requirements are met; S7. When the algorithm training of the entire link is completed, the trained model is exported and used to infer the digital pathology slices, locate and classify the corresponding tissues, and export their data reports.

[0006] Therefore, by creating a link algorithm, operators that affect inference speed and memory usage can be removed while maintaining its accuracy; the storage space of computing nodes can be managed to ensure the stability of computing node services; by creating a link algorithm, users can train data sets in a highly customized manner in combination with various downstream tasks to generate models that meet their needs.

[0007] Furthermore, the link algorithm is an improved RTMDet model, which includes: Backbone consists of CSPNextBlock and Neck consists of CSPNextPAFPN; The S2 comprises the following steps: S2-1. When digital pathology slide data is used as the input of the target detection network, the size of the input digital pathology slide data is cropped or resized according to the memory of the hardware built by the RTMDet model. The RTMDet model network is improved. A higher-resolution shallow feature map is introduced in the C2 stage of Backbone. The DSPPE module is designed in the C5 stage to make it more suitable for capturing the receptive field of cells. The feature map is sent to the CSP module through a connection operation at the Neck for convolution. After the feature map is downsampled by 32 times, the HAD module is added to further enhance the features. S2-2. RTMDet model output and loss design: The output of the object detection model includes: object analogy and object location. Therefore, in S2-1, the purpose of cropping or resizing the input digital pathology slide data is because digital pathology slide scans usually have a high resolution.

[0008] Furthermore, the S2-1 includes the following steps: S2-1-1, CARAFE upsampling operator: In the Neck feature fusion network path, the plug-and-play CARAFE operator is used, which predicts the convolution kernel required for upsampling through convolution operation. , kernels at different positions can flexibly combine the semantic information of feature maps , thus obtaining more accurate upsampling results; S2-1-2, DSPPE expanded spatial pyramid pooling fast module: uses convolutional pooling with different expansion rates to improve SPPE and restricts the output to a square receptive field; S2-1-3, after Neck, the output feature maps are: the input downsampling size The HAD module is added after the feature map is downsampled to 32 times to enhance the ability of the RTMDet model to interpret contextual information. The model consists of a Convolution reduces the channel dimension and module parameters, and the output after dimensionality reduction is fed in parallel to three convolution branches and one residual attention branch. The expansion rates of the three parallel convolutions are , while residual attention consists of spatial attention and channel attention. Therefore, in S2-1-1, the purpose of using the plug-and-play CARAFE operator is: in the path of the Neck feature fusion network, feature maps of different resolutions usually need to be upsampled to the same resolution before connection. However, the upsampling commonly used in FPN (such as: nearest neighbor interpolation method, deconvolution, sub-pixel convolution) has problems such as being affected by background interference, easy loss of key cell information, and introduction of invalid information; in S2-1-2, the purpose of using convolution pooling with different expansion rates to improve SPPE is: since the receptive field of the output of ordinary convolution operation is fixed, the extracted feature area is also limited, and maximum pooling limits the spatial positioning information.

[0009] Furthermore, in S2-1-1, the convolution kernel The expression is: ; in: represents the kernel prediction module, which consists of a channel compression, a content encoder, and a kernel normalization layer; for l-centered neighborhood; Semantic information The expression is: ; in, is a content-aware reorganization module that converts Chinese Neighborhood and Kernel Reorganization, including = ; 'for The center position of the target , r= ; In S2-1-2, the maximum distance between two non-zero values in the feature map The expression is: ; in, is the receptive field radius of the i-th layer; In S2-1-3, the DSPPF module multi-scale feature fusion output The calculation formula is: ; in, represents a dilated convolution with a dilation rate of r, It is a normal convolution. represents the feature concatenation operation, and ECA is the efficient channel attention module.

[0010] Furthermore, in S2-2, the dynamic soft label allocation policy function The expression is: ; in: 、 、 They are 1, 3, and 1 respectively; 、 、 Represent the classification loss function, regression loss function, and regional loss function respectively; Represents cross-entropy loss; P represents the classification probability predicted by the model; Represents the intersection-over-union ratio of the predicted box and the true box ( ), IoU = the intersection area of the predicted box and the real box / the union area of the predicted box and the real box, Represents the geometric alignment metric between the predicted box and the true box; is the center coordinate of the prediction box; is the center coordinate of the real box; for the region prior, the soft center prior is used Instead of a fixed center prior.

[0011] Furthermore, the platform resources are scheduled as follows: In S1, each GPU is assigned a unique URI and encapsulated in a resource pool; In S2, each user's request is added to a queue. The system simulates the user's input data and increases the batch size from 2 until the largest integer and even batch size is found, and then allocates GPU resources to it. In S3, when all GPU resources are occupied, the queue is locked and no new requests are accepted until GPU resources are released.

[0012] Furthermore, the step S3 includes the following steps: S3-1. Image compression based on HEVC: When constructing an image pyramid, each layer of the image is considered as multiple When the camera takes a picture of the slide, each tile is put into the GPU and compressed using HEVC. Each time it is compressed, the pyramid level to which the tile belongs and the coordinates of the layer are recorded. When the entire image pyramid is built, each layer of the pyramid is also compressed accordingly, and the final output is a pathology image in .mpi format. S3-2. Image scaling and slicing during training and inference: S3-2-1. If the user slice data uses the .mpi format based on S3-1, use the corresponding decoder to first obtain the corresponding layer according to the mpp according to the level and coordinates recorded during compression, and then decode the corresponding tile according to the coordinate order; S3-2-2, if the user slice data format is not .mpi format, the image is converted into Cut into several tiles, and set the overlap area to 10%-20% to prevent feature points from being cut off; S3-3, adaptive screen resolution decoding method: for the digital pathology slices read, measure their length and width, and Divide it integerly and estimate how many tiles it can be divided into. Record the corresponding coordinates. Based on the center of the display screen and the ratio of the screen resolution to the original two-tile resolution, select the tile with the corresponding coordinates for decoding and presentation. Among them: In S3-2-1 and S3-2-2, during training, the tiles are scaled to the training size specified by the RTMDet model for training, and during inference, the tiles are merged. Thus, combined with HEVC, the storage space of digital pathology slices is compressed, reducing the storage pressure of the server; the image is scaled to Or cut into many The training and reasoning of images are accelerated by the adaptive screen resolution tile decoding method to speed up the browsing of images.

[0013] Furthermore, the S4 includes the following steps: S4-1. Model architecture: Improved based on Sam, including image encoder, hint encoder, and mask encoder; S4-2, loss function L: use the unweighted sum of cross-entropy loss and dice loss; The calculation formula of the loss function L is: ; Among them: In S4-1, the image encoder is composed of Vit-Base, and the prompt encoder encodes the point and box prompt data into sparse embedding and dense embedding. If the input is a point and a box, the output shape is , the mask encoder uses the embedding of the image encoder and the hint encoder to generate multiple segmentation masks using the Transformer architecture and simultaneously evaluate the quality of each mask; In S4-2, S and G are the segmentation results and the ground truth. 、 Voxel The predicted segmentation results and Ground truth, represents the cross entropy loss function; Denotes the Dice loss function. This can significantly improve the labeling efficiency.

[0014] Furthermore, the platform protects data privacy as follows: In S6, the platform creates organizations and imports data into them. Organizations are independent of each other unless the creator allows data sharing between them. Furthermore, when an organization creator invites other members to join, members outside the organization cannot view the organization's datasets. This creates a data distribution center that can separate data, allowing datasets to be segmented based on the organization or project they are responsible for. Different organizations or projects cannot view their datasets, while also providing organizations and project leaders with the ability to join and propose members.

[0015] Furthermore, the link algorithm construction includes the following steps: S2-A, new algorithm; S2-B, select the output label and data source for the algorithm. If the algorithm is the first module in the cascade, the data source is the dataset in the organization; otherwise, it comes from the previous level algorithm; S2-C: Use the annotation tools provided by the platform to annotate. Once the organization has annotated and reviewed datasets for at least three, training can begin. S2-D, adjust the confidence and use the trained algorithm to filter the original data set for the next level algorithm; S2-E, repeat S2-A, execute the algorithm to mark the new line, then the result of the previous algorithm will be used as the input source of the new algorithm, and repeat SB and SD until the pathological tissue of the user's interest is detected or located in the original data set.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By creating a link algorithm, operators that affect inference speed and memory usage can be removed while maintaining their accuracy; the storage space of computing nodes can be managed to ensure the stability of computing node services; by creating a link algorithm, users can train data sets in a highly customized manner in combination with various downstream tasks to generate models that meet their needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below with reference to the accompanying drawings and examples.

[0018] Figure 1 Flowchart of the diagnostic method of digital pathological sections based on the link algorithm of the present invention; Figure 2 is an overview diagram of the improved RTMDet model of Example 1; Figure 3 is the CARAFE upsampling operator graph of the present invention; Figure 4 It is the fast module diagram of the expanded spatial pyramid pooling of the present invention; Figure 5 This is the HDA hybrid dilated attention module diagram of the present invention; Figure 6 It is the editing module diagram of the present invention; Figure 7 It is the toolbar diagram of the present invention; Figure 8 This is an example diagram of manually annotated polygons of the present invention; Figure 9 This is an example diagram of the automatic annotation of the present invention; Figure 10 This is a diagram illustrating frame selection and click selection of the present invention. DETAILED DESCRIPTION

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0020] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, features defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0021] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0022] like Figures 1 to 10 FIG. 1 is a preferred embodiment of the present invention. The diagnostic method of digital pathological sections based on the link algorithm of this embodiment includes the following steps: S1. Create tissue and upload digital pathology slide datasets; S2, create link algorithm; S3, using annotation tools to annotate the digital pathology slide data in S1; S4, review the digital pathology slide data that have been annotated in S3; S5, setting corresponding parameters for the link algorithm created by S2, and training the link algorithm using the digital pathology slide data reviewed and annotated by S4; S6. If the training effect of the link algorithm in S5 meets the requirements, repeat S2 as needed and cascade new modules. If the training effect of the link algorithm in S5 does not meet the requirements, repeat S3 and S4 to supplement the data and continue to enrich the data set until the requirements are met; S7. Once the entire chain algorithm is trained, the trained model is exported and used to infer digital pathology slides, locate and classify the corresponding tissues, and export their data reports. Thus, by creating a chain algorithm, operators that affect inference speed and memory usage are removed while maintaining accuracy; the storage space of computing nodes can be managed to ensure the stability of computing node services; and by creating a chain algorithm, users can train datasets in a highly customized manner in combination with various downstream tasks to generate models that meet their needs.

[0023] Specifically, in S3, the marked shapes include but are not limited to: circle, square, polygon, SAM, SAM2, etc.

[0024] In this embodiment, the link algorithm is an improved RTMDet model, which includes: Backbone consists of CSPNextBlock and Neck consists of CSPNextPAFPN; The S2 comprises the following steps: S2-1. When digital pathology slide data is used as the input of the target detection network, the size of the input digital pathology slide data is cropped or resized according to the memory of the hardware built by the RTMDet model. The RTMDet model network is improved. A higher-resolution shallow feature map is introduced in the C2 stage of Backbone. The DSPPE module is designed in the C5 stage to make it more suitable for capturing the receptive field of cells. The feature map is sent to the CSP module through a connection operation at the Neck for convolution. After the feature map is downsampled by 32 times, the HAD module is added to further enhance the features. S2-2. RTMDet model output and loss design: The output of the target detection model includes: target analogy and target location; The S2-1 includes the following steps: S2-1-1, CARAFE upsampling operator: In the Neck feature fusion network path, the plug-and-play CARAFE operator is used, which predicts the convolution kernel required for upsampling through convolution operation. , kernels at different positions can flexibly combine the semantic information of feature maps , thus obtaining more accurate upsampling results; S2-1-2, DSPPE expanded spatial pyramid pooling fast module: uses convolutional pooling with different expansion rates to improve SPPE and restricts the output to a square receptive field; S2-1-3, after Neck, the output feature maps are: the input downsampling size The HAD module is added after the feature map is downsampled to 32 times to enhance the ability of the RTMDet model to interpret contextual information. The model consists of a Convolution reduces the channel dimension and module parameters, and the output after dimensionality reduction is fed in parallel to three convolution branches and one residual attention branch. The expansion rates of the three parallel convolutions are , while the residual attention consists of spatial attention and channel attention; In S2-1-1, the convolution kernel The expression is: ; in: represents the kernel prediction module, which consists of a channel compression, a content encoder, and a kernel normalization layer; for l-centered neighborhood; Semantic information The expression is: ; in, is a content-aware reorganization module that converts Chinese Neighborhood and Kernel Reorganization, including = ; 'for The center position of the target , r= ; In S2-1-2, the maximum distance between two non-zero values in the feature map The expression is: ; in, is the receptive field radius of the i-th layer; In S2-1-3, the DSPPF module multi-scale feature fusion output The calculation formula is: ; in, represents a dilated convolution with a dilation rate of r, It is a normal convolution. represents the feature concatenation operation, and ECA is the efficient channel attention module; In S2-2, the dynamic soft label allocation policy function The expression is: ; in: 、 、 They are 1, 3, and 1 respectively; 、 、 Represent the classification loss function, regression loss function, and regional loss function respectively; Represents cross-entropy loss; P represents the classification probability predicted by the model; Represents the intersection-over-union ratio of the predicted box and the true box ( ), IoU = the intersection area of the predicted box and the real box / the union area of the predicted box and the real box, Represents the geometric alignment metric between the predicted box and the true box; is the center coordinate of the prediction box; is the center coordinate of the real box; for the region prior, the soft center prior is used Instead of a fixed center prior. Therefore, in S2-1, the purpose of cropping or resizing the input digital pathology slice data is because digital pathology slice scans usually have a higher resolution. In S2-1-1, the purpose of using the plug-and-play CARAFE operator is that in the path of the Neck feature fusion network, feature maps of different resolutions usually need to be upsampled to the same resolution before being connected. However, the upsampling commonly used in FPN (such as the nearest neighbor interpolation method, deconvolution, and sub-pixel convolution) has problems such as being affected by background interference, easy loss of key cell information, and introduction of invalid information. In S2-1-2, the purpose of using convolution pooling with different expansion rates to improve SPPE is that the receptive field of the output of ordinary convolution operations is fixed, the extracted feature area is also limited, and maximum pooling limits the spatial positioning information.

[0025] Specifically, in the first layer, a Ordinary convolution, then, applies three , the convolution with expansion rates of 1, 2, and 3 are used to extract feature maps of different receptive fields. Finally, the feature maps of four different receptive fields are spliced and reduced in dimension, and then The attention module provides output, and the input of the feature map in the DSPPE module is as follows: ; ; ; ; ; Among them: the subscript number indicates the number of layers, and They represent the core size and expansion rate of the roll respectively.

[0026] Specifically, ; .

[0027] In this embodiment, the platform resources are scheduled as follows: In S1, each GPU is assigned a unique URI and encapsulated in a resource pool; In S2, each user's request is added to a queue. The system simulates the user's input data and increases the batch size from 2 until the largest integer and even batch size is found, and then allocates GPU resources to it. In S3, when all GPU resources are occupied, the queue is locked and no new requests are accepted until GPU resources are released.

[0028] In this embodiment, S3 includes the following steps: S3-1. Image compression based on HEVC: When constructing an image pyramid, each layer of the image is considered as multiple When the camera takes a picture of the slide, each tile is put into the GPU and compressed using HEVC. Each time it is compressed, the pyramid level to which the tile belongs and the coordinates of the layer are recorded. When the entire image pyramid is built, each layer of the pyramid is also compressed accordingly, and the final output is a pathology image in .mpi format. S3-2. Image scaling and slicing during training and inference: S3-2-1. If the user slice data uses the .mpi format based on S3-1, use the corresponding decoder to first obtain the corresponding layer according to the mpp according to the level and coordinates recorded during compression, and then decode the corresponding tile according to the coordinate order; S3-2-2, if the user slice data format is not .mpi format, the image is converted into Cut into several tiles, and set the overlap area to 10%-20% to prevent feature points from being cut off; S3-3, adaptive screen resolution decoding method: for the digital pathology slices read, measure their length and width, and Divide it integerly and estimate how many tiles it can be divided into. Record the corresponding coordinates. Based on the center of the display screen and the ratio of the screen resolution to the original two-tile resolution, select the tile with the corresponding coordinates for decoding and presentation. Among them: In S3-2-1 and S3-2-2, during training, the tiles are scaled to the training size specified by the RTMDet model for training, and during inference, the tiles are merged. Thus, combined with HEVC, the storage space of digital pathology slices is compressed, reducing the storage pressure of the server; the image is scaled to Or cut into many The training and reasoning of images are accelerated by the adaptive screen resolution tile decoding method to speed up the browsing of images.

[0029] In this embodiment, S4 includes the following steps: S4-1. Model architecture: Improved based on Sam, including image encoder, hint encoder, and mask encoder; S4-2, loss function L: use the unweighted sum of cross-entropy loss and dice loss; The calculation formula of the loss function L is: ; Among them: In S4-1, the image encoder is composed of Vit-Base, and the prompt encoder encodes the point and box prompt data into sparse embedding and dense embedding. If the input is a point and a box, the output shape is , the mask encoder uses the embedding of the image encoder and the hint encoder to generate multiple segmentation masks using the Transformer architecture and simultaneously evaluate the quality of each mask; In S4-2, S and G are the segmentation results and the ground truth. 、 Voxel The predicted segmentation results and Ground truth, represents the cross entropy loss function; Denotes the Dice loss function. This can significantly improve the labeling efficiency.

[0030] In this embodiment, the platform protects data privacy as follows: In S6, the platform creates organizations and imports data into them. Organizations are independent of each other unless the organization creator allows data sharing between them. Furthermore, while the organization creator can invite other members to join, members outside the organization cannot view the organization's datasets. This creates a data distribution center that can separate data, allowing datasets to be segmented based on the organization or project they are responsible for. Different organizations or projects cannot view their datasets, while also providing organizations and project leaders with the ability to join and propose members.

[0031] In this embodiment, the link algorithm construction includes the following steps: S2-A, new algorithm; S2-B, select the output label and data source for the algorithm. If the algorithm is the first module in the cascade, the data source is the dataset in the organization; otherwise, it comes from the previous level algorithm; S2-C: Use the annotation tools provided by the platform to annotate. Once the organization has annotated and reviewed datasets for at least three, training can begin. S2-D, adjust the confidence and use the trained algorithm to filter the original data set for the next level algorithm; S2-E, repeat S2-A, execute the algorithm annotation update, then the result of the previous level algorithm execution will be used as the input source of the new algorithm, and repeat SB and SD until the pathological tissue of the user's concern is detected or located in the original data set.

[0032] In summary, the present invention creates a link algorithm to remove operators that affect inference speed and memory usage while maintaining its accuracy; it can manage the storage space of computing nodes to ensure the stability of computing node services; and by creating a link algorithm, users can train data sets in a highly customized manner in combination with various downstream tasks to generate models that meet their needs.

[0033] The above description is intended to serve as a guide for the preferred embodiments of the present invention. Based on the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of the present invention. The technical scope of the present invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A diagnostic method for digital pathological sections based on a link algorithm, characterized in that: The following steps are involved: S1. Create tissue and upload digital pathology slide datasets; S2, create link algorithm; S3, using annotation tools to annotate the digital pathology slide data in S1; S4, review the digital pathology slide data that have been annotated in S3; S5, setting corresponding parameters for the link algorithm created by S2, and training the link algorithm using the digital pathology slide data reviewed and annotated by S4; S6. If the training effect of the link algorithm in S5 meets the requirements, repeat S2 as needed and cascade new modules. If the training effect of the link algorithm in S5 does not meet the requirements, repeat S3 and S4 to supplement the data and continue to enrich the data set until the requirements are met; S7. When the algorithm training of the entire link is completed, the trained model is exported and used to infer the digital pathology slices, locate and classify the corresponding tissues, and export their data reports.

2. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The link algorithm is an improved RTMDet model, which includes: Backbone consists of CSPNextBlock and Neck consists of CSPNextPAFPN; The S2 comprises the following steps: S2-1. When digital pathology slide data is used as the input of the target detection network, the size of the input digital pathology slide data is cropped or resized according to the memory of the hardware built by the RTMDet model. The RTMDet model network is improved. A higher-resolution shallow feature map is introduced in the C2 stage of Backbone. The DSPPE module is designed in the C5 stage to make it more suitable for capturing the receptive field of cells. The feature map is sent to the CSP module through a connection operation at the Neck for convolution. After the feature map is downsampled by 32 times, the HAD module is added to further enhance the features. S2-2. RTMDet model output and loss design: The output of the target detection model includes: target analogy and target location.

3. The diagnostic method for digital pathological sections based on the link algorithm according to claim 2, characterized in that: The S2-1 includes the following steps: S2-1-1, CARAFE upsampling operator: In the Neck feature fusion network path, the plug-and-play CARAFE operator is used, which predicts the convolution kernel required for upsampling through convolution operation. , kernels at different positions can flexibly combine the semantic information of feature maps , thus obtaining more accurate upsampling results; S2-1-2, DSPPE expanded spatial pyramid pooling fast module: uses convolutional pooling with different expansion rates to improve SPPE and restricts the output to a square receptive field; S2-1-3, after Neck, the output feature maps are: the input downsampling size The HAD module is added after the feature map is downsampled to 32 times to enhance the ability of the RTMDet model to interpret contextual information. The model consists of a Convolution reduces the channel dimension and module parameters, and the output after dimensionality reduction is fed in parallel to three convolution branches and one residual attention branch. The expansion rates of the three parallel convolutions are , and the residual attention consists of spatial attention and channel attention.

4. The diagnostic method for digital pathological sections based on the link algorithm according to claim 3, characterized in that: In S2-1-1, the convolution kernel The expression is: ； in: represents the kernel prediction module, which consists of a channel compression, a content encoder, and a kernel normalization layer; for l-centered neighborhood; Semantic information The expression is: ； in, is a content-aware reorganization module that converts Chinese Neighborhood and Kernel Reorganization, including = ; 'for The center position of the target , r= ; In S2-1-2, the maximum distance between two non-zero values in the feature map The expression is: ； in, is the receptive field radius of the i-th layer; In S2-1-3, the DSPPF module multi-scale feature fusion output The calculation formula is: ； in, represents a dilated convolution with a dilation rate of r, is a normal convolution, represents the feature concatenation operation, and ECA is the efficient channel attention module.

5. The diagnostic method for digital pathological sections based on the link algorithm according to claim 2, characterized in that: In S2-2, the dynamic soft label allocation policy function The expression is: ； in: 、 、 They are 1, 3, and 1 respectively; 、 、 Represent the classification loss function, regression loss function, and regional loss function respectively; Represents cross-entropy loss; P represents the classification probability predicted by the model; Represents the intersection-over-union ratio of the predicted box and the true box ( ), IoU = the intersection area of the predicted box and the real box / the union area of the predicted box and the real box, Represents the geometric alignment metric between the predicted box and the true box; is the center coordinate of the prediction box; is the center coordinate of the real box; for the region prior, the soft center prior is used Instead of a fixed center prior.

6. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The platform resources are scheduled as follows: In S1, each GPU is assigned a unique URI and encapsulated in a resource pool; In S2, each user's request is added to a queue. The system simulates the user's input data and increases the batch size from 2 until the largest integer and even batch size is found, and then allocates GPU resources to it. In S3, when all GPU resources are occupied, the queue is locked and no new requests are accepted until GPU resources are released.

7. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The S3 includes the following steps: S3-1. Image compression based on HEVC: When constructing an image pyramid, each layer of the image is considered as multiple When the camera takes a picture of the slide, each tile is put into the GPU and compressed using HEVC. Each time it is compressed, the pyramid level to which the tile belongs and the coordinates of the layer are recorded. When the entire image pyramid is built, each layer of the pyramid is also compressed accordingly, and the final output is a pathology image in .mpi format. S3-2. Image scaling and slicing during training and inference: S3-2-1. If the user slice data uses the .mpi format based on S3-1, use the corresponding decoder to first obtain the corresponding layer according to the mpp according to the level and coordinates recorded during compression, and then decode the corresponding tile according to the coordinate order; S3-2-2, if the user slice data format is not .mpi format, the image is converted into Cut into several tiles, and set the overlap area to 10%-20% to prevent feature points from being cut off; S3-3, adaptive screen resolution decoding method: for the digital pathology slices read, measure their length and width, and Divide it integerly and estimate how many tiles it can be divided into. Record the corresponding coordinates. Based on the center of the display screen and the ratio of the screen resolution to the original two-tile resolution, select the tile with the corresponding coordinates for decoding and presentation. Among them: In S3-2-1 and S3-2-2, during training, the tiles are scaled to the training size specified by the RTMDet model for training, and during inference, the tiles are merged.

8. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The S4 comprises the following steps: S4-1. Model architecture: Improved based on Sam, including image encoder, hint encoder, and mask encoder; S4-2, loss function L: use the unweighted sum of cross-entropy loss and dice loss; The calculation formula of the loss function L is: ； Among them: In S4-1, the image encoder is composed of Vit-Base, and the prompt encoder encodes the point and box prompt data into sparse embedding and dense embedding. If the input is a point and a box, the output shape is , the mask encoder uses the embedding of the image encoder and the hint encoder to generate multiple segmentation masks using the Transformer architecture and simultaneously evaluate the quality of each mask; In S4-2, S and G are the segmentation results and the ground truth. 、 Voxel The predicted segmentation results and Groundtruth, represents the cross entropy loss function; represents the Dice loss function.

9. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The platform protects data privacy as follows: In S6, the platform imports data into the organization by creating the organization; Organizations are independent of each other unless the founder of the organization allows data to be shared between organizations; Among them: The organization creator invites other members to join the organization. Members who are not in the organization cannot view the datasets in the organization.

10. The diagnostic method for digital pathological sections based on the link algorithm according to claim 1, characterized in that: The link algorithm construction includes the following steps: S2-A, new algorithm; S2-B, select the output label and data source for the algorithm. If the algorithm is the first module in the cascade, the data source is the dataset in the organization; otherwise, it comes from the previous level algorithm; S2-C: Use the annotation tools provided by the platform to annotate. Once the organization has annotated and reviewed datasets for at least three, training can begin. S2-D, adjust the confidence and use the trained algorithm to filter the original data set for the next level algorithm; S2-E, repeat S2-A, execute the algorithm annotation update, then the result of the previous level algorithm execution will be used as the input source of the new algorithm, and repeat SB and SD until the pathological tissue of the user's concern is detected or located in the original data set.

Images