Pathological section analysis method, electronic device and readable storage medium based on deep learning
Through deep learning technology, combined with Bill-Lambert conversion method and DPU-Net network, the problem of nucleus segmentation and classification in pathological sections is solved, the accurate segmentation of nucleus and classification of pathological sections is achieved, and the diagnostic efficiency is improved.
Patent Information
- Application Number
- CN202110955682.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-19
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-08-19
AI Technical Summary
The prior art is difficult to accurately segment the nucleus in pathological sections, especially when the nucleus overlaps, contacts and impurities exist, affecting diagnostic efficiency.
Using deep learning-based pathological slice analysis method, color normalization is performed through Bill-Lambert conversion method and sparse matrix decomposition method, feature images are extracted using the encoder of the DPU-Net network, combined with nuclear segmentation and classification branching, and using distance The graph branch processed overlapping nuclei, and extracted the characteristics of the nuclear aggregation region through the DBSCAN clustering algorithm, and finally classified using a width learning model.
The accurate segmentation of the cell nucleus and the accurate classification of pathological sections are achieved, which reduces the pressure on pathologists to diagnose and improves diagnostic efficiency.
Smart Images

Figure CN113628201B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision technology and relates to a pathological section analysis method based on deep learning, an electronic device and a readable storage medium. Background Art
[0002] In recent years, with the rapid development of scanning technology, especially the emergence of full-field digital slices, the development of automatic analysis of histopathological images in the computer field has been greatly promoted. At present, artificial intelligence technology represented by deep learning has become one of the hottest research fields, and combining it with histopathological image analysis has attracted the attention of countless scientific researchers.
[0003] Automated analysis of cell nucleus information, including segmentation and classification, is of great significance for the automated diagnosis of pathological sections. Among them, the task of cell nucleus segmentation occupies an important position in the entire automated analysis of tissue pathology images. At the same time, cell nucleus segmentation is also a difficult point. Due to the limitations of the production process of pathological sections, cell nucleus overlap and contact cannot be avoided, and impurities around the cell nucleus and cytoplasmic areas will affect the cell nucleus segmentation. If the cell nucleus can be accurately segmented and its contour information can be obtained on this basis, it will be able to effectively help pathologists extract cell nucleus information and make effective diagnoses; cell nucleus classification and segmentation are only the extraction of cell nucleus information on pathological sections. If this information can be used for further analysis, the classification information of the entire pathological section can be obtained, which greatly reduces the diagnostic pressure of pathologists. Summary of the invention
[0004] In order to achieve the above objectives, an embodiment of the present invention provides a pathological section analysis method based on deep learning, which can accurately segment the cell nuclei in the pathological sections, and realize the accurate classification of the pathological sections by combining the pixel classification results and the characteristics of the cell nucleus aggregation area.
[0005] The embodiments of the present invention also provide an electronic device and a readable storage medium.
[0006] The technical solution adopted by the present invention is a pathological section analysis method based on deep learning, which specifically includes the following steps:
[0007] S1, color normalization of RGB pathological sections using Beer-Lambert transformation and sparse matrix decomposition;
[0008] S2, using the encoder of the DPU-Net network to extract the feature image in the slice after color normalization, and based on the feature image, perform cell nucleus segmentation and cell nucleus classification on the pathological slice, and preliminarily divide the pathological slice into normal slices and suspected abnormal slices;
[0009] S3, dividing the suspected abnormal sections into cell nucleus aggregation areas and extracting the features of the cell nucleus aggregation areas;
[0010] S4, using the cell nucleus clustering area features as input, the width learning model is used to perform the final classification of suspected abnormal slices.
[0011] Furthermore, the DPU-Net network includes an encoder and a decoder;
[0012] The encoder includes a feature extraction layer and four dual-stream connection modules connected in sequence, the feature extraction layer includes a convolution layer, a normalization layer and an activation layer connected in sequence, the dual-stream connection module includes three groups of normalization layers, activation layers and convolution layers connected in sequence, and the processing process of the dual-stream connection module is as follows:
[0013] y=PC([x[bw:],F(x)[bw:],x[:bw]+F(x)[:bw]])
[0014] Where PC(·) represents the mean pooling operation and convolution operation, y represents the output result, x represents the input feature, F(x) represents the intermediate variable, + represents the superposition operation in residual learning, and [·] represents dense connection;
[0015] The decoder includes four UP Block modules and a Softmax layer corresponding to the dual-stream connection module, and the UP Block module includes an upsampling layer and two groups of convolution layers, a normalization layer and an activation layer connected in sequence.
[0016] Furthermore, the preliminary classification process of the pathological sections in S2 is as follows:
[0017] S2-1, performing cell nucleus segmentation on the feature image, and preliminarily dividing it into the cell nucleus inner region, the cell nucleus outer region and the cell nucleus boundary region;
[0018] S2-2, calculating the distance value from each pixel point in the inner region of the cell nucleus to the outer region of the cell nucleus, a distance value of 1 indicates that the pixel point is in the center region of the cell nucleus, a distance value of 0 indicates that the pixel point is in the boundary region of the cell nucleus, and a distance value of (0, 1) indicates that the pixel point is in the inner region of the cell nucleus. The pixel point with a distance value of 0 is used as a segmentation line to determine the boundary region of the overlapping cell nuclei, and the cell nucleus segmentation is completed;
[0019] S2-3, classifying each pixel in the feature image into pixels outside the cell nucleus, normal pixels, and abnormal pixels;
[0020] S2-4, counting the number of normal pixels and the number of abnormal pixels in the boundary area of each cell nucleus in the cell nucleus segmentation result, if the number of normal pixels is greater than the number of abnormal pixels, the cell nucleus is considered to be a normal cell nucleus, otherwise it is considered to be an abnormal cell nucleus;
[0021] S2-5, count the number of abnormal cell nuclei in the pathological section. If the number of abnormal cell nuclei is less than 5% of the total number of cell nuclei, the section is considered to be a normal section, otherwise it is a suspected abnormal section.
[0022] Furthermore, the S3 uses the DBSCAN clustering algorithm to divide the suspected abnormal slices into cell nucleus aggregation areas, and uses mathematical analysis to statistically analyze the characteristics of the cell nucleus aggregation areas, wherein the characteristics include the average cell nucleus area, the average cell nucleus density, the area ratio of the area with a cell nucleus density greater than 50%, the number of cell nucleus aggregation areas and the regional characteristics of the number of cell nuclei.
[0023] Furthermore, the classification process of S4 is as follows:
[0024] S4a, taking the cell nucleus clustering area features in the form of text data as input to construct feature nodes, and connecting the feature nodes to generate a feature node set Z N , Z N =[Z1,Z2,…,Z i ,…,Z N ], for Z N Merge to get the feature matrix;
[0025] The feature node Where X represents the characteristics of the cell nucleus aggregation area, φ represents the function of constructing feature nodes, Respectively represent the offset matrix and offset vector of the i-th group of feature nodes, Z i represents the i-th group of feature nodes, and N represents the total number of feature node groups;
[0026] S4b, feature enhancement is performed on the feature nodes to obtain enhanced nodes, the enhanced nodes are divided into M modules, each module includes m groups of enhanced nodes, and the last group of enhanced nodes of each module is extracted to form an enhanced matrix;
[0027] The calculation formula of the enhanced node is as follows:
[0028]
[0029] Among them, H jk represents the kth group of enhanced nodes in the jth module, j = 1, ..., M, k = 1, ..., m, Indicates the activation function of the enhanced node. Respectively represent the randomly generated H jkThe offset matrix and offset vector, H 1m represents the mth group of enhanced nodes in the first module, H (j-1)m represents the mth group of enhanced nodes in the j-1th module, H j(k-1) represents the k-1th group of enhanced nodes in the jth module;
[0030] S4c, calculate the final classification label of the suspected non-normal slice based on the feature matrix, enhancement matrix and weight matrix.
[0031] An electronic device comprises a processor, a memory and a communication bus, wherein the processor and the memory communicate with each other via the communication bus;
[0032] Memory, used to store computer programs;
[0033] The processor is used to implement the above method steps when executing the program stored in the memory.
[0034] A computer-readable storage medium stores a computer program, and the computer program implements the above method steps when executed by a processor.
[0035] The beneficial effects of the present invention are:
[0036] 1. The embodiment of the present invention uses a dual-stream connection module to extract features from feature images, realizes the reuse of shallow features and deep features, uses dense connection pathways to explore new features, and makes the obtained features more representative of the information of pathological sections; at the same time, the cell nucleus segmentation branch and the distance map branch are used to jointly segment the cell nucleus, so as to obtain accurate contour information of single cell nuclei and overlapping and contacting cell nuclei, and the preliminary classification results of the pathological sections obtained by combining the classification results of each pixel point are more accurate;
[0037] 2. Based on the preliminary classification results of pathological sections, the embodiment of the present invention extracts the features of the cell nucleus aggregation area and uses the width learning system model to accurately classify it. During the entire classification process, the embodiment of the present invention combines the detailed information in the pathological sections and the macro information of the cell nucleus aggregation area, and the obtained classification results are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0039] Figure 1is a flow chart of an embodiment of the present invention.
[0040] Figure 2 This is the model diagram of the DPU-Net network.
[0041] Figure 3 This is the processing diagram of the dual-stream connection module.
[0042] Figure 4 This is the structural diagram of the DPU-Net network decoder.
[0043] Figure 5 It is a network model diagram of the width learning system.
[0044] Figure 6 This is a comparison chart before and after color normalization. DETAILED DESCRIPTION
[0045] 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.
[0046] Example
[0047] like Figure 1 As shown, the pathological section analysis method based on deep learning specifically includes the following steps:
[0048] Step 1: Preprocess the pathological slices in RGB format and use the encoder in the DPU-Net network to extract features from the image. The specific process includes:
[0049] Step 1a, using Beer-Lambert transformation method and sparse matrix decomposition method, performing color normalization operation on RGB format pathological slice with width and height of w×h to obtain a first image;
[0050] Because different operators use different amounts of reagents when HE staining pathological sections, the colors presented are very different. When performing deep learning, color information is a critical feature. In order to avoid unnecessary errors caused by differences in color information, color normalization is required to normalize the color space of all sections. When performing color normalization, the RGB matrix of the pathological section is first converted into an optical density matrix through the Beer-Lambert transformation method. The specific conversion formula is as follows:
[0051]
[0052] Then, the optical density matrix is decomposed by sparse non-negative matrix to obtain two matrices, one of which is the RGB optical density component matrix containing each dyeing prototype, and the other is the dyeing density matrix. Finally, the dyeing density matrix of the image to be normalized and the dyeing prototype matrix of the target image are multiplied, and then the Beer-Lambert inverse transformation is performed to obtain the normalized first image. The result of color normalization is shown as follows Figure 6 As shown;
[0053] Step 1b, inputting the first image into the encoder of the DPU-Net network to obtain the corresponding feature image, wherein the encoder is connected in sequence by a feature extraction layer and four dual-stream connection modules, which are respectively used to extract image features from shallow to deep, and finally obtain a feature image covering shallow features and deep features of the image;
[0054] Step 2, input the feature image into the cell nucleus segmentation module in the DPU-Net network decoder for decoding to obtain the cell nucleus segmentation result;
[0055] The cell nucleus segmentation module in the DPU-Net network decoder includes two branches: the cell nucleus segmentation branch and the distance map branch. The feature map is input into the cell nucleus segmentation branch to obtain the pixel-level semantic segmentation result, that is, the feature image is preliminarily divided into the cell nucleus internal area, the cell nucleus external area and the cell nucleus boundary area;
[0056] The feature image is input into the distance map branch, and the distance value from each pixel in the inner area of the cell nucleus to the outer area of the cell nucleus is calculated. The final cell nucleus segmentation result is obtained based on the cell nucleus semantic segmentation result and the distance map.
[0057] The decoder uses the Sigmod function as the activation function, and the function formula is as follows: The output range of the function is between 0 and 1. In the cell nucleus segmentation branch, the output of the sigmoid function represents the probability that the pixel belongs to a certain category. The pixel is finally divided into the category corresponding to the maximum probability. In the distance map branch, its output represents the distance value from each pixel in the internal area of the cell nucleus to the external area of the cell nucleus. Since the distance from the central area of the cell nucleus to the external area is the longest, and the distance from the boundary of the cell nucleus and the external area of the cell nucleus to the external area is the shortest, when the output result of the distance map branch is 1, it represents the central area of the cell nucleus, when the output result is 0, it represents the boundary or external area of the cell nucleus, and when the output result is between 0 and 1, it represents the internal area of the cell nucleus.
[0058] The cell nucleus segmentation branch can divide the feature image into the cell nucleus internal area, the cell nucleus external area and the cell nucleus boundary area. However, if two or more cell nuclei overlap, the cell nucleus segmentation branch cannot effectively determine the boundary area of each cell nucleus at the overlap, and can only identify the overlap as the cell nucleus internal area. Therefore, the distance map branch is needed to segment the overlapping cell nuclei and calculate the distance value from each pixel point in the cell nucleus internal area to the cell nucleus external area. When there are multiple cell nucleus center areas in the overlapping area, multiple maximum value areas will be obtained. There are multiple segmentation lines composed of zero value points between the multiple maximum value areas, which are the boundary areas of multiple overlapping cell nuclei, so as to achieve accurate segmentation of overlapping cell nuclei.
[0059] Step 3: Input the feature image into the cell nucleus classification module in the DPU-Net network decoder for decoding, classify each pixel of the feature image, and finally divide it into pixels outside the cell nucleus, normal pixels, and abnormal pixels;
[0060] Step 4, counting the number of normal pixels and abnormal pixels in the boundary area of each cell nucleus in the cell nucleus segmentation result. If the number of normal pixels is greater than the number of abnormal pixels, the cell nucleus is considered to be a normal cell nucleus, otherwise it is considered to be an abnormal cell nucleus;
[0061] Step 5, counting the number of abnormal cell nuclei in the pathological section, if the number of abnormal cells is less than 5% of the total number of cell nuclei, the pathological section is considered to be a normal section, otherwise it is considered to be a suspected abnormal section, and the suspected abnormal section is divided into a cell nucleus aggregation area;
[0062] The domain radius was set to 35 mm, and the threshold of the number of cell nuclei in the domain was set to 5. The center point of each cell nucleus was calculated based on the cell nucleus contour information in the cell nucleus segmentation result. Based on this, the DBSCAN clustering algorithm was used to divide the pathological sections to obtain the cell nucleus aggregation area. The pixel points in the cell nucleus aggregation area were mathematically analyzed to obtain the characteristics of the cell nucleus aggregation area, including the average cell nucleus area, the average cell nucleus density, the area ratio of the area with a cell nucleus density greater than 50%, the number of cell nucleus aggregation areas, and the regional characteristics of the number of cell nuclei.
[0063] In step 6, the cell nucleus clustering area features of the suspected abnormal slices are used as input, and the suspected abnormal slices are finally classified using the dense cascade-based width learning system model to divide the suspected abnormal slices into normal slices and abnormal slices.
[0064] The specific classification process of step 6 is as follows:
[0065] Step 6a, taking the features of the cell nucleus aggregation area in the form of text data as input to construct feature nodes, generate a feature node set, and then merge them to obtain a feature matrix;
[0066] Step 6b, perform feature enhancement on the feature matrix to obtain enhanced nodes, divide the enhanced nodes into M modules, each module contains m groups of enhanced nodes, connect the modules using a dense connection method, and extract the last group of enhanced nodes in each module as the enhanced matrix;
[0067] Step 6c, calculate the final classification label of the pathological slice according to the feature matrix, the enhancement matrix and the weight matrix W obtained in the previous training, and complete the classification of the suspected abnormal slices.
[0068] The existing technology starts more from the network model structure and improves the segmentation effect of the feature image through a complex model structure. The DPU-Net network described in the embodiment of the present invention performs detailed segmentation of overlapping cell nuclei by adding distance graph branches, avoiding the influence of cell nucleus overlap and contact as well as the cytoplasm region and impurities around the cell nucleus on the cell nucleus segmentation. It can accurately segment the cell nucleus and obtain its contour information, effectively helping pathologists to extract cell nucleus information.
[0069] The DPU-Net network is based on the U-Net network and uses a dual-stream connection module to replace the block in the original U-Net network. The original block is mainly used for image feature extraction. It is easy to forget the shallow features extracted earlier during the extraction process, and the search ability for new features is weak; the dual-stream connection module inherits the advantages of the two major networks of ResNet and DenseNet, combines shallow features with deep features through the residual branch path to achieve feature reuse, and explores new features through the dense connection path of the densely connected network, effectively improving the feature extraction capability, and the obtained image features are more comprehensive.
[0070] The overall structure of the DPU-Net network is as follows Figure 2 As shown in FIG. 1 , it includes an encoder for feature extraction and a decoder for realizing functions. The encoder includes a feature extraction layer and four dual-stream connection modules (DPN Block) connected in sequence. The feature extraction layer is used to perform preliminary feature extraction on the input data, and the dual-stream connection module is used to reuse the features extracted by the previous layer. The processing process of the dual-stream connection module is as follows: Figure 3As shown in the figure, the input feature x is first subjected to three consecutive normalization (BN), activation (activation function is ReLu), and convolution (Conv) operations to obtain the intermediate variable F(x), then x[bw:] and F(x)[bw:] are subjected to residual learning (i.e., superposition operation), and the residual learning result is densely connected with x[:bw] and F(x)[:bw], and finally the dense connection result is subjected to mean pooling operation (Avg Pool) and convolution operation to obtain the output result y. PC(·) is used to represent the mean pooling operation and convolution operation. The processing process of the dual-stream connection module in the encoder is as follows:
[0071] y=PC([x[bw:],F(x)[bw:],x[:bw]+F(x)[:bw]])
[0072] Figure 2 The C in the middle stands for Connected, which means fusing the shallow features obtained by the encoder with the deep features. This can effectively avoid the overwriting of shallow information acquired earlier due to the increase in the number of layers in the deep learning system, thereby improving the processing effect of the neural network.
[0073] The decoder part includes a cell nucleus segmentation module and a cell nucleus classification module. The cell nucleus segmentation module includes a cell nucleus segmentation branch and a distance map branch. The cell nucleus segmentation branch, the distance map branch and the cell nucleus classification module have similar structures. Figure 4 As shown, there is an Up Block module in the decoder that corresponds to the dual-stream connection module of the encoder. The UP Block module includes an upsampling layer and two groups of convolution layers, a normalization layer and an activation layer connected in sequence. It performs decoding operations through upsampling, convolution operations and normalization functions, and then activates through the sigmoid function. The sigmoid function can map the value of the pixel point to a value of [0, 1], and the sum of these values is 1. The output of the sigmoid function in the cell nucleus segmentation branch and the cell nucleus classification branch represents the probability that the pixel point belongs to a certain category. At the final output, the pixel point is divided into the category corresponding to the maximum probability, completing the segmentation and classification of the pixel points in the feature image. The output of the sigmoid function in the distance map branch represents the distance value from each pixel point in the internal area of the cell nucleus to the external area of the cell nucleus.
[0074] The network structure of the densely connected width learning system is as follows Figure 5 As shown in the figure, the width learning system model first extracts the features of the input data to obtain feature nodes. Where X represents the input data, that is, the features extracted from the cell nucleus aggregation area, φ represents the function of constructing feature nodes, that is, the sparse autoencoder function based on Lasso regression, Respectively represent the offset matrix and offset vector of the i-th group of feature nodes, Z irepresents the i-th group of feature nodes, N represents the total number of feature node groups, and N groups of feature nodes are connected to obtain the final feature node set Z N , Z N =[Z1,Z2,…,Z i ,…,Z N ].
[0075] The feature nodes are enhanced to obtain the corresponding enhanced nodes. The enhanced nodes are divided into M modules. Each module contains m groups of enhanced nodes. The modules are connected in a dense connection manner. For each group of enhanced nodes in each module, the input data includes the feature node set Z N , the last set of enhanced nodes of all previous modules. If this module is the first module, its input data is only the feature node set Z N , the m groups of enhanced nodes in each module are connected in a fully connected manner. The formula for obtaining enhanced nodes is as follows:
[0076]
[0077] H jk represents the kth group of enhanced nodes in the jth module, j = 1, ..., M, k = 1, ..., m, Indicates the activation function of the enhanced node. Respectively represent the randomly generated H jk The offset matrix and offset vector, H 1m represents the mth group of enhanced nodes in the first module, H (j-1)m represents the mth group of enhanced nodes in the j-1th module, H j(k-1) represents the k-1th group of enhanced nodes in the jth module.
[0078] The feature node set Z N Merge into feature matrix F. In order to avoid overfitting, extract the mth group of enhanced nodes and combined enhanced node set H of each module. M , H M =[H 1m ,H 2m ,…,H jm ,…,H Mm ], the enhanced node set H M Merge into the enhanced matrix E′, connect the feature matrix F and the enhanced matrix E′ according to the column vector to obtain the matrix D, and the final label data Y=DW, where W is the weight matrix obtained through model training.
[0079] The embodiment of the present invention first uses a DPU-Net network to divide pathological slices into normal slices and suspected abnormal slices. The DPU-Net network uses a dual-stream connection module to extract features, which can combine shallow features and deep features to achieve feature reuse, and explore new features through dense connection pathways to make the obtained features more comprehensive; at the same time, three branches are constructed in the decoder part: a cell nucleus segmentation branch, a distance map branch, and a cell nucleus classification branch. The cell nucleus segmentation branch is used to roughly segment the cell nuclei in the pathological slices, and the distance map branch is used to finely segment the overlapping areas in the cell nucleus segmentation map to obtain more accurate cell nucleus segmentation results, and then the pixel-level classification results are combined to achieve preliminary classification of the pathological slices, and finally, the statistical characteristics of the cell nucleus aggregation area in the pathological slices are combined, and the densely connected width learning system is used to perform the final classification.
[0080] The cell nucleus segmentation result obtained by the embodiment of the present invention is more accurate, avoiding the influence of cell nucleus contact, overlap or cytoplasm and impurities on the cell nucleus segmentation result. On this basis, the pixel-level classification results and the macroscopic characteristics of the cell nucleus aggregation area are combined to perform step-by-step classification. The obtained classification is more accurate and can effectively help pathologists extract cell nucleus information.
[0081] The present invention also includes an electronic device, including a memory and a processor, wherein the memory is used to store various computer program instructions, and the processor is used to execute the computer program instructions to complete all or part of the above steps; the electronic device can communicate with one or more external devices, and can also communicate with one or more devices that enable users to interact with the electronic device, and / or communicate with any device that enables the electronic device to communicate with one or more other computing devices. The electronic device can also communicate with one or more networks (such as local area networks, wide area networks and / or public networks) through a network adapter.
[0082] The present invention also includes a computer-readable medium storing a computer program that can be executed by a processor. The computer-readable medium may include but is not limited to magnetic storage devices, optical disks, digital versatile disks, smart cards and flash memory devices. In addition, the readable storage medium described in the present invention can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" includes but is not limited to wireless channels and various other media (and / or storage media) that can store, contain and / or carry codes and / or instructions and / or data.
[0083] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.
Claims
1. A pathological section analysis method based on deep learning, characterized in that: The specific steps include: S1, color normalization of RGB pathological sections using Beer-Lambert transformation and sparse matrix decomposition; S2, using the encoder of the DPU-Net network to extract the feature image in the slice after color normalization, and based on the feature image, perform cell nucleus segmentation and cell nucleus classification on the pathological slice, and preliminarily divide the pathological slice into normal slices and suspected abnormal slices; S3, dividing the suspected abnormal sections into cell nucleus aggregation areas and extracting the features of the cell nucleus aggregation areas; S4, using the cell nucleus clustering area features as input, the width learning model is used to perform the final classification of suspected abnormal slices; The DPU-Net network includes an encoder and a decoder; The encoder includes a feature extraction layer and four dual-stream connection modules connected in sequence, the feature extraction layer includes a convolution layer, a normalization layer and an activation layer connected in sequence, the dual-stream connection module includes three groups of normalization layers, activation layers and convolution layers connected in sequence, and the processing process of the dual-stream connection module is as follows: ; in Represents mean pooling operation and convolution operation, Indicates the output result, represents the input features, represents the intermediate variable, represents the superposition operation in residual learning, Indicates dense connection; The decoder includes four UP Block modules and a Softmax layer corresponding to the dual-stream connection module, and the UP Block module includes an upsampling layer and two groups of convolution layers, a normalization layer, and an activation layer connected in sequence; The S3 uses the DBSCAN clustering algorithm to divide suspected abnormal sections into cell nucleus aggregation areas, and uses mathematical analysis to count the characteristics of the cell nucleus aggregation areas, the characteristics including the average cell nucleus area, the average cell nucleus density, the area ratio of the area with a cell nucleus density greater than 50%, the number of cell nucleus aggregation areas and the regional characteristics of the number of cell nuclei.
2. The pathological section analysis method based on deep learning according to claim 1, characterized in that: The preliminary classification process of pathological sections in S2 is as follows: S2-1, performing cell nucleus segmentation on the feature image, and preliminarily dividing it into the cell nucleus inner region, the cell nucleus outer region and the cell nucleus boundary region; S2-2, calculating the distance value from each pixel point in the inner region of the cell nucleus to the outer region of the cell nucleus, a distance value of 1 indicates that the pixel point is in the center region of the cell nucleus, a distance value of 0 indicates that the pixel point is in the boundary region of the cell nucleus, and a distance value of (0, 1) indicates that the pixel point is in the inner region of the cell nucleus. The pixel point with a distance value of 0 is used as a segmentation line to determine the boundary region of the overlapping cell nuclei, and the cell nucleus segmentation is completed; S2-3, classifying each pixel in the feature image into pixels outside the cell nucleus, normal pixels, and abnormal pixels; S2-4, counting the number of normal pixels and the number of abnormal pixels in the boundary area of each cell nucleus in the cell nucleus segmentation result, if the number of normal pixels is greater than the number of abnormal pixels, the cell nucleus is considered to be a normal cell nucleus, otherwise it is considered to be an abnormal cell nucleus; S2-5, count the number of abnormal cell nuclei in the pathological section. If the number of abnormal cell nuclei is less than 5% of the total number of cell nuclei, the pathological section is considered to be a normal section, otherwise it is a suspected abnormal section.
3. The pathological section analysis method based on deep learning according to claim 1, characterized in that: The classification process of S4 is as follows: S4a, taking the cell nucleus aggregation area features in the form of text data as input to construct feature nodes, and connecting the feature nodes to generate a feature node set ,right Merge to get the feature matrix; The feature node ,in Indicates the characteristics of the nuclear clustering area, Represents the function of constructing feature nodes, Respectively represent The offset matrix and offset vector of the group feature node, Indicates Group feature nodes, Indicates the total number of feature node groups; S4b, feature enhancement is performed on the feature nodes to obtain enhanced nodes, the enhanced nodes are divided into M modules, each module includes m groups of enhanced nodes, and the last group of enhanced nodes of each module is extracted to form an enhanced matrix; The calculation formula of the enhanced node is as follows: ; in Indicates of the modules Group Enhancement Node, , Indicates the activation function of the enhanced node. Respectively represent random generation The offset matrix and offset vector, Indicates the first module Group Enhancement Node, Indicates of the modules Group Enhancement Node, Indicates of the modules Group Enhancement Node; S4c, calculate the final classification label of the suspected non-normal slice based on the feature matrix, enhancement matrix and weight matrix.
4. An electronic device, characterized in that: It includes a processor, a memory and a communication bus, wherein the processor and the memory communicate with each other via the communication bus; Memory, used to store computer programs; A processor, for implementing the method steps described in any one of claims 1 to 3 when executing a program stored in a memory.
5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 3 are implemented.
