Heterogeneous cell detection method and device based on deep learning
By constructing a heterogeneous cell detection model, using cell-level annotation and regional learning networks to optimize the loss function, the problems of low accuracy and high annotation cost in the existing technology are solved, and efficient heterogeneous cell detection and pathological diagnosis are achieved.
Patent Information
- Application Number
- CN202310121657.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-15
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2043-02-15
AI Technical Summary
When detecting heterogeneous cells, existing deep learning methods are difficult to effectively obtain the global regional distribution information of cells in pathological images, resulting in low detection accuracy and high regional labeling cost, which affects the accuracy of inflammatory cell detection.
By constructing a heterogeneous cell detection model, using cell-level annotation and cell position likelihood maps, combining the regional learning network and region guidance module, the implicit regional information is extracted and loss function optimization is performed to achieve accurate detection of heterogeneous cells.
It improves the accuracy of heterogeneous cell detection, reduces the labeling cost, improves the efficiency of pathological diagnosis, and promotes the further development of pathological research.
Smart Images
Figure CN116309354B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of cell detection, and specifically, to a method and device for detecting heterogeneous cells based on deep learning. Background Art
[0002] In the medical diagnosis process, histopathological diagnosis, which relies on morphology, is the primary means of pathology and is often considered the "gold standard" for identifying lesions and guiding clinical treatment selection. However, pathological diagnosis requires pathologists to identify subtle pathological changes in highly complex tissue images. As clinical demand increases, the workload and intensity of pathologists are also increasing. In recent years, deep learning technology has gradually matured and is increasingly used in many medical scenarios. Combining deep learning with pathology can assist pathologists in their diagnosis and improve their diagnostic efficiency. This is also a new development direction for the integration of pathology diagnosis and new technologies.
[0003] Cell analysis is a fundamental and crucial task in pathological diagnosis. Cancer cells, tumor cells, and various inflammatory cells are closely associated with the diagnosis of several diseases. Cell analysis is commonly used in disease screening and pathological typing. Automating the detection of tissue pathological cells using deep learning methods can significantly reduce the burden on pathologists.
[0004] In recent years, with the rapid development of deep learning in medical imaging, convolutional neural networks have been applied to various cell detection tasks in tissue pathology images. However, the distribution of cells in pathology images is heterogeneous, that is, different types of cells have similar characteristics but are distributed in different regions. Taking endometrial tissue pathology images as an example, they usually contain stromal and epithelial regions. The stromal region contains four types of inflammatory cells closely related to endometritis and reproductive problems, including lymphocytes, plasma cells, eosinophils, and neutrophils. However, there are some cells in the epithelial region that have similar characteristics to the inflammatory cells in the stroma. As a result, deep learning methods that analyze cells in each region equally often misclassify epithelial cells as inflammatory cells, thereby affecting the accuracy of inflammatory cell detection and adversely affecting subsequent diagnosis.
[0005] To accurately detect cells with heterogeneous distributions, existing cell detection methods often use region-level annotation as a guide, constructing a two-stage model that performs cell detection and region segmentation separately, thus incorporating region-level information into detection. However, due to the variable shape of regions and unclear boundary representation, the cost of region-level annotation is relatively high, making the application of such methods relatively complex. Furthermore, many have proposed using attention mechanisms to capture the spatial distribution information of cells. However, existing attention mechanisms are mostly based on convolution, which can only capture local relationships. Furthermore, the non-local block used to capture global information simply uses learned deterministic relationships (i.e., weighted sums) as weights to aggregate features. This approach has weak global information mining capabilities and lacks sufficient adaptability. Therefore, these methods are unable to effectively obtain global regional distribution information of cells in pathological images, and are therefore unable to effectively detect cells with heterogeneous distributions. Summary of the Invention
[0006] In order to overcome at least one deficiency in the prior art, the embodiments of the present application provide a method and device for detecting heterogeneous cells based on deep learning.
[0007] In a first aspect, a method for constructing a heterogeneous cell detection model based on deep learning is provided, comprising:
[0008] Obtaining model building data, which includes original images of various types of cells with heterogeneous distribution, cell position likelihood maps, and cell-level annotations;
[0009] The heterogeneous cell detection model is trained based on the original image to obtain a trained heterogeneous cell detection model; the heterogeneous cell detection model includes a cell detection network, a region learning network, and a region guidance module; the cell detection network is used to obtain cell features based on the original image; the region learning network is used to obtain target region prediction results based on the original image; and the region guidance module is used to obtain cell prediction results based on the cell features and the target region prediction results.
[0010] The loss function of the heterogeneous cell detection model is determined based on the cell position likelihood map and cell-level annotations.
[0011] In one embodiment, determining a loss function of a heterogeneous cell detection model based on a cell position likelihood map and cell-level annotations includes:
[0012] Input the cell-level annotation into the region acquisition module to obtain the region mask;
[0013] Obtaining a regional loss function according to the regional mask and the target region prediction result;
[0014] The detection loss function is obtained based on the cell position likelihood map and the cell prediction results;
[0015] According to the regional loss function and the detection loss function, the loss function of the heterogeneous cell detection model is determined.
[0016] In one embodiment, the cell-level annotation is input into the region acquisition module to obtain the region mask, including:
[0017] The location points of each cell in the cell-level annotation constitute a point set S;
[0018] According to the point set S and the loss function parameter α, the Alpha-shape algorithm is used to process and obtain the boundary;
[0019] Fill the boundary internally to obtain the region mask.
[0020] In one embodiment, a regional loss function is obtained based on the region mask and the target region prediction result, including:
[0021]
[0022] Among them, L region is the regional loss function, Ω r is the pixel set of the region mask, r j is the jth pixel in the pixel set, x is the input image, f(x) is the target region prediction result, and r is the region mask.
[0023] In one embodiment, a detection loss function is obtained based on the cell position likelihood map and the cell prediction result, including:
[0024] L det =-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p)
[0025] Among them, L det is the detection loss function, α and γ are loss function parameters, p is the cell prediction result, and y is the cell position likelihood map.
[0026] In one embodiment, determining the loss function of the heterogeneous cell detection model based on the region loss function and the detection loss function includes:
[0027] L total =L det +λL region
[0028] Among them, L total is the loss function of the heterogeneous cell detection model, L det is the detection loss function, Lregion is the regional loss function, and λ is the coefficient of the regional loss function.
[0029] In one embodiment, the cell detection network includes a shared encoder and a first decoder, and the region learning network includes a shared encoder and a second decoder.
[0030] In a second aspect, a device for constructing a heterogeneous cell detection model based on deep learning is provided, comprising:
[0031] A model building data acquisition unit, used to acquire model building data, the model building data including original images of various types of cells with heterogeneous distribution, cell position likelihood maps and cell-level annotations;
[0032] A model training unit is used to train a heterogeneous cell detection model based on the original image to obtain a trained heterogeneous cell detection model; the heterogeneous cell detection model includes a cell detection network, a region learning network, and a region guidance module; the cell detection network is used to obtain cell features based on the original image; the region learning network is used to obtain target region prediction results based on the original image; and the region guidance module is used to obtain cell prediction results based on the cell features and the target region prediction results.
[0033] The loss function determination unit is used to determine the loss function of the heterogeneous cell detection model based on the cell position likelihood map and cell-level annotations.
[0034] In a third aspect, a method for detecting heterogeneous cells based on deep learning is provided, comprising:
[0035] Input each image block of the original image to be identified into the heterogeneous cell detection model in sequence to obtain the prediction result corresponding to the image block;
[0036] The prediction results corresponding to all image blocks are spliced together to obtain the cell detection results of the original image to be identified;
[0037] The heterogeneous cell detection model is obtained according to the above-mentioned deep learning-based heterogeneous cell detection model construction method.
[0038] In a fourth aspect, a heterogeneous cell detection device based on deep learning is provided, comprising:
[0039] The prediction unit is used to input each image block of the original image to be identified into the heterogeneous cell detection model in sequence to obtain the prediction result corresponding to the image block; the heterogeneous cell detection model is obtained according to the above-mentioned deep learning-based heterogeneous cell detection model construction method.
[0040] The splicing unit splices the prediction results corresponding to all image blocks to obtain the cell detection results of the original image to be identified.
[0041] Compared with the prior art, this application has the following beneficial effects:
[0042] (1) Inflammatory cells in endometrial tissue pathological sections are characterized by heterogeneous distribution. Different types of cells have similar characteristics but are distributed in different regions. This application mines the implicit regional information from cell-level annotations, thereby adding regional constraints to the detection results of heterogeneous cells and making corrections, thereby effectively improving the accuracy of heterogeneous cell detection.
[0043] (2) Compared with existing heterogeneous cell detection methods, this application only requires cell-level annotation and does not require further region-level annotation, which greatly reduces the annotation cost required by existing methods.
[0044] (3) The precise detection of inflammatory cells in endometrial pathological sections is beneficial for the quantitative analysis of various inflammatory cells and the diagnosis of diseases such as endometritis in clinical practice, thereby promoting further research by pathologists. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] The present application may be better understood by referring to the following description in conjunction with the accompanying drawings, which together with the following detailed description are incorporated into and form a part of this specification. In the drawings:
[0046] Figure 1 A flowchart of a method for constructing a heterogeneous cell detection model based on deep learning according to an embodiment of the present application is shown;
[0047] Figure 2 A schematic diagram of heterogeneous cell detection model training according to an embodiment of the present application is shown;
[0048] Figure 3 A schematic diagram of a regional learning network according to an embodiment of the present application is shown;
[0049] Figure 4 A schematic diagram of a regional guidance module according to an embodiment of the present application is shown;
[0050] Figure 5 A structural block diagram of a device for constructing a heterogeneous cell detection model based on deep learning according to an embodiment of the present application is shown. DETAILED DESCRIPTION
[0051] Exemplary embodiments of the present application are described below with reference to the accompanying drawings. For the sake of clarity and conciseness, not all features of actual embodiments are described in this specification. However, it should be understood that in the process of developing any such actual embodiment, many implementation-specific decisions may be made to achieve the developer's specific goals, and these decisions may vary from one implementation to another.
[0052] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, the accompanying drawings only show the device structure closely related to the solution according to the present application, while other details that are not closely related to the present application are omitted.
[0053] It should be understood that the present application is not limited to the described embodiments due to the following description with reference to the accompanying drawings. In this document, where feasible, the embodiments may be combined with each other, features between different embodiments may be replaced or borrowed, and one or more features may be omitted in one embodiment.
[0054] The present invention provides a method for constructing a heterogeneous cell detection model based on deep learning. Figure 1 A flowchart of a method for constructing a heterogeneous cell detection model based on deep learning according to an embodiment of the present application is shown. Figure 1 , methods include:
[0055] Step S1, obtaining model construction data, the model construction data including original images of various types of cells with heterogeneous distribution, cell position likelihood maps and cell-level annotations;
[0056] In this step, H&E-stained endometrial pathological sections are first collected, and multiple image blocks are captured at a 20× field of view magnification of the full-field image of the H&E-stained digital section to form a data set, that is, the original images of various types of cells with heterogeneous distribution. Specifically, before scanning the pathological sections, the high-throughput scanner is first adjusted to a 20× magnification. In this way, the full-field image of the digital section obtained by scanning can clearly distinguish the size, outline, and basic structure of various types of cells and tissues, facilitating further category identification. However, due to the huge pixel size of the digital pathological section, if directly input into the processor, it will cause memory overflow. Therefore, it is necessary to capture image blocks at a 20× field of view magnification of the full-field image of the digital section to form a data set. When capturing image blocks, care should be taken to ensure that their size is at least larger than 512×512 to accommodate the size of the model input image. If the size of the image block is too small, the global heterogeneous distribution information it contains will be reduced, which is not conducive to the subsequent learning of the regional features of heterogeneous cell distribution.
[0057] Then, for all the acquired image blocks, several experienced pathologists used the point labeling method to manually label and verify the four types of inflammatory cells, including lymphocytes, plasma cells, eosinophils and neutrophils, to ensure the accuracy of the labeling and reduce the subjectivity of the labeling, thereby obtaining cell-level labeling.
[0058] Finally, the cell-level table annotations are preprocessed to generate a cell position likelihood map. In this step, the manually annotated annotations are preprocessed using a cell position likelihood map to represent the difference between the manually annotated cell positions and the actual cell positions. A cell position likelihood map is a map where the manually annotated cell positions are peaked, while the remaining pixel values gradually decrease following a Gaussian distribution. In this case, cells are not labeled with a single point, but rather with a circular likelihood aperture whose pixel values follow a Gaussian distribution.
[0059] After the above processing, there is a correspondence between the original image, cell position likelihood map and cell-level annotation. After obtaining the model construction data, all image data are randomly divided into training set and test set according to 8:2 for subsequent model training and testing.
[0060] Step S2: Training the heterogeneous cell detection model based on the model construction data to obtain a trained heterogeneous cell detection model. The heterogeneous cell detection model includes a cell detection network, a region learning network, and a region guidance module. The cell detection network is used to obtain cell features based on the original image; the region learning network is used to obtain target region prediction results based on the original image; and the region guidance module is used to obtain cell prediction results based on the cell features and the target region prediction results. Here, the cell prediction results refer to the detection results of various types of cells with heterogeneous distributions.
[0061] Step S3: determining the loss function of the heterogeneous cell detection model based on the cell position likelihood map and cell-level annotations.
[0062] Specifically, Figure 2 A schematic diagram of heterogeneous cell detection model training according to an embodiment of the present application is shown, wherein the cell detection network includes a shared encoder and a first decoder (decoder-1), and the region learning network includes a shared encoder and a second decoder (decoder-2); here, the original image is input into the shared encoder for feature extraction, and the first decoder and the second decoder respectively implement the decoding of cell features and region features, and output cell features and target region prediction results, respectively; the region guidance module uses the target region prediction results to correct the cell features and output the final cell prediction results. Here, the region guidance module is used to make the network focus on the target region and ignore similar cells in the non-target region, guiding the cell detection network to further focus on the cell features in the target region, thereby improving the cell detection network's ability to distinguish similar cells in different regions, effectively avoiding false positive misjudgments, and improving the accuracy of cell detection with heterogeneous distribution.
[0063] During the model training process, the model was trained based on the NVIDIA GeForce GTX 1080Ti GPU, with 500 training iterations and a batch size of 4. The Adam optimizer was used, with an initial learning rate of 0.001, a weight decay of 1e-4, and a decay rate of 0.1. All modules were trained in an end-to-end manner to complete the training and obtain the optimal model.
[0064] The above embodiments of the present application take into account the heterogeneous distribution of inflammatory cells in endometrial tissue pathological sections. Different types of cells have similar characteristics but are distributed in different regions. By mining the implicit regional information from the cell-level annotations, regional constraints are added to the detection results of heterogeneous cells, and then corrections are made to effectively improve the accuracy of heterogeneous cell detection.
[0065] In one embodiment, the shared encoder and the first decoder form a simple and effective UNet network as a cell detection network. The UNet network is composed of multiple layers of convolution, pooling, downsampling, upsampling and feature fusion operations. In this embodiment, the last convolution layer of the U-Net is removed to obtain a feature map with 64 channels, which is used as a preliminary cell detection result for correction in subsequent processes.
[0066] In one embodiment, Figure 3 A schematic diagram of a regional learning network according to an embodiment of the present application is shown. Figure 3 , 512, 256, 128, and 64 are the number of channels in the feature map. During the encoding process, multiple max pooling and convolution operations reduce the image size but increase the number of channels. To reduce computational complexity, we first use 1×1 convolution for channel fusion, then upsample the feature map and concatenate it with the underlying features of the corresponding layer in the encoder. This is because the multi-layer pooling operations in the encoder lose detailed image information, which can be supplemented by concatenating with the underlying features. Finally, a 1×1 convolution prediction layer is used to output the target region prediction result.
[0067] In one embodiment, Figure 4 A schematic diagram of a regional guidance module according to an embodiment of the present application is shown. The regional guidance module is implemented by an activation function, matrix element-by-element multiplication, and convolution operations. It is used to use the target area identified by the regional learning network to integrate and correct the preliminary detection results obtained by the cell detection network, adjust the feature weights of different regions, and make the network focus on the target distribution area and ignore similar cells in non-distribution areas, thereby avoiding false positives and improving detection accuracy. Figure 4, the feature map output by the first decoder of the cell detection network is recorded as F1, and the target region prediction result output by the regional learning network is activated by the Sigmoid function to obtain the regional weight matrix. In order to correct the cell detection results in different regions, the feature map F1 and the regional weight matrix are fused by element-by-element multiplication to obtain F final Then, in F final Perform 1×1 convolution and Softmax activation on the network to obtain the cell prediction result, called P cell .
[0068] In one embodiment, the loss function of the heterogeneous cell detection model is determined based on the cell-level annotations and the cell position likelihood map, which may specifically include:
[0069] Step S31, inputting the cell-level annotation into the region acquisition module to obtain a region mask;
[0070] In this step, the position points of each cell in the cell-level annotation constitute a point set S; based on the point set S and the loss function parameter α, the Alpha-shape algorithm is used to process and obtain the boundary; the boundary is filled inside to obtain the true cell distribution area of the original image, that is, the region mask.
[0071] Step S32: Obtain the regional loss function L according to the regional mask and regional prediction results. region :
[0072]
[0073] Among them, L region is the regional loss function, Ω r is the pixel set of the region mask, r j is the jth pixel in the pixel set, x is the input image, f(x) is the target region prediction result, and r is the region mask.
[0074] Step S33: Obtain the detection loss function L according to the cell position likelihood map and the cell prediction result. det :
[0075] L det =-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p)
[0076] Among them, L det is the detection loss function, where α and γ are loss function parameters, p is the cell prediction result, and y is the cell position likelihood map. In this embodiment, α = [5, 5, 5, 15] and γ = 2.
[0077] Step S34: Determine the loss function L of the heterogeneous cell detection model based on the regional loss function and the detection loss function. total :
[0078] L total =L det +λL region
[0079] Among them, L total is the loss function of the heterogeneous cell detection model, L det is the detection loss function, L region is the regional loss function, λ is the coefficient of the regional loss function, here, λ = 0.1.
[0080] In a second aspect, based on the same inventive concept as the method for constructing a heterogeneous cell detection model based on deep learning provided in the embodiment of the present application, the embodiment of the present application also provides a device for constructing a heterogeneous cell detection model based on deep learning. Figure 5 The structure block diagram of the device for constructing a heterogeneous cell detection model based on deep learning according to an embodiment of the present application is shown, and the device includes:
[0081] A model building data acquisition unit 51 is used to acquire model building data, wherein the model building data includes original images of various types of cells with heterogeneous distribution, cell position likelihood maps, and cell-level annotations;
[0082] The model training unit 52 is used to train the heterogeneous cell detection model based on the original image to obtain a trained heterogeneous cell detection model; the heterogeneous cell detection model includes a cell detection network, a region learning network, and a region guidance module; the cell detection network is used to obtain cell features based on the original image; the region learning network is used to obtain target region prediction results based on the original image; and the region guidance module is used to obtain cell prediction results based on the cell features and the target region prediction results.
[0083] a loss function determination unit 53, configured to determine a loss function of a heterogeneous cell detection model based on the cell position likelihood map and the cell-level annotations;
[0084] The above embodiments of the present application take into account the heterogeneous distribution of inflammatory cells in endometrial tissue pathological sections. Different types of cells have similar characteristics but are distributed in different regions. By mining the implicit regional information from the cell-level annotations, regional constraints are added to the detection results of heterogeneous cells, and then corrections are made to effectively improve the accuracy of heterogeneous cell detection.
[0085] The implementation functions of each module in the heterogeneous cell detection model construction device based on deep learning provided in the embodiment of the present application are consistent with the heterogeneous cell detection model construction method based on deep learning and will not be described in detail.
[0086] In a third aspect, the present application also provides a method for detecting heterogeneous cells based on deep learning, comprising:
[0087] Each image block of the original image to be identified is sequentially input into the heterogeneous cell detection model to obtain a prediction result corresponding to the image block; the heterogeneous cell detection model is obtained according to the heterogeneous cell detection model construction method based on deep learning in the above embodiment.
[0088] The prediction results corresponding to all image blocks are spliced together to obtain the cell detection results of the original image to be identified.
[0089] In a fourth aspect, the present application also provides a heterogeneous cell detection device based on deep learning, comprising:
[0090] The prediction unit is used to sequentially input each image block of the original image to be identified into the heterogeneous cell detection model to obtain a prediction result corresponding to the image block; the heterogeneous cell detection model is obtained according to the deep learning-based heterogeneous cell detection model construction method of the above embodiment.
[0091] The splicing unit splices the prediction results corresponding to all image blocks to obtain the cell detection results of the original image to be identified.
[0092] In summary, this application has the following beneficial effects:
[0093] (1) Inflammatory cells in endometrial tissue pathological sections are characterized by heterogeneous distribution. Different types of cells have similar characteristics but are distributed in different regions. This application mines the implicit regional information from cell-level annotations, thereby adding regional constraints to the detection results of heterogeneous cells and making corrections, thereby effectively improving the accuracy of heterogeneous cell detection.
[0094] (2) Compared with existing heterogeneous cell detection methods, this application only requires cell-level annotation and does not require further region-level annotation, which greatly reduces the annotation cost required by existing methods.
[0095] (3) The precise detection of inflammatory cells in endometrial pathological sections is beneficial for the quantitative analysis of various inflammatory cells and the diagnosis of diseases such as endometritis in clinical practice, thereby promoting further research by pathologists.
[0096] The above descriptions are merely examples of various embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any modifications or substitutions that can be readily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.
Claims
1. A method for constructing a heterogeneous cell detection model based on deep learning, characterized in that: include: Obtaining model construction data, the model construction data including original images of various types of cells with heterogeneous distribution, cell position likelihood maps, and cell-level annotations; the cell position likelihood map refers to a map formed by recording the manually annotated cell positions in the cell-level annotations as peaks and the remaining pixel values gradually decreasing according to a Gaussian distribution; Training a heterogeneous cell detection model based on the original image to obtain a trained heterogeneous cell detection model; the heterogeneous cell detection model includes a cell detection network, a region learning network, and a region guidance module; The cell detection network is used to obtain cell features based on the original image; The region learning network is used to obtain a target region prediction result based on the original image; the region guidance module is used to obtain a cell prediction result based on the cell characteristics and the target region prediction result; A loss function of the heterogeneous cell detection model is determined according to the cell position likelihood map and the cell-level annotation.
2. The method according to claim 1, wherein in, Determining a loss function of a heterogeneous cell detection model according to the cell position likelihood map and the cell-level annotation includes: Inputting the cell-level annotation into a region acquisition module to obtain a region mask; Obtaining a regional loss function according to the regional mask and the target region prediction result; Obtaining a detection loss function according to the cell position likelihood map and the cell prediction result; A loss function of the heterogeneous cell detection model is determined according to the region loss function and the detection loss function.
3. The method according to claim 2, wherein Inputting the cell-level annotation into a region acquisition module to obtain a region mask includes: The position points of each cell in the cell-level annotation constitute a point set S; According to the point set S and the loss function parameter α, the Alpha-shape algorithm is used to process and obtain the boundary; The boundary is filled internally to obtain a region mask.
4. The method according to claim 2, wherein Obtaining a regional loss function according to the regional mask and the target region prediction result includes: Among them, L region is the regional loss function, Ω r is the pixel set of the region mask, r j is the jth pixel in the pixel set, x is the input image, f(x) is the target region prediction result, and r is the region mask.
5. The method according to claim 2, wherein The detection loss function is obtained according to the cell position likelihood map and the cell prediction result, include: L det =-αy(1-p) γ log(p)-(1-α)(1-y)p γ log(1-p) Among them, L det is the detection loss function, α and γ are loss function parameters, p is the cell prediction result, and y is the cell position likelihood map.
6. The method according to claim 2, wherein Determining the loss function of the heterogeneous cell detection model according to the region loss function and the detection loss function includes: THE total =L det +λL region Among them, L total is the loss function of the heterogeneous cell detection model, L det is the detection loss function, L region is the regional loss function, and λ is the coefficient of the regional loss function.
7. The method according to claim 1, wherein The cell detection network includes a shared encoder and a first decoder, and the region learning network includes the shared encoder and a second decoder.
8. A device for constructing a heterogeneous cell detection model based on deep learning, characterized in that: include: a model construction data acquisition unit, configured to acquire model construction data, wherein the model construction data includes original images of various types of cells with heterogeneous distribution, cell position likelihood maps, and cell-level annotations; the cell position likelihood map is a map formed by recording the manually annotated cell positions in the cell-level annotations as peaks and the remaining pixel values gradually decreasing along a Gaussian distribution; A model training unit, configured to train a heterogeneous cell detection model based on the original image to obtain a trained heterogeneous cell detection model; the heterogeneous cell detection model comprises a cell detection network, a region learning network, and a region guidance module; The cell detection network is used to obtain cell features based on the original image; The region learning network is used to obtain a target region prediction result based on the original image; the region guidance module is used to obtain a cell prediction result based on the cell characteristics and the target region prediction result; A loss function determination unit is used to determine the loss function of the heterogeneous cell detection model based on the cell position likelihood map and the cell-level annotation.
9. A method for detecting heterogeneous cells based on deep learning, characterized in that: include: Input each image block of the original image to be identified into the heterogeneous cell detection model in sequence to obtain the prediction result corresponding to the image block; Splicing the prediction results corresponding to all image blocks to obtain the cell detection result of the original image to be identified; The heterogeneous cell detection model is obtained according to the deep learning-based heterogeneous cell detection model construction method according to any one of claims 1 to 7.
10. A heterogeneous cell detection device based on deep learning, characterized in that: include: A prediction unit, configured to sequentially input each image block of the original image to be identified into the heterogeneous cell detection model to obtain a prediction result corresponding to the image block; The heterogeneous cell detection model is obtained according to the method for constructing a heterogeneous cell detection model based on deep learning according to any one of claims 1 to 7; The splicing unit splices the prediction results corresponding to all image blocks to obtain the cell detection result of the original image to be identified.