Method and device for detecting cell nuclei in histopathological images based on iterative correction
By constructing a location generation network and an iterative correction network that incorporate deep supervision mechanisms, the problems of inaccurate nucleus detection and difficulty in removing redundant targets in pathological images are solved, achieving high-precision nucleus detection and simplified suppression range selection.
Patent Information
- Application Number
- CN202111418890.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-26
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-11-26
AI Technical Summary
In existing technologies, the detection and localization of cell nuclei in pathological images is inaccurate, and the selection of the inhibition range during non-maximum inhibition is difficult, making it difficult to effectively remove redundant targets.
An iterative correction-based approach is adopted to construct a location generation network and an iterative correction network that incorporate a deep supervision mechanism. The error is gradually reduced through iterative correction to obtain the refined candidate point locations, and redundant targets are removed through a non-maximum suppression algorithm.
It improves the accuracy and localization of cell nucleus detection, enhances the distinguishability between different cell nucleus candidates, and simplifies the selection of inhibition range in the non-maximum inhibition process.
Smart Images

Figure CN114187243B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing and pattern recognition, and more specifically, to a method and device for detecting cell nuclei in histopathological images based on iterative correction. Background Art
[0002] Histopathology images are examinations of microscopic sections composed of cells, tissues, etc., which play a vital role in the diagnosis of various cancers and have been widely used in clinical practice and pathological research. Accurate analysis of breast histopathology images is very important for pathologists to grade breast cancer and evaluate prognostic factors. For example, the Nottingham histological scoring system is associated with tubule formation, nuclear pleomorphism, and mitotic counts, where nuclear pleomorphism can indicate the evolution of breast cancer and mitotic counts can assess the invasiveness of breast tumors. For quantitative analysis of histopathology images, features are usually quantified based on the appearance and shape of individual cells and the distribution and arrangement of cell clusters. The need to identify a large number of objects in numerous digital slides makes this task time-consuming and labor-intensive, and prone to manual processing errors. Therefore, it is very demanding to build computational models that can automatically and accurately detect common cellular-level objects from histological images.
[0003] Computer-aided diagnosis methods can make objective judgments using quantitative methods, which helps to improve the efficiency and accuracy of pathological routine and eliminate a lot of redundancy. However, due to the diversity of cell nuclei in size, shape, contrast and texture, as well as the high diversity of tissue morphology in different cell nucleus types and different pathological conditions, the automatic detection of cell nuclei is a challenging task. In addition, cell nuclei can overlap and cluster, which brings further complexity to the detection task, among which traditional target detection methods cannot achieve satisfactory performance. In recent years, thanks to the development of deep learning technology and large-scale biomedical data annotation, computational pathology has made great progress. In order to predict the location of cell nuclei and identify their types, the Faster R-CNN network structure for natural image target detection is introduced into the processing of histopathological images. Although RCNN may show superior performance in some specific cases, due to the significant differences between natural images and medical images, Faster R-CNN designed and trained for natural images in more cases is not fully suitable for cell-level target detection tasks in histological images. In order to reduce the redundant targets generated by detection, non-maximum suppression steps are usually adopted at these candidate locations. Since the overlap and clustering of cell nuclei make it difficult to distinguish adjacent targets, it poses a major challenge to the non-maximum suppression process, i.e., if the suppression coverage is too large, it will be disturbed by the candidates of adjacent objects; if the suppression coverage is too small, the redundancy cannot be completely eliminated, resulting in false negative and false positive detections. Summary of the invention
[0004] The main technical problem solved by the present invention is to provide a method and device for cell nucleus detection in histopathological images based on iterative correction, in view of the inaccuracy of positioning in cell nucleus detection in current pathological images and the difficulty in selecting non-maximum inhibition scales in eliminating redundant candidates.
[0005] According to one aspect of the present invention, a method for detecting cell nuclei in histopathological images based on iterative correction comprises the following steps:
[0006] Collect training sample blocks;
[0007] Construct a position generation network that incorporates deep supervision mechanisms;
[0008] Inputting the training sample block into the position generation network to obtain candidate points of the cell nucleus and offset estimates;
[0009] Obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation;
[0010] Construct an iterative correction network;
[0011] Inputting the augmented input features into the iterative correction network, gradually reducing the error through iterative correction, and obtaining the refined candidate point positions;
[0012] After the training is completed, a trained position generation network model and a trained iterative correction network model are obtained;
[0013] The tissue pathology image to be detected is sequentially input into the trained position generation network model and the trained iterative correction network model, and redundant targets are removed by a non-maximum suppression algorithm to obtain a cell nucleus detection result.
[0014] Preferably, the training sample block includes positive samples and negative samples;
[0015] The positive sample is sampled by taking the position of the center point of the artificial label of the cell nucleus in the tissue pathology image database as a reference point, performing a random offset, and sampling the image block with the offset position as the center, and the positive sample includes the category information and the center offset information of the image block;
[0016] The negative samples are collected by random sampling and only contain the category information of the image blocks.
[0017] Preferably, the position generation network comprises: a backbone network, a side output layer and a weighted fusion layer;
[0018] The backbone network is CNN;
[0019] There are multiple side output layers, and the input end of each side output layer is respectively connected to each pooling layer and the last convolution layer of CNN;
[0020] The output end of each side output layer is connected to the weighted fusion layer.
[0021] Preferably, the step of obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation comprises:
[0022] For each pathological image in the training sample, the original candidate point image block is obtained by taking the candidate point as the center;
[0023] Each pathological image in the training sample is translated according to the offset estimation and then intercepted to obtain an offset image block;
[0024] An augmented input feature is obtained according to the original candidate point image block and the offset image block.
[0025] Preferably, the step of obtaining augmented input features according to the original candidate point image block and the offset image block comprises:
[0026] Performing differential subtraction between the original candidate point image block and the offset image block to obtain a differential image block;
[0027] The differential image block is used as new information and connected in series with the original candidate point image block block by block to obtain augmented input features.
[0028] Preferably, the step of inputting the augmented input features into the iterative correction network, gradually reducing the error through iterative correction, and obtaining the refined candidate point positions comprises:
[0029] Using the augmented input features as input of the iterative correction network, using the residual between the last estimated offset and the true value as the true value, and obtaining a predicted value of the current offset;
[0030] Taking the predicted value of the current offset as input, updating the residual according to the current estimated offset, and using the updated residual as the new true value for the next round of model training;
[0031] At the same time, the new predicted offset is translated to obtain a new differential image block, which is then concatenated with the original candidate point image block block by block to obtain a new augmented input matrix;
[0032] The new augmented input features are used as the input of the next round of model, so as to gradually reduce the error through iterative correction and obtain the refined candidate point positions.
[0033] Preferably, the position generation network and the iterative correction network are trained to minimize the objective function using back propagation and stochastic gradient descent optimization methods.
[0034] According to another aspect of the present invention, a device for detecting cell nuclei in histopathological images based on iterative correction comprises the following modules:
[0035] A training sample block collection module, used for collecting training sample blocks;
[0036] The location generation network building module is used to build a location generation network that incorporates a deep supervision mechanism;
[0037] A candidate point acquisition module, used for inputting the training sample block into the position generation network to obtain candidate points of the cell nucleus and offset estimates;
[0038] An input feature augmentation module, used for obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation;
[0039] Iterative correction network construction module, used to construct an iterative correction network;
[0040] A candidate point refinement module, used for inputting the augmented input features into the iterative correction network, gradually reducing the error through iterative correction, and obtaining a refined candidate point position;
[0041] A detection model acquisition module, used to obtain a trained position generation network model and a trained iterative correction network model after training is completed;
[0042] The cell nucleus detection module is used to input the tissue pathology image to be detected into the trained position generation network model and the trained iterative correction network model in sequence, and remove redundant targets through a non-maximum suppression algorithm to obtain a cell nucleus detection result.
[0043] The beneficial effects brought by the technical solution provided by the present invention are:
[0044] A position generation network with deep supervision mechanism is constructed to integrate the position expression ability into the learning of image patches, and realize the classification and positioning tasks of cell nucleus detection at the same time. This image patch learning method can provide a flexible and effective method for generating cell nucleus candidates. Unlike the traditional RPN fixed sampling method, this method can provide enough positive samples to learn various situations of nuclei. The deep supervision mechanism is suitable for the nucleus detection model and provides a rich hierarchical representation. Each layer of the network is responsible for generating predictions of a certain scale, so as to obtain consistent contributions from each layer. An iterative correction network is constructed. The iterative correction strategy can not only obtain higher nucleus positioning accuracy, but also make multiple candidate nuclei of the same target tend to their true values, enhance the distinguishability between different cell nucleus candidates, and facilitate the selection of suppression range in the non-maximum suppression process. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0046] Figure 1 A network structure diagram of the application of the iterative correction method in the detection of cell nuclei in breast cancer tissue pathology images according to an embodiment of the present invention;
[0047] Figure 2 is an example diagram of a training sample block in an embodiment of the present invention, wherein Figure 2 (a) is a positive sample block, Figure 2 (b) is the negative sample block;
[0048] Figure 3 is a network structure of a location generation network (LPN) in an embodiment of the present invention;
[0049] Figure 4 4 is a structural diagram of a device for detecting cell nuclei in histopathological images based on iterative correction in an embodiment of the present invention. DETAILED DESCRIPTION
[0050] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0051] The cell nucleus detection model used in the cell nucleus detection method for histopathological images based on iterative correction provided in this embodiment consists of two parts, namely, a position generation network (LPN) and an iterative correction network (ICN). The position generation network selects CNN as the backbone network and adds a side output layer under a deep supervision mechanism to generate candidate positions. The iterative correction network refines the position by iteratively merging difference clues and feedback error prediction.
[0052] The cell nucleus detection method of histopathological images can be used for cell nucleus detection of various histopathological images, such as Figure 1As shown, this embodiment applies the method to cell nucleus detection in breast cancer tissue pathology images, specifically including the following steps:
[0053] Step 1: Collection of training sample blocks:
[0054] In this embodiment, the size of the training sample block is set to 40×40 pixels, the positive-negative sample ratio is 1 to 2, the positive sample sampling is based on the artificial label center point position of the cell nucleus in the breast cancer tissue pathology image database as the reference point, and a random offset (within the range of 0 to 10 pixels) is performed, and the image block is sampled with the offset position as the center. The positive sample contains the category and center offset information of the image block; the negative sample is generated by random sampling, and it is ensured that the center distance of the negative sample is greater than 30 pixels from the nearest artificial label (positive sample), and the negative sample only contains the category information of the image block.
[0055] like Figure 2 As shown, Figure 2 (a) is a positive sample block, Figure 2 (b) is a negative sample block. The positioning capability is increased in the block-based learning process. The positive samples are sampled after adding random offsets to the manually marked nucleus center positions, and the offsets are saved for position regression. Negative samples are randomly generated, and the center of the negative samples is ensured to be far away from all manually marked nucleus centers (greater than 30 pixels). Because negative samples do not contain any nuclei or only part of the nuclei, they do not participate in position regression.
[0056] Step 2: Build a location generation network (LPN) with deep supervision mechanism to obtain candidate points of the cell nucleus.
[0057] The training sample blocks are input into the position generation network (LPN). The input data in the training phase is the image block, and the category judgment of the image block and the prediction of the center offset of the positive sample are learned. In the test phase, the trained model is convolved on the entire image to generate an anchor point array. This process is implemented using a fully convolutional network structure; each anchor point has a category score and a center offset estimate, where the positive anchor point is used as a candidate point for the cell nucleus.
[0058] like Figure 3As shown, in the position generation network (LPN), CNN is used as the backbone network for feature extraction, and the side output layer is connected to each pooling layer and the last convolution layer, and supervised learning is implemented for each side output layer. In each side output layer path, a SE block is first inserted to recalibrate the channel features, and then a convolution layer with a kernel size of 1×1 is used to combine the features of different channels. Finally, a full convolution layer is used as a classifier to predict the category and center offset of each training block. Finally, a weighted fusion layer is used to combine side outputs from different scales. The position generation network (LPN) takes the above image blocks for training as input and outputs the category label and center offset of each block. Then, the network training is implemented on the entire image by using a full convolution architecture to obtain anchors on the entire image and output the classification map and the prediction of the center offset of the positive anchor. The candidate points of the cell nucleus are generated by the positive anchor points. Two groups of 40×40 pixel blocks are cropped from the original image, centered on the candidate position and the offset position respectively. The two groups of blocks are subtracted to obtain their difference blocks. The candidate blocks, difference blocks, and the corresponding residuals between the estimated offsets and their true values are fed into an iterative correction network (ICN) to refine the kernel locations.
[0059] Step 3: Obtain new information and augment input features.
[0060] For each image, the original candidate point image block is captured with the candidate point as the center, and then the offset image block is captured after translation according to the offset estimation. The two groups of image blocks are differentially subtracted, and the generated differential image blocks are used as new information and are concatenated block by block with the original candidate point image blocks to obtain augmented input features.
[0061] Step 4: Construct an iterative correction network (ICN) to refine the candidate point locations.
[0062] The iterative correction network (ICN) takes the augmented input features as input, and the residual between the last estimated offset and its true value as the prediction target. Then, the residual is updated according to the current estimate and used as the new true value for the next round of model training. At the same time, it translates according to the newly predicted offset to obtain a new differential image block, which is concatenated with the original candidate point image block block by block to obtain a new augmented input feature as the input for the next round of model training. Through iterative correction, the error is gradually reduced, making the prediction result closer to the actual situation.
[0063] Step 5: Minimize the objective function through standard back propagation and stochastic gradient descent (SGD) optimization methods to finally obtain two trained network models.
[0064] Specifically, a position generation network (LPN) is supervised by training data containing the category and center offset information of each image patch. The trained LPN model is applied to the entire image to generate candidate locations of cell nuclei, and the nearest artificial label center is assigned to each candidate nucleus based on the Euclidean distance. The residual is calculated from the estimated position and its true position and used as the initial correction value of the iterative correction network (ICN). In each iteration of the ICN, the correction is updated by the current estimate.
[0065] The objective function of the position generation network consists of two parts: classification loss and position regression loss:
[0066] The classification loss is used to perform binary classification of image patches, i.e., with kernel and without kernel, and is defined by the cross entropy function, which is expressed as follows:
[0067]
[0068] Where c represents the true value of the binary label of the input image block, that is, the non-nucleus image block is 0 and the nucleus image block is 1. is the predicted kernel probability in [0, 1] obtained using the Softmax classifier, i is the index, represents the predicted kernel probability in [0, 1] obtained by the Softmax classifier when the index is i, c i Represents the true value of the binary label of the input image block when the index is i;
[0069] The regression loss is defined by the Euclidean L2 norm and is expressed as follows:
[0070]
[0071] Among them, 1(·) is the indicator function, indicating that only positive anchor points are involved in the regression calculation, (x, y) is the true value of the center offset of the cell nucleus in the image block, is the offset estimated by the location generation network, i is the index, (x i ,y i ) is the true value of the center offset of the cell nucleus in the image block when the index is i, is the offset estimated by the position generation network when the index is i;
[0072] The overall objective function of the position generation network is calculated by combining the side output layer and the weighted fusion layer as follows:
[0073]
[0074] Where W represents all parameters that need to be learned by the position generation network, k represents different side output layers, fuse represents the weighted fusion layer, and λ is a hyperparameter for balancing classification and regression tasks, which is set to 2 in this embodiment. Through standard back propagation and stochastic gradient descent (SGD) optimization methods, the objective function is minimized with respect to the parameter W, and finally two trained network models are obtained.
[0075] Step 6: Input the tissue pathology images to be detected into the two trained network models in sequence, use the non-maximum suppression algorithm to remove redundant targets, and finally output the cell nucleus detection results.
[0076] In this embodiment, the performance evaluation index of the method for detecting cell nuclei by combining LPN and ICN models in the present invention is compared with five other traditional methods. The five traditional methods are SAE (Stacked Autoencoder), SSAE (Stacked sparse autoencoder), SR-CNN (Super-Resolution Convolutional Neural Network), RPN (region proposal network), SPRN (similarity-based region proposal networks). In addition, the method for detecting cell nuclei by combining LPN and ICN models is compared with the method for detecting cell nuclei by using LPN alone. The comparison results are shown in Table 1:
[0077] Table 1 Comparison of evaluation indicators of the method of the present invention and other methods
[0078]
[0079] The comparison results show that the method of detecting cell nuclei using LPN and ICN models adopted in the present invention is significantly superior to other methods in terms of performance evaluation indicators such as Precision, Recall, F-score, and AP (average accuracy).
[0080] like Figure 4 As shown, this embodiment also provides a tissue pathology image cell nucleus detection device based on iterative correction, including the following modules:
[0081] The training sample block acquisition module 1 is used to acquire training sample blocks;
[0082] Position generation network construction module 2, used to construct a position generation network that incorporates a deep supervision mechanism;
[0083] A candidate point acquisition module 3 is used to input the training sample block into the position generation network to obtain candidate points of the cell nucleus and offset estimates;
[0084] An input feature augmentation module 4 is used to obtain augmented input features according to the candidate points of the cell nucleus and the offset estimation;
[0085] An iterative correction network construction module 5, used for constructing an iterative correction network;
[0086] The candidate point refinement module 6 is used to input the augmented input features into the iterative correction network, gradually reduce the error through iterative correction, and obtain the refined candidate point position;
[0087] A detection model acquisition module 7 is used to obtain a trained position generation network model and a trained iterative correction network model after the training is completed;
[0088] The cell nucleus detection module 8 is used to input the tissue pathology image to be detected into the trained position generation network model and the trained iterative correction network model in sequence, and remove redundant targets through the non-maximum suppression algorithm to obtain the cell nucleus detection result.
[0089] This embodiment provides a deep network structure including a position generation network and an iterative correction network to detect cell nuclei in breast cancer tissue pathology images. First, candidate objects and their preliminary positions are generated; then the positioning is refined. A novel block learning method is designed to obtain high-quality nuclear candidates, in which, in addition to the category, position expressions are added to the image block information, realizing a multi-task learning process for cell nucleus classification and positioning. At the same time, a deep supervision mechanism is introduced into the task, providing a rich hierarchical representation. In order to improve the accuracy of cell nucleus positioning, an iterative correction strategy is proposed to make the prediction gradually approach the actual situation, significantly improve the positioning accuracy of the cell nucleus center, and facilitate the selection of the suppression range in the non-maximum suppression step.
[0090] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0091] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0092] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for detecting cell nuclei in histopathological images based on iterative correction, characterized in that: The following steps are involved: Collect training sample blocks; Construct a position generation network that incorporates deep supervision mechanisms; Inputting the training sample block into the position generation network to obtain candidate points of the cell nucleus and offset estimates; Obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation; Construct an iterative correction network; Using the augmented input features as input of the iterative correction network, using the residual between the last estimated offset and the true value as the true value, and obtaining a predicted value of the current offset; Taking the predicted value of the current offset as input, updating the residual according to the current estimated offset, and using the updated residual as the new true value for the next round of model training; At the same time, the new predicted offset is translated to obtain a new differential image block, which is then concatenated with the original candidate point image block block by block to obtain a new augmented input matrix; The new augmented input features are used as the input of the next round of model, so as to gradually reduce the error through iterative correction and obtain the refined candidate point position; After the training is completed, a trained position generation network model and a trained iterative correction network model are obtained; The tissue pathology image to be detected is sequentially input into the trained position generation network model and the trained iterative correction network model, and redundant targets are removed by a non-maximum suppression algorithm to obtain a cell nucleus detection result.
2. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 1, characterized in that: The training sample block includes positive samples and negative samples; The positive sample is sampled by taking the position of the center point of the artificial label of the cell nucleus in the tissue pathology image database as a reference point, performing a random offset, and sampling the image block with the offset position as the center, and the positive sample includes the category information and the center offset information of the image block; The negative samples are collected by random sampling and only contain the category information of the image blocks.
3. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 1, characterized in that: The position generation network includes: a backbone network, a side output layer and a weighted fusion layer; The backbone network is CNN; There are multiple side output layers, and the input end of each side output layer is respectively connected to each pooling layer and the last convolution layer of CNN; The output end of each side output layer is connected to the weighted fusion layer.
4. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 1, characterized in that: The step of obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation comprises: For each pathological image in the training sample, the original candidate point image block is obtained by taking the candidate point as the center; Each pathological image in the training sample is translated according to the offset estimation and then intercepted to obtain an offset image block; An augmented input feature is obtained according to the original candidate point image block and the offset image block.
5. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 4, characterized in that: The step of obtaining the augmented input feature according to the original candidate point image block and the offset image block comprises: Performing differential subtraction between the original candidate point image block and the offset image block to obtain a differential image block; The differential image block is used as new information and connected in series with the original candidate point image block block by block to obtain augmented input features.
6. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 1, characterized in that: In the process of training the position generation network and the iterative correction network, back propagation and stochastic gradient descent optimization methods are used to minimize the objective function.
7. The method for detecting cell nuclei in histopathological images based on iterative correction according to claim 1, characterized in that: In the step of training the location generation network, the objective function of the location generation network consists of two parts: classification loss and location regression loss: The classification loss is used to perform binary classification of image patches, i.e., with kernel and without kernel, and is defined by the cross entropy function, which is expressed as follows: Where c represents the true value of the binary label of the input image block, that is, the non-nucleus image block is 0 and the nucleus image block is 1. is the predicted kernel probability in [0, 1] obtained using the Softmax classifier, i is the index, represents the predicted kernel probability in [0, 1] obtained by the Softmax classifier when the index is i, c i Represents the true value of the binary label of the input image block when the index is i; The regression loss is defined by the Euclidean L2 norm, expressed as follows: Among them, 1(·) is the indicator function, indicating that only positive anchor points are involved in the regression calculation, (x, y) is the true value of the center offset of the cell nucleus in the image block, is the offset estimated by the location generation network, i is the index, (x i ,y i ) is the true value of the center offset of the cell nucleus in the image block when the index is i, is the offset estimated by the position generation network when the index is i; The overall objective function of the position generation network is calculated by combining the side output layer and the weighted fusion layer as follows: Where W represents all the parameters that the position generation network needs to learn, k represents different side output layers, fuse represents the weighted fusion layer, and λ is a hyperparameter for balancing classification and regression tasks.
8. A device for detecting cell nuclei in histopathological images based on iterative correction, characterized in that: Includes the following modules: A training sample block collection module, used for collecting training sample blocks; The location generation network building module is used to build a location generation network that incorporates a deep supervision mechanism; A candidate point acquisition module, used for inputting the training sample block into the position generation network to obtain candidate points of the cell nucleus and offset estimates; An input feature augmentation module, used for obtaining augmented input features according to the candidate points of the cell nucleus and the offset estimation; Iterative correction network construction module, used to construct an iterative correction network; A candidate point refinement module is used to use the augmented input features as inputs of the iterative correction network, and use the residual between the last estimated offset and the true value as the true value to obtain a predicted value of the current offset; use the predicted value of the current offset as input, update the residual according to the current estimated offset, and use the updated residual as the new true value for the next round of model training; translate according to the newly predicted offset to obtain a new differential image block, and connect it block by block in series with the original candidate point image block to obtain a new augmented input matrix; use the new augmented input features as inputs of the next round of models, so as to gradually reduce the error through iterative correction and obtain a refined candidate point position; A detection model acquisition module, used to obtain a trained position generation network model and a trained iterative correction network model after training is completed; The cell nucleus detection module is used to input the tissue pathology image to be detected into the trained position generation network model and the trained iterative correction network model in sequence, and remove redundant targets through a non-maximum suppression algorithm to obtain a cell nucleus detection result.
Citation Information
Patent Citations
Method and device for the inclusion of additive and multiplicative correction variables in a continuous fuel feed system
DE3642476A1
Method and system for displaying dynamic condition of pedestrians
WO2018214974A1
