A Deep Learning-Based Method for Detecting Giant Cells in Ovarian Cancer Polyploid Tumors from H&E Images

Abstract

A kind of H&E image ovarian cancer polyploid giant cell (PGCCs) detection method based on deep learning, including constructing specific data set, image annotation and data division, using OCDet model to carry out feature learning and optimization, and model training and evaluation.The method first establishes the H&E staining image data set containing PGCCs, then accurately labels the image and divides it into training, verification and test set.OCDet model takes CSPDarkNet as the core, combines ECA mechanism, focuses on the deep learning and re-encoding of pathological semantic features.Through the training data set, the model updates parameters through back propagation and gradient descent, optimizes to identify PGCCs features.The verification set is used for model tuning, and the test set is used for evaluating the performance of the model.The automatic detection technology of the application can help doctors improve the diagnosis efficiency and reduce the error, provide an important reference for clinical treatment and prognosis evaluation, and show significant clinical application value.

Description

Technical Field

[0001] This invention relates to histopathological image processing, and in particular to a method for detecting giant cells in ovarian cancer polyploid tumors based on deep learning-based H&E images. Background Technology

[0002] Ovarian cancer is one of the most dangerous gynecological malignancies threatening women's lives. Pathological diagnosis is the gold standard for determining the occurrence and development of tumors, and hematoxylin and eosin (H&E) staining is a fundamental staining method widely used in histopathological diagnosis. This technique utilizes the basic properties of hematoxylin and the acidic properties of eosin to stain different parts of the cell blue-purple or red. Through this color contrast, pathologists can clearly observe subtle changes in cell structure, providing important evidence for disease diagnosis. Typically, pathologists need to rely on morphological hematoxylin and eosin (H&E) pathological images to identify cells in various types of ovarian tissue, including polyploid giant cancer cells (PGCCs). Polyploid giant cancer cells are a special subpopulation of cancer cells characterized by single or multiple large nuclei, which are usually irregular, and are of great significance for tumor diagnosis and staging. However, the detection of polyploid giant cancer cells in H&E staining images is a complex and time-consuming process.

[0003] As a commonly used staining technique in pathological diagnosis, H&E staining images can provide rich information on cell morphology and structure. However, due to the significant morphological differences between PGCCs and traditional cells, their detection and analysis is a challenging task for pathologists.

[0004] While traditional ovarian cancer diagnosis boasts high accuracy, it suffers from drawbacks such as high subjectivity, operational complexity, and time-consuming nature. However, the rapid development of deep learning technology has opened up new possibilities for automated disease diagnosis with deep learning-based computer-aided systems. By training deep neural network models, the automatic detection and classification of lesions in medical images can be achieved, thereby improving diagnostic accuracy and efficiency. Deep learning models have been widely applied to medical image detection tasks, making it possible to automatically and accurately detect target cells in various pathological images. Therefore, applying deep learning technology to the detection of polyploid tumor giant cells in H&E-stained images has significant research value and development potential. Through deep learning algorithms, the characteristic information of polyploid tumor giant cells in H&E-stained images can be automatically learned and identified, enabling rapid and accurate detection of these cells, thus assisting pathologists in making more accurate tumor diagnoses and staging.

[0005] However, research on using deep learning technology to detect polyploid tumor giant cells in H&E staining images is still in its early stages, and many challenges and problems remain to be solved. First, how to design suitable deep learning models and optimize model parameters to improve the accuracy and stability of detecting polyploid tumor giant cells. Second, in H&E pathological images of ovarian cancer, the morphological heterogeneity often leads to semantic ambiguity in the detection model. The morphological differences of polyploid tumor giant cells and the complexity of staining images reduce the signal-to-noise ratio, requiring the detection model to construct a more complex feature space. Third, due to the large variation in the number and morphology of polyploid tumor giant cells and their relatively small number, the model needs to learn highly heterogeneous pathological features based on a limited number of samples.

[0006] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The main objective of this invention is to address the problems existing in the background art described above and to provide a method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect of the present invention, a method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning H&E images includes the following steps:

[0010] S1. Construct a dataset of H&E-stained images containing polyploid tumor giant cells (PGCCs);

[0011] S2. Label the H&E stained images to mark the location of polyploid tumor giant cells, and divide the labeled dataset into training set, validation set and test set;

[0012] S3. Construct an OCDet model based on deep features for channel attention recoding and fusion. The OCDet model uses CSPDarkNet as the feature extraction backbone and combines the attention mechanism ECA to realize the relationship modeling and recoding of pathological semantic features of different channels.

[0013] S4. The OCDet model is trained using the labeled training dataset. The model parameters are updated using the backpropagation algorithm and gradient descent optimizer, enabling the model to learn the feature representation of PGCCs. The model is tuned using the validation set during training, and evaluated using the test set after training. The trained model is used for automatic detection of ovarian cancer polyploid tumor giant cells in H&E images.

[0014] In a second aspect of the invention, a computer-readable storage medium stores a computer program, characterized in that the computer program, when executed by a processor, implements the deep learning-based method for detecting giant cells in H&E images of ovarian cancer polyploid tumors.

[0015] In a third aspect of the present invention, a computer program product includes a computer program, characterized in that the computer program, when executed by a processor, implements the deep learning-based H&E image detection method for ovarian cancer polyploid tumor giant cells.

[0016] The present invention has the following beneficial effects:

[0017] This invention proposes a novel method for detecting polyploid tumor giant cells (PGCCs) in ovarian cancer using deep learning technology in hematoxylin and eosin (H&E) stained images. This method achieves efficient and accurate automated detection, possessing significant clinical implications and application value. By constructing and training a target detection model, this invention enables the automatic detection and identification of PGCCs in H&E stained images. Specifically, the proposed detection model, OCDet, learns the morphological and structural features of cells in H&E stained images, particularly focusing on modeling and optimizing the specific morphology and structure of PGCCs. Through training and optimization, this detection model can accurately distinguish PGCCs from other cells and automatically label the location and number of PGCCs in the image. Automated detection not only improves diagnostic efficiency and reduces human error but also provides more valuable reference information for subsequent clinical treatment and prognostic assessment. Furthermore, the method of this invention is scalable and customizable, adaptable to the detection needs of other target cells under different disease types and staining conditions.

[0018] Compared with existing methods for detecting giant cells in polyploid tumors, the main advantages of this invention are:

[0019] 1. Without expending a lot of manpower and resources to construct large datasets, precise modeling and optimization of the special morphology and structure of polyploid tumor giant cells were carried out, enabling accurate and efficient detection of polyploid tumor giant cells in H&E stained images.

[0020] 2. Once trained and optimized, the OCDet model can rapidly process large amounts of H&E image data, enabling automated detection of polyploid tumor giant cells. This significantly improves the efficiency of pathological diagnosis, reduces the workload of doctors, and allows more cases to be diagnosed in a shorter time.

[0021] 3. The designed deep learning model exhibits excellent scalability. With the continuous accumulation of new H&E image data, the model can be continuously trained and optimized, further improving its ability to identify polyploid tumor giant cells. Furthermore, this model can also be applied to the detection of other cell or tissue types, providing more possibilities and value for pathological diagnosis.

[0022] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0023] Figure 1 This is an example diagram of a polyploid tumor giant cell according to an embodiment of the present invention.

[0024] Figure 2 This is an overall network framework diagram of an embodiment of the present invention.

[0025] Figure 3 This is a flowchart of the ECA-RER and ECA-MREF multi-scale feature extraction and fusion process according to an embodiment of the present invention.

[0026] Figure 4 This is a comparison of the detection results of this invention with other models. Detailed Implementation

[0027] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0028] This invention provides a method for detecting giant cells in ovarian cancer polyploid tumors using H&E images based on deep learning, comprising the following steps:

[0029] S1. Construct a dataset of H&E-stained images containing polyploid tumor giant cells (PGCCs);

[0030] S2. Label the H&E stained images to mark the location of polyploid tumor giant cells, and divide the labeled dataset into training set, validation set and test set;

[0031] S3. Construct an OCDet model based on deep features for channel attention recoding and fusion. The OCDet model uses CSPDarkNet as the feature extraction backbone and combines the attention mechanism ECA to realize the relationship modeling and recoding of pathological semantic features of different channels.

[0032] S4. The OCDet model is trained using the labeled training dataset. The model parameters are updated using the backpropagation algorithm and gradient descent optimizer, enabling the model to learn the feature representation of PGCCs. The model is tuned using the validation set during training, and evaluated using the test set after training. The trained model is used for automatic detection of ovarian cancer polyploid tumor giant cells in H&E images.

[0033] The present invention proposes an innovative deep learning model, OCDet, for detecting ovarian cancer polyploid tumor giant cells in hematoxylin and eosin (H&E) pathological images.

[0034] In a preferred embodiment, the OCDet model comprises the following sequentially connected parts: a backbone network, including a feature extraction model CSPDarkNet, which extracts deep features from the input H&E stained image and constructs a mapping relationship from the input image to multi-resolution feature maps; a connection part, Neck, which further processes and optimizes the features extracted by the backbone network, including enhancing salient features and removing redundant features through the ECA-RER module, and efficiently aggregating and fusing multi-scale features through the ECA-MREF module; and a detection head, which uses feature maps of different resolutions to regress detection boxes, obtains the final detection result of polyploid tumor giant cells, and optimizes the model performance by comparing it with the ground truth through a loss function.

[0035] In a more preferred embodiment, such as Figure 2 and Figure 3 As shown, the processing of the OCDet model's backbone network includes: performing data augmentation operations to improve the quality of pathological images and enhance the feature representation of the images; adjusting the pathological images to a uniform size, such as 640 pixels × 640 pixels, to meet the input requirements of the feature extraction model; and inputting the resized images into the CSPDarkNet model for feature extraction, which converts the images into a set of multi-resolution feature maps to capture pathological information at different levels.

[0036] like Figure 2 and Figure 3As shown, the processing of the Neck connection part of the OCDet model includes: strengthening salient features and removing or suppressing redundant features in the feature map extracted from the Backbone through the ECA-RER module to optimize feature representation and achieve a refined feature embedding process; aggregating the refined features output by the ECA-RER module layer by layer through the ECA-MREF module to achieve effective feature fusion at multiple scales, resulting in a set of feature maps corresponding to different resolutions, enhancing the model's ability to identify polyploid tumor giant cells; by utilizing the two modules of the Neck connection part, the OCDet model can further improve the feature expression ability, providing richer and more accurate feature information for the final detection task.

[0037] like Figure 3 As shown, in a preferred embodiment, the specific designs of the ECA-RER module and the ECA-MREF module are as follows:

[0038] The feature maps from the backbone network (CSPDarkNet) are three feature maps at different resolutions, which our model needs to enhance through feature processing. The ECA-RER module includes multiple ECA mechanisms that perform global average pooling on the three feature maps from the backbone. Each feature map is converted into a one-dimensional vector containing global information for that channel. Then, one-dimensional convolutions are used to capture the correlation between channels and generate weight coefficients for each channel. In particular, the kernel size of the one-dimensional convolution is adaptively determined by a nonlinear mapping of the channel dimension C, which ensures that feature maps with different numbers of channels can obtain the most suitable cross-channel interactions. Subsequently, these weight coefficients are applied to the original three feature maps respectively, and the weighting operation is achieved through element-wise multiplication. The weighted feature maps, while preserving the original information, enhance the feature representation of important channels and suppress unimportant channels, achieving embedding refinement at a specific resolution. The ECA-MREF module includes multiple ECA mechanisms and bidirectional cross-resolution fusion operations. Lower-level feature maps have higher resolution and contain more detailed information, while higher-level feature maps have lower resolution but stronger semantic expressive power. ECA-MREF fuses feature maps of different resolutions in a bidirectional manner. This means that it can not only fuse high-level semantic information into low-level feature maps to enhance their semantic expressive power, but also pass low-level detailed information to high-level feature maps to supplement their detailed information. Multi-resolution embedding fusion is mainly achieved through two approaches—upsampling and downsampling. Upsampling interpolates the low-resolution feature map to match its size with the upper-level feature map, then connects and fuses them. The input feature map is divided into two parts, and then cross-connected between these two parts (multiple convolutional residual blocks), and finally merged (concatenation operation). This cross-stage connection method helps the model learn richer feature combinations and improves the model's generalization ability.

[0039] The processing of the OCDet model's detection head includes: using multiple detection heads to process feature maps of different resolutions obtained from the connection part (Neck) to regress the detection boxes of polyploid tumor giant cells and obtain predicted detection results. Further, the obtained detection results are compared with the ground truth annotations, and an error calculation is performed using a loss function. This loss function is a combination of detection loss and classification loss to optimize the accuracy of the detection boxes and classification. Specifically, the detection loss quantifies the similarity between the predicted and ground truth boxes by calculating the intersection of the predicted and ground truth boxes and dividing by the union. Simultaneously, a difference loss is introduced to optimize the prediction probability in the form of cross-entropy, enabling the network to more accurately identify the target location. The classification loss uses different calculation methods depending on the presence or absence of labels. For cases with labels, a weighted cross-entropy loss is used; for cases without labels, a logarithmic loss function with adjusted weight coefficients is used. Through this processing method of the detection head, the OCDet model can effectively and accurately locate and classify polyploid tumor giant cells, while continuously optimizing the model's detection performance under the guidance of the loss function.

[0040] A key innovative design of this invention includes the OCDet model and its components, which improves the model's detection performance through optimized feature representation and multi-scale feature fusion. The OCDet model achieves efficient and accurate detection of polyploid tumor giant cells by combining the Backbone, Neck, and Head components. Specifically, in the Neck component, this invention designs two modules: ECA-RER and ECA-MREF. The ECA-RER module optimizes feature representation by enhancing saliency features and removing redundant features; the ECA-MREF module enhances the model's ability to detect polyploid tumor giant cells through multi-scale feature fusion.

[0041] The above embodiments propose an innovative deep learning-based method for detecting ovarian cancer polyploid tumor giant cells in H&E images. This method significantly improves the efficiency and accuracy of pathological image diagnosis by constructing an automated OCDet model. The OCDet model cleverly integrates CSPDarkNet as a backbone for deep feature extraction. Furthermore, the ECA-RER module enhances feature saliency and removes redundant features, while the ECA-MREF module achieves effective aggregation and fusion of multi-scale features. This structural design enables the model to not only accurately capture the special morphological and structural features of polyploid tumor giant cells, but also learn highly heterogeneous pathological features based on a small number of samples, solving the problems of high subjectivity, complex operation, and long time consumption in traditional detection processes. As an auxiliary diagnostic tool, the detection method of this invention can effectively reduce the workload of doctors, enabling more cases to be diagnosed quickly. It also has good scalability and customizability, adapting to the cell detection needs of different disease types and staining conditions, providing an efficient and accurate solution for the detection and auxiliary diagnosis of ovarian cancer polyploid tumor giant cells in H&E images.

[0042] The following describes specific embodiments of the present invention.

[0043] The method of this invention mainly includes the following steps: First, plots are extracted from whole-slide images of ovarian cancer and a dataset is constructed through manual annotation; second, the OCDet model is trained by loading the training set to detect polyploid tumor giant cells in H&E-stained images; finally, the model is evaluated and optimized using the test set to achieve good generalization and stable practicality of the final model.

[0044] A method for detecting ovarian cancer polyploid tumor giant cells in hematoxylin-eosin images using deep learning technology includes:

[0045] Step S1 involves constructing a rich dataset of H&E-stained images containing polyploid tumor giant cells, and randomly cropping image patches from the whole slide images in the dataset. This process may require manual intervention to filter and correct the data. Finally, data preprocessing is performed, including image enhancement, noise removal, and color normalization, to improve image quality and consistency.

[0046] Step S2 involves professional pathologists annotating the H&E images to clearly identify the locations of giant cells in polyploid tumors. This process uses VOC annotation format to suit the requirements of the proposed model. Finally, the annotated dataset is divided into training, validation, and test sets. The training set is used for model training, the validation set for model tuning, and the test set for final model evaluation.

[0047] Step S3 involves designing a deep feature-based channel attention recoding and fusion model, OCDet, based on the morphological and structural characteristics of polyploid tumor giant cells. This model uses CSPDarkNet as the feature extraction backbone and incorporates an attention mechanism (ECA) to model and recode the relationships between pathological semantic features of different channels, thereby improving model performance.

[0048] Step S4 trains the OCDet model using a labeled training dataset. The model parameters are updated using backpropagation and gradient descent optimizers, enabling the model to fully learn the feature representations of PGCCs. The trained model is evaluated using an independent test dataset, calculating metrics such as accuracy, recall, and F1 score. Performance tuning is then performed based on the evaluation results.

[0049] The specific implementation process is as follows:

[0050] Step S1-1: Collect H&E pathological image data of polyploid tumor giant cells of ovarian cancer from Peking University Third Hospital. This batch of data contains 13 whole slide images (WSIs).

[0051] Step S1-2 divides the WSIs into 512*512 pixel patches, and randomly selects 197, 66, and 66 patches for dividing the training, validation, and test sets in S1-1. The patches are manually selected and corrected to ensure that each patch contains the target cells and has minimal background interference. These patches cover the morphological diversity of polyploid tumor giant cells.

[0052] Step S2-1: Experienced pathologists annotate each H&E patch sequentially, ensuring as few omissions or errors as possible. Annotation information includes features such as the location and size of giant cells in polyploid tumors. Finally, XML files are used as the storage format for the annotation information, facilitating subsequent model reading and processing.

[0053] Step S2-2 divides the labeled dataset into a training set, a validation set, and a test set using a 6:2:2 ratio. The training set is used for model training and optimization, the validation set is used to monitor the model training process and adjust hyperparameters, and the test set is used to evaluate the model's performance.

[0054] Step S3-1 proposes OCDet for the detection of polyploid tumor giant cells in morphological H&E images. OCDet's workflow consists of three parts: Backbone, Neck, and Head. In the Backbone processing flow, the pathological image is first data augmented, then resized to 640*640 pixels, and finally input into the feature extraction model CSPDarkNet. This process constructs the following mapping:

[0055]

[0056] Where x represents a tile. Denotes the dataset, f n This represents feature maps at three resolutions.

[0057] Step S3-2, in the Neck section, involves efficiently embedding refinement and feature fusion using the three feature maps of different resolutions obtained from the Backbone, processed by the ECA-RER and ECA-MREF modules. The ECA-RER module emphasizes salient features extracted from the Backbone and removes redundant features, thereby optimizing the feature representation. The refinement process is defined as a mapping from input features to weighted features:

[0058]

[0059] The ECA-MREF module performs efficient hierarchical aggregation of the updated features, resulting in three feature maps corresponding to different resolutions.

[0060] Step S3-3, in the Head section, OCDet uses multiple detection heads to perform bounding box regression on feature maps of different resolutions and obtains the final detection result for polyploid tumor giant cells. This result is used to calculate the loss with Ground Truth, and the loss function consists of detection loss and classification loss. The detection loss is defined as:

[0061]

[0062]

[0063]

[0064] The classification loss is defined as:

[0065]

[0066] Where D box It is a prediction box, G box This is the Ground Truth, where q is the label, α and γ are weighting coefficients, and y is the general distribution value. L dfl The significance is to optimize the probabilities of the two positions (one left and one right) closest to the label y in the form of cross-entropy, so that the network can focus on the target position more quickly.

[0067] Step S4-1 loads the labeled training dataset and the OCDet model with initialized weights. Training is divided into two phases: a freeze phase and a unfreeze phase, both with a batch size of 16. In the freeze phase, only some network parameters are fine-tuned, with 50 epochs. In the unfreeze phase, all OCDet parameters are trained to better fit the data, with 150 epochs. Simultaneously, OCDet training uses the SGD optimizer with an initial learning rate of 0.01.

[0068] Step S4-2 inputs the training data into the model and performs forward propagation through each layer of the model to obtain the prediction results. The loss is calculated based on the model's prediction results and the true labels. By calculating the gradient of the loss, the model parameters are updated using the backpropagation algorithm and the optimizer's strategy. The process of forward propagation, loss calculation, backpropagation, and parameter update is repeated until a preset number of training epochs is reached or other stopping conditions are met (such as performance on the validation set no longer improving).

[0069] Step S4-3 loads an independent test dataset and uses the trained model to make predictions on the test dataset. Based on the prediction results and the true labels, calculate the model's evaluation metrics, such as precision, recall, F1-Score, and mAP@0.5. These metrics comprehensively evaluate the model's performance on the target detection task of polyploid tumor giant cells. Optimize the model's performance based on the evaluation results. First, analyze the types and causes of errors the model makes on the test set; then, implement targeted optimization measures to address these issues, such as adjusting the model structure, modifying the loss function, and adding data augmentation; finally, retrain the model and evaluate its performance until satisfactory results are achieved.

[0070] experiment

[0071] The superiority of the method is discussed using both quantitative and qualitative metrics. Quantitative metrics include precision, recall, F1-score, and mAP@0.5, all of which can be used to evaluate model performance. To avoid data leakage, tiles from the test set are used for evaluation. The results are shown in Table 1 below (bold text represents the best results).

[0072] Table 1

[0073]

[0074] As shown in the table, the OCDet model exhibits significant advantages across multiple key metrics. For recall, OCDet achieves 77.03%, ranking first among all models. This performance demonstrates that the model of this invention can not only accurately identify real targets but also comprehensively cover all real targets as much as possible, reducing false negatives. For F1-Score and mAP@0.5, the model of this invention achieves a high score of 0.72 and a mAP@0.5, respectively, ranking first among all models. These metrics fully demonstrate the superiority of OCDet in detecting polyploid cancer giant cells.

[0075] Qualitative metrics provide a clear visual representation of the results, as shown in the figure. Regarding the detection of polyploid tumor giant cells, although Faster R-CNN did detect all polyploid cells, it incorrectly identified three targets as one and then detected it again, which clearly does not meet the detection requirements. The other three models (YOLO V7, SSD, and RetinaNet) had varying numbers of missed targets. Only OCDet accurately identified all target cells. Comparative analysis revealed that OCDet provides more comprehensive and accurate detection of target cells in morphological H&E pathological images of ovarian cancer than other methods.

[0076] In summary, this invention proposes an innovative deep learning model, OCDet, for detecting ovarian cancer polyploid tumor giant cells in hematoxylin and eosin (H&E) pathological images. Compared with existing methods, this invention has the following advantages:

[0077] 1. Without expending a lot of manpower and resources to construct large datasets, precise modeling and optimization of the special morphology and structure of polyploid tumor giant cells were carried out, enabling accurate and efficient detection of polyploid tumor giant cells in H&E stained images.

[0078] 2. Once trained and optimized, the OCDet model can rapidly process large amounts of H&E image data, enabling automated detection of polyploid tumor giant cells. This significantly improves the efficiency of pathological diagnosis, reduces the workload of doctors, and allows more cases to be diagnosed in a shorter time.

[0079] 3. The proposed deep learning model exhibits good scalability. With the continuous accumulation of new H&E image data, the model can be continuously trained and optimized, further improving its ability to identify polyploid tumor giant cells. Furthermore, this model can also be applied to the detection of other cell or tissue types, providing more possibilities and value for pathological diagnosis.

[0080] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0081] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0082] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0083] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk drive or magnetic tape drive. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0084] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0085] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0086] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0087] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0089] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0090] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0091] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0092] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting giant cells in ovarian polyploid tumors using H&E images based on deep learning, characterized in that, Includes the following steps: S1. Construct a dataset of H&E-stained images containing polyploid tumor giant cells (PGCCs); S2. Label the H&E stained images to mark the location of polyploid tumor giant cells, and divide the labeled dataset into training set, validation set and test set; S3. Construct an OCDet model based on deep features for channel attention recoding and fusion. The OCDet model uses CSPDarkNet as the feature extraction backbone and combines the attention mechanism ECA to realize the relationship modeling and recoding of pathological semantic features of different channels. The OCDet model comprises the following sequentially connected parts: a backbone network, including the feature extraction model CSPDarkNet, which extracts deep features from the input H&E stained image and constructs a mapping relationship from the input image to multi-resolution feature maps; a connection part, the Neck, which further processes and optimizes the features extracted by the backbone network, including enhancing salient features and removing redundant features through the ECA-RER module, and aggregating and fusing multi-scale features through the ECA-MREF module; and a detection head, which uses feature maps of different resolutions to regress detection boxes, obtains the final detection results of polyploid tumor giant cells, and optimizes the model performance by comparing the loss function with the ground truth annotation Ground Truth. For the ECA-RER module, a global average pooling operation is performed on each feature map from the backbone to convert the multidimensional feature map into a one-dimensional feature vector. The correlation between channels is captured by one-dimensional convolution, generating weight coefficients for each channel. The kernel size of the one-dimensional convolution is adaptively determined based on the channel dimension C of the feature map through a nonlinear mapping to achieve optimal interaction across channels. The weight coefficients obtained from the ECA-RER module are applied to the original feature map and weighted by element-wise multiplication to enhance the features of important channels and suppress unimportant channels, thus achieving refined feature embedding. For the ECA-MREF module, a bidirectional cross-resolution fusion operation is adopted to fuse low-level high-resolution feature maps with high-level low-resolution feature maps. Size matching is achieved through upsampling and downsampling, and the semantic expressiveness and detailed information of the feature maps are enhanced through cross-connection and merging operations. The cross-connection consists of multiple convolutional residual blocks to promote the model to learn richer feature combinations and improve the model's generalization ability. S4. The OCDet model is trained using the labeled training dataset. The model parameters are updated using the backpropagation algorithm and gradient descent optimizer, enabling the model to learn the feature representation of PGCCs. The model is tuned using the validation set during training, and evaluated using the test set after training. The trained model is used for automatic detection of ovarian cancer polyploid tumor giant cells in H&E images.

2. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in claim 1, characterized in that, In step S1, patches are randomly cut from the whole glass slide image and manually screened and corrected; the patches are preprocessed, including image enhancement, noise removal and color standardization.

3. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in claim 1 or 2, characterized in that, The processing of the OCDet model's backbone network includes: Perform data augmentation operations to improve the quality of pathological images and enhance the feature representation of the images; Adjust the pathology images to a uniform size; The resized image is input into the CSPDarkNet model for feature extraction. This model converts the image into a set of multi-resolution feature maps, thereby capturing pathological information at different levels.

4. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in claim 3, characterized in that, The processing of the Neck connection part of the OCDet model includes: The ECA-RER module is used to enhance salient features and remove or suppress redundant features in the feature map extracted from the backbone to optimize feature representation and achieve a refined feature embedding process. The refined features output by the ECA-RER module are aggregated layer by layer through the ECA-MREF module to achieve the fusion of multi-scale features and obtain a set of feature maps corresponding to different resolutions, thereby enhancing the model's ability to identify polyploid tumor giant cells.

5. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in claim 3, characterized in that, The processing of the detection head (Head) of the OCDet model includes: Multiple detection heads are used to process feature maps of different resolutions obtained from the connecting part Neck to achieve detection box regression of polyploid tumor giant cells and obtain predicted detection results; The obtained detection results are compared with the ground truth annotation, and the error is calculated using a loss function, which is a combination of detection loss and classification loss. The detection loss quantifies the similarity between the predicted box and the ground truth box by calculating the intersection of the predicted box and the ground truth box and dividing by the union. It also introduces a difference loss to optimize the prediction probability in the form of cross-entropy so that the network can more accurately identify the target location. The classification loss is calculated differently depending on whether a label is present or not. For cases where a label is present, a weighted cross-entropy loss is used, while for cases where no label is present, a logarithmic loss function with weight coefficients adjusted is used.

6. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in any one of claims 1 to 2, characterized in that, In step S4, the training process of the OCDet model includes two stages: Freeze phase: In this phase, only some parameters of the model are fine-tuned to stabilize the basic performance of the model. The number of training iterations in this phase is a set number of cycles. Thawing phase: After the freezing phase is complete, all parameters of the model are used for training to further improve the model's fit to the data and detection performance. The number of training iterations in this phase is more cycles.

7. The method for detecting giant cells of ovarian cancer polyploid tumors based on deep learning in H&E images as described in claim 6, characterized in that, During the freezing and unfreezing phases, the model is trained using a uniform batch size, and the model parameters are updated using the backpropagation algorithm and the gradient descent optimizer.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based method for detecting giant cells of ovarian cancer polyploid tumors in H&E images as described in any one of claims 1-7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based method for detecting giant cells of ovarian cancer polyploid tumors in H&E images as described in any one of claims 1-7.

Images