Breast cancer pathological image mitosis detection method based on deep convolutional neural network
Through the deep convolutional neural network, the breast cancer pathological image detection method is automatically extracted, which solves the problem of time-consuming and labor-intensive and inaccurate traditional methods, and achieves high-precision and stable detection of mitotic cells.
Patent Information
- Application Number
- CN202510366916.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2025-07-11
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional breast cancer pathological image mitosis detection methods are time-consuming and labor-intensive and susceptible to subjective factors of the observer. Existing image processing methods are difficult to accurately extract the characteristics of mitotic cells, resulting in inconsistency and inaccuracy of the detection results.
Deep convolutional neural network-based detection methods are adopted, including data collection and labeling, preprocessing, network construction and training, post-processing of detection results and model verification, combined with transfer learning and custom modules, and automatic detection of mitotic cells is performed using cross-entropy loss function and IoU loss function.
It significantly improves the accuracy and robustness of mitotic cell detection, can accurately locate cells in complex backgrounds, reduce interference from human factors, and improve the stability and efficiency of detection results.
Smart Images
Figure CN120298349A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network. Background Art
[0002] In breast cancer pathological diagnosis, the detection of mitotic cells plays a crucial role in evaluating the proliferative activity of tumors, predicting prognosis, and formulating treatment plans. Traditional mitosis detection methods rely on pathologists to manually observe pathological sections under a microscope and mark all mitotic cells. However, this method is not only time-consuming and laborious but also easily affected by the subjective factors of observers, resulting in inconsistencies and inaccuracies in detection results.
[0003] In recent years, with the rapid development of computer vision and deep learning technologies, automatic detection methods based on image processing have gradually become a research hotspot. Although some existing image processing methods, such as those based on morphological processing, feature extraction, and machine learning classifiers, have improved the detection efficiency of mitotic cells to a certain extent, these methods still face many challenges.
[0004] Breast cancer pathological images usually contain a large amount of complex background information, such as a large number of cell overlaps, uneven staining, and noise interference, which makes it difficult for traditional image processing methods to accurately extract the features of mitotic cells. Secondly, mitotic cells have different morphological and texture features at different stages, which makes it very difficult to design a general detection algorithm. Finally, existing detection methods usually rely on manually designed features, which often cannot comprehensively capture the complex changes of cells, thus limiting the further improvement of detection accuracy.
[0005] In response to this, the inventor proposes a method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network to solve the above problems. Summary of the Invention
[0006] The purpose of the present invention is to provide a method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network to solve the problems raised in the above background art.
[0007] To achieve the above purpose, the present invention provides the following technical solutions:
[0008] A method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network includes the following steps:
[0009] S1. Data collection and annotation: Collect a dataset containing breast cancer pathological images and annotate the mitotic cells in the images. The content of the annotation includes position information and category information;
[0010] S2. Data preprocessing: Preprocess the dataset to obtain preprocessed images;
[0011] S3. Network construction and training: Build a mitosis detection model based on a deep convolutional neural network, optimize the network structure by combining transfer learning and custom modules, and use training data to train the model. Use the cross-entropy loss function to detect mitotic cell categories and combine the IoU loss function to locate cell positions;
[0012] S4. Post-processing of detection results: Post-process the output results of the detection model, including non-maximum suppression and false positive filtering, to improve the accuracy and robustness of the detection;
[0013] S5. Model validation and evaluation: Validate the model performance on an independent test dataset and data from different sources, and evaluate the model's performance in terms of sensitivity, specificity, and accuracy;
[0014] S6. System deployment: Apply the trained model to the clinical environment, combine visualization tools to display detection results, and support real-time inference and continuous optimization.
[0015] Preferably, the data preprocessing step includes cropping, scaling, normalizing, and data augmentation of the images. The data augmentation operations include random rotation, horizontal flipping, color jittering, and synthetic data generation based on the generative adversarial network GAN.
[0016] Preferably, the normalization normalizes the image pixel values to a unified range, and the formula for the normalization is:
[0017]
[0018] where I_raw: original image pixel value;
[0019] μ: mean of the image;
[0020] σ: standard deviation of the image;
[0021] I_norm: normalized image pixel value.
[0022] Preferably, the data augmentation increases the diversity of the training dataset and improves the generalization ability of the model by applying operations such as rotation, flipping, and cropping to the images:
[0023]
[0024] (x, y): original coordinates of the image;
[0025] (x′, y′): new coordinates after rotation;
[0026] θ: rotation angle.
[0027] Preferably, the mitosis detection model uses a deep convolutional neural network to extract spatial features from the input image for the mitosis detection model to learn local features of mitotic cells, where the expression of the deep convolutional neural network is:
[0028]
[0029] where x: input image or feature map;
[0030] w: convolutional kernel, filter;
[0031] y: output feature map;
[0032] (i,j): coordinates of the output feature map;
[0033] m,n: coordinates of the convolutional kernel.
[0034] Preferably, the post-processing step of the detection result uses the non-maximum suppression algorithm to merge detection boxes and filters false positive detection results by a rule-based or simple classifier method.
[0035] Preferably, the non-maximum suppression is used to process multiple overlapping bounding boxes, and only the box with the highest score is retained to eliminate redundant detection results, where the expression of the non-maximum suppression is:
[0036]
[0037] If the intersection over union IoU of two bounding boxes is greater than the set threshold of 0.5, the box with the lower score is suppressed;
[0038] A: current detection box;
[0039] B: other detection boxes overlapping with the current detection box;
[0040] IoU: intersection over union of two bounding boxes;
[0041] IoU threshold for judging whether to perform suppression.
[0042] Preferably, the cross-entropy loss function is used for classification tasks. In the classification of mitotic cells, it measures the difference between the predicted class and the true class. The expression of the cross-entropy loss function is:
[0043]
[0044] where C: number of classes;
[0045] y i: true label;
[0046] Probability value predicted by the model;
[0047] L: Loss function value.
[0048] Preferably, the model performance is evaluated by the mean average precision mAP, and its expression is:
[0049]
[0050] Where N: Number of categories;
[0051] APi: Average precision of the i-th category.
[0052] Compared with the prior art, the beneficial effects of the present invention are:
[0053] (1) By using a deep convolutional neural network, the system of the present invention can automatically extract complex features from breast cancer pathological images, such as the edges, shapes, textures, and color distributions of cells. These deep features enable the model to identify mitotic cells in subtle changes that are difficult to capture by traditional manual annotation, thereby significantly improving the detection accuracy.
[0054] (2) The present invention overcomes the difficulty of detecting mitotic cells in complex pathological images by traditional image processing methods through deep learning, and can accurately locate mitotic cells in the presence of noise and complex backgrounds. Experimental results show that the model has achieved significant improvements in both sensitivity and precision, especially in the performance of cell details and variations, and can effectively reduce the interference of human factors and improve the stability of detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a flowchart of the method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0057] Embodiment 1:
[0058] Please refer to Figure 1 As shown, the method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network includes the following steps:
[0059] S1. Data collection and annotation: Collect a dataset containing breast cancer pathological images and annotate the mitotic cells in the images. The content of the annotation includes position information and category information.
[0060] S2. Data preprocessing: Preprocess the dataset to obtain preprocessed images.
[0061] S3. Network construction and training: Build a mitosis detection model based on a deep convolutional neural network, combine transfer learning and custom modules to optimize the network structure, and use training data for model training. Use the cross-entropy loss function to detect the mitotic cell category and combine the IoU loss function to locate the cell position.
[0062] S4. Post-processing of detection results: Post-process the output results of the detection model, including non-maximum suppression and false positive filtering, to improve the accuracy and robustness of the detection.
[0063] S5. Model validation and evaluation: Validate the model performance on an independent test dataset and data from different sources, and evaluate the model's performance in terms of sensitivity, specificity, and accuracy.
[0064] S6. System deployment: Apply the trained model to the clinical environment, combine visualization tools to display the detection results, and support real-time inference and continuous optimization.
[0065] Specifically, the data preprocessing steps include cropping, scaling, normalizing, and data augmentation of the images to improve the image quality and data diversity. The data augmentation operations include random rotation, horizontal flipping, color jittering, and synthetic data generation based on the generative adversarial network GAN.
[0066] Specifically, the normalization normalizes the image pixel values to a unified range (usually [0,1] or [-1,1]) to accelerate model training and reduce the difference in data distribution. The formula for normalization is:
[0067]
[0068] where I_raw: the original image pixel value;
[0069] μ: the mean of the image;
[0070] σ: the standard deviation of the image;
[0071] I_norm: the normalized image pixel value;
[0072] The purpose of normalization is to make the mean of the image data 0 and the variance 1, thereby accelerating the convergence speed of model training, improving the stability of training, and reducing overfitting.
[0073] Specifically, the data augmentation increases the diversity of the training dataset and improves the generalization ability of the model by applying operations such as rotation, flipping, and cropping to the images:
[0074]
[0075] (x, y): Original coordinates of the image;
[0076] (x′, y′): New coordinates after rotation;
[0077] θ: Rotation angle;
[0078] The rotation and flipping augmentation operations increase the diversity of the training samples, help the model learn cell features from different angles and directions, thereby improving the robustness and generalization ability of the model, and preventing overfitting.
[0079] Specifically, the mitosis detection model uses a deep convolutional neural network to extract spatial features from the input image for the mitosis detection model to learn local features (such as edges, shapes, etc.) of mitotic cells and improve the detection accuracy. The expression of the deep convolutional neural network is as follows:
[0080]
[0081] where x: Input image or feature map;
[0082] w: Convolution kernel, filter;
[0083] y: Output feature map;
[0084] (i, j): Coordinates of the output feature map;
[0085] m, n: Coordinates of the convolution kernel.
[0086] Specifically, the post-processing step of the detection results uses the non-maximum suppression algorithm to merge the detection boxes and filters the false positive detection results by a rule-based or simple classifier method.
[0087] Specifically, the non-maximum suppression is used to process multiple overlapping bounding boxes and only keep the box with the highest score, thereby eliminating redundant detection results. The expression of the non-maximum suppression is as follows:
[0088]
[0089] If the intersection over union IoU of two bounding boxes is greater than the set threshold of 0.5, the box with the lower score is suppressed;
[0090] A: Current detection box;
[0091] B: Other detection boxes overlapping with the current detection box;
[0092] IoU: Intersection over Union of two bounding boxes;
[0093] IoU threshold for determining whether to perform suppression;
[0094] NMS reduces the interference of multiple overlapping boxes, retains the most representative detection results, improves the detection accuracy, and avoids the model from repeatedly detecting the same mitotic cell.
[0095] Specifically, the cross - entropy loss function is used for classification tasks. In the classification of mitotic cells, it measures the difference between the predicted class and the true class. The expression of the cross - entropy loss function is:
[0096]
[0097] where C: number of classes;
[0098] yi: true label;
[0099] Probability value predicted by the model;
[0100] L: loss function value;
[0101] The cross - entropy loss function can effectively penalize the mispredicted classes, thus guiding the model to optimize the parameters and improve the classification accuracy. In the breast cancer cell detection task, it helps the model learn how to correctly classify mitotic cells.
[0102] Specifically, the model performance is evaluated by the mean average precision mAP, and its expression is:
[0103]
[0104] where N: number of classes;
[0105] APi: average precision of the i - th class;
[0106] mAP comprehensively considers the detection accuracies of multiple classes and is a key indicator for measuring the performance of the mitosis detection system, which can reflect the overall performance of the model on multiple detection classes.
[0107] As can be seen from the above, by using a deep convolutional neural network, the system can automatically extract complex features from breast cancer pathology images, such as the edges, shapes, textures, and color distributions of cells. These deep features enable the model to identify mitotic cells in subtle changes that are difficult to capture by traditional manual annotation, thus significantly improving the detection accuracy.
[0108] The system overcomes the difficulties in detecting mitotic cells in complex pathological images by traditional image processing methods through deep learning, and can accurately locate mitotic cells in the presence of noise and complex backgrounds. Experimental results show that the model has achieved significant improvements in both sensitivity and accuracy, especially in the performance under cell details and variations, and can effectively reduce the interference of human factors and improve the stability of detection results.
[0109] The automated system can provide fast and accurate results, reducing the workload of pathologists and avoiding errors and inconsistencies in manual annotation. At the same time, the model can process a large number of images, overcoming the limitations brought by human factors.
[0110] By combining transfer learning and convolutional neural networks, the model training and inference processes have significantly improved in speed. The time for data processing and mitotic cell detection has been significantly shortened, meeting the requirements for high efficiency in a clinical environment.
[0111] The system can analyze each image and output detection results within seconds, greatly improving the work efficiency of pathologists. Especially in high-throughput screening, it can quickly provide auxiliary decision-making for pathologists, saving a large amount of manual time.
[0112] Based on the detection of breast cancer pathological images, the system can be extended to the detection of pathological images of other types of cancers, with broad application prospects. Since the model is based on deep learning, it can continue to be optimized and adapted to new pathological features with the introduction of more data.
[0113] Example 2:
[0114] This example is based on a deep convolutional neural network (DCNN) for detecting mitotic cells in breast cancer pathological images, aiming to improve the accuracy and robustness of mitotic cell detection using deep learning algorithms.
[0115] 1. Data collection and annotation
[0116] In this example, breast cancer pathological image datasets from two different hospitals were selected for training and testing. Each image has a resolution of 1024×1024 pixels, and the positions and quantities of mitotic cells in the images were annotated.
[0117] Dataset 1: From Hospital A, a total of 500 images, with approximately 10,000 mitotic cells annotated.
[0118] Dataset 2: From Hospital B, a total of 400 images, with approximately 8,000 mitotic cells annotated.
[0119] The annotation form includes: the position of each cell (bounding box) and its classification information (mitotic cells vs non-mitotic cells).
[0120] 2. Data preprocessing
[0121] Data preprocessing includes image normalization, data augmentation, and cropping operations to improve the training effect and generalization ability of the model.
[0122] Image normalization: Convert the pixel values of the image from [0, 255] to the range [0, 1], and use the following formula for normalization:
[0123]
[0124] where μ = 128 and σ = 64 are the mean and standard deviation of the training images.
[0125] Data augmentation: Apply the following augmentation operations to each image:
[0126] Random rotation: The angle range is [0°, 90°].[[]]END]]
[0127] Random flipping: Horizontal flipping and vertical flipping.
[0128] Random cropping: Crop 80% of the image area.
[0129] Random brightness adjustment: The brightness change range is [0.8, 1.2].[[]]END]]
[0130] The goal of these augmentation operations is to increase the diversity of the data, prevent overfitting, and ensure that the model has strong recognition ability for mitotic cells under various deformations and angles.
[0131] 3. Network construction and training
[0132] In this embodiment, a deep convolutional neural network (DCNN) based on transfer learning is used. ResNet-50 is selected as the base model, and a custom module is added on it to optimize the detection of mitotic cells.
[0133] Network architecture:
[0134] Use the pre-trained ResNet-50 network for feature extraction.
[0135] Add 3 convolutional layers and 2 fully connected layers behind it to further extract the detailed features of the cells.
[0136] Add a multi-scale feature extraction module with convolutional kernel sizes of [3×3, 5×5, 7×7] to adapt to mitotic cells of different sizes.
[0137] Training parameters:
[0138] Batch Size: 32.
[0139] Learning Rate: 0.001.
[0140] Optimization algorithm: Adam optimizer.
[0141] Loss function: Cross-entropy loss function.
[0142] Epochs: 50.
[0143] Training effect:
[0144] During the training process, the loss function began to decline steadily after 30 epochs, indicating that the model was gradually converging.
[0145] After 50 epochs, the losses of both the training set and the validation set were close to 0.15, indicating that the model already had good performance.
[0146] 4. Post-processing of detection results
[0147] After the model outputs the results, non-maximum suppression (NMS) is used to suppress the redundant bounding boxes to reduce duplicate detections and improve the detection accuracy.
[0148] NMS parameters:
[0149] IoU threshold: 0.5.
[0150] Score threshold for detection boxes: 0.7 (only the boxes with scores greater than this threshold will be retained).
[0151] Post-processing effect:
[0152] NMS successfully removed 15%-20% of the redundant detection boxes, reduced the false positives, and improved the detection accuracy.
[0153] 5. Model verification and evaluation
[0154] In this embodiment, the model was verified on an independent test data set, and the evaluation metric used was the mean average precision (mAP).
[0155] Test data set: 100 images were selected from each of Hospital A and Hospital B for testing.
[0156] Evaluation results:
[0157] Mean average precision (mAP): 0.85.
[0158] Verification effect:
[0159] The model shows high performance on datasets from different sources, with high sensitivity and the ability to accurately detect most mitotic cells.
[0160] The accuracy is good, with few false detections and missed detections, and the model has strong clinical adaptability.
[0161] 6. System Deployment
[0162] Finally, the trained model was integrated into a pathological image detection platform. This platform has the following functions:
[0163] Real-time inference: After uploading a pathological image, the system can detect mitotic cells in real time and display the results.
[0164] Visualization tool: It provides a heat map to show the positions and distributions of mitotic cells, helping pathologists quickly locate suspicious areas.
[0165] Feedback mechanism: Pathologists can provide feedback on the detection results of the model to further optimize the model.
[0166] Platform deployment effect:
[0167] In a clinical environment, the model can complete the detection of a pathological image within 5 seconds, significantly improving the work efficiency of pathologists.
[0168] By combining with the feedback from pathologists, the system can be continuously optimized to further improve the accuracy.
[0169] As can be seen from the above, in this embodiment, through data collection, preprocessing, training and optimization of the deep convolutional neural network, a detection system for mitotic cells in breast cancer pathological images was successfully constructed. After verification, the model performs excellently on different datasets, accurately detecting mitotic cells, and verifying its high efficiency in actual clinical applications through NMS post-processing and evaluation metrics.
[0170] Example Three:
[0171] This embodiment introduces a lung cancer pathological image cell classification and tumor detection system based on a deep convolutional neural network (DCNN), aiming to assist pathologists in the early diagnosis of lung cancer through computer vision technology. This system uses a deep learning model for cell classification and tumor detection in lung cancer pathological images, and improves the diagnostic accuracy and reduces misdiagnosis and missed diagnosis in actual applications.
[0172] 1. Data Collection and Annotation
[0173] In this embodiment, a lung cancer pathological image dataset from multiple hospitals, including pathological section images of benign and malignant tumors, was selected. The resolution of each image is 1024×1024 pixels, and detailed annotation was carried out.
[0174] Dataset 1: From Hospital A, consisting of 600 images, labeled with different types of cells including normal cells, cancer cells, benign tumor cells, and malignant tumor cells.
[0175] Dataset 2: From Hospital B, consisting of 400 images, covering the pathological features of lung cancer at different stages.
[0176] The annotation content includes the location, type (cancer cell / benign cell) of each cell, and its histological category.
[0177] 2. Data Preprocessing
[0178] Data preprocessing includes image normalization and data augmentation to improve the robustness and generalization ability of the network.
[0179] Image Normalization: Normalize the pixel values of each image to the range [0, 1] using the following formula:
[0180]
[0181] where μ = 128 and σ = 64 are the mean and standard deviation of the image.
[0182] Data Augmentation: Adopt the following augmentation techniques to increase the diversity of training data:
[0183] Random Rotation: The rotation angle range is [0°, 360°].
[0184] Random Cropping: Crop different regions of the image while retaining the core region of the image.
[0185] Random Brightness Adjustment: Adjust the brightness range to [0.7, 1.3].
[0186] Random Flipping: Include horizontal and vertical flipping.
[0187] The purpose of augmenting the data is to enable the model to handle different image backgrounds, lesion degrees, and shooting angles, and improve the robustness in actual diagnosis.
[0188] 3. Network Construction and Training
[0189] In this embodiment, a deep convolutional neural network (DCNN) based on VGG-16 is used for cell classification and tumor detection. VGG-16 is a classic convolutional neural network architecture that can extract high-level features in images after multiple layers of convolution and pooling operations.
[0190] Network Architecture:
[0191] Use the convolutional layer and pooling layer of the VGG-16 network to extract low-level features of the image.
[0192] The cell type classification is further performed through a fully connected layer.
[0193] A tumor detection module is added to detect the tumor region and generate bounding boxes.
[0194] Training parameters:
[0195] Batch size: 64.
[0196] Learning rate: 0.001, and a learning rate decay strategy is adopted.
[0197] Optimizer: Adam optimizer.
[0198] Loss function: Multiclass cross-entropy loss function.
[0199] Epochs: 50.
[0200] Training effect:
[0201] After 50 training epochs, the training loss of the model gradually decreases, and the validation loss tends to be stable, indicating that the model can fit the training data well.
[0202] An early stopping strategy is used during training to avoid overfitting.
[0203] 4. Post-processing of detection results
[0204] After the model outputs the detection results, the non-maximum suppression (NMS) algorithm is used to suppress the redundant bounding boxes to reduce the redundant detection boxes and improve the detection accuracy.
[0205] NMS parameters:
[0206] IoU threshold: 0.5.
[0207] Score threshold: 0.75.
[0208] Post-processing effect:
[0209] NMS removes approximately 20% of the redundant boxes, significantly reduces false detections, and improves the detection accuracy.
[0210] The accuracy of the bounding boxes is improved, and the detection of the tumor region is more accurate.
[0211] 5. Model validation and evaluation
[0212] In this embodiment, the model is validated on an independent test dataset, and the evaluation metric is the mean average precision (mAP).
[0213] Test dataset: 200 images from Hospital A and Hospital B were selected for testing, including different types of cells and tumor regions.
[0214] Evaluation results:
[0215] Mean Average Precision (mAP): 0.82.
[0216] Verification effect:
[0217] The model can maintain high detection accuracy on different data sources, especially in the detection of tumor regions, and shows excellent performance.
[0218] It has high sensitivity, ensuring that most cancer cells can be accurately identified and reducing the missed diagnosis rate.
[0219] 6. System deployment
[0220] The trained deep convolutional neural network model has been integrated into an intelligent pathological image analysis platform, which has the following functions:
[0221] Real-time inference: After uploading a lung cancer pathological image, the system can complete cell classification and tumor detection within 5 seconds.
[0222] Visualization tool: It provides a heat map to show the distribution of cancer cells and the location of tumor regions, assisting pathologists to quickly locate abnormal regions.
[0223] Feedback mechanism: Pathologists can annotate and modify the detection results to further optimize the system.
[0224] Platform deployment effect:
[0225] The application of this system in the clinical environment has improved the diagnostic efficiency. Doctors can quickly obtain auxiliary diagnostic results and shorten the diagnostic time.
[0226] The system can provide real-time feedback to help pathologists verify the detection results and improve the accuracy and credibility.
[0227] As can be seen from the above, this embodiment demonstrates a lung cancer pathological image cell classification and tumor detection system based on a deep convolutional neural network, which can efficiently and accurately identify cell types in lung cancer pathological images and detect tumor regions. Through techniques such as data augmentation and non-maximum suppression, the system can process pathological images with high noise and complex backgrounds and has strong robustness. After model verification, the system has excellent performance on different datasets and can provide important support for the early diagnosis of lung cancer.
[0228] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.
[0229] In the drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved, and other structures can refer to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other.
[0230] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network, characterized in that, It includes the following steps: S1. Data collection and annotation: Collect a dataset containing breast cancer pathological images, and annotate the mitotic cells in the images. The content of the annotation includes position information and category information. S2. Data preprocessing: Preprocess the dataset to obtain preprocessed images. S3. Network construction and training: Based on a deep convolutional neural network, construct a mitosis detection model. Combine transfer learning and custom modules to optimize the network structure, and use training data to train the model. Use the cross-entropy loss function to detect the mitotic cell category, and combine the IoU loss function to locate the cell position. S4. Post-processing of detection results: Post-process the output results of the detection model, including non-maximum suppression and false positive filtering, to improve the accuracy and robustness of the detection. S5. Model validation and evaluation: Validate the model performance on an independent test dataset and data from different sources, and evaluate the model's performance in terms of sensitivity, specificity, and accuracy. S6. System deployment: Apply the trained model to the clinical environment, combine visualization tools to display the detection results, and support real-time inference and continuous optimization.
2. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 1, wherein The data preprocessing step includes cropping, scaling, normalizing, and data augmentation of the images. The data augmentation operations include random rotation, horizontal flipping, color jittering, and synthetic data generation based on the generative adversarial network GAN.
3. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 2, characterized in that The normalization normalizes the image pixel values to a unified range. The formula for the normalization is: where Iraw: the original image pixel value; μ: the mean of the image; σ: the standard deviation of the image; Inorm: the normalized image pixel value.
4. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 2, wherein The data augmentation increases the diversity of the training dataset by applying rotation, flipping, and cropping operations to the images, and improves the generalization ability of the model. (x,y): the original coordinates of the image; (x′,y′): the new coordinates after rotation; θ: the rotation angle.
5. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 1, wherein The mitosis detection model uses a deep convolutional neural network to extract the spatial features in the input image, which are used for the mitosis detection model to learn the local features of mitotic cells. The expression of the deep convolutional neural network is: where x: the input image or feature map; w: the convolutional kernel, filter; y: the output feature map; (i,j): the coordinates of the output feature map; m,n: the coordinates of the convolutional kernel.
6. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 1, characterized in that The post-processing step of the detection results uses the non-maximum suppression algorithm to merge the detection boxes, and filters the false positive detection results by a rule-based or simple classifier method.
7. The mitosis detection method for breast cancer pathological images based on a deep convolutional neural network according to claim 1, characterized in that The non-maximum suppression is used to process multiple overlapping bounding boxes, and only the box with the highest score is retained, thereby eliminating redundant detection results. The expression of the non-maximum suppression is: If the intersection over union (IoU) of two bounding boxes is greater than the set threshold of 0.5, then suppress the box with the lower score. A: the current detection box; B: other detection boxes overlapping with the current detection box; IoU: the intersection over union of two bounding boxes; The IoU threshold used to determine whether to perform suppression.
8. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 1, characterized in that, The cross-entropy loss function is used for classification tasks. In the classification of mitotic cells, it measures the difference between the predicted class and the true class. The expression of the cross-entropy loss function is as follows: where C: the number of classes; yi: the true label; Probability value predicted by the model; L: the value of the loss function.
9. The method for detecting mitosis in breast cancer pathological images based on a deep convolutional neural network according to claim 1, wherein The performance of the model is evaluated using the mean average precision (mAP), and its expression is as follows: where N: the number of classes; APi: the average precision of the i-th class.