Image segmentation identification method and system based on coupling model
By adopting a coupled model-based image segmentation recognition method in breast cancer detection, combining gamma distribution processing and U-Net model, the problem of excessive work burden and error analysis in early breast cancer detection in the prior art is solved, and high-accurate breast cancer image segmentation and classification are achieved.
Patent Information
- Application Number
- CN202411713460.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-14
AI Technical Summary
Prior Art In early breast cancer detection, doctors need to review medical images for a long time, resulting in excessive workload and the possibility of analysis errors, and there are errors in manual testing.
The image segmentation recognition method based on the coupled model is adopted, and the breast ultrasound images are segmented and classified by splicing the YOLOv10 model and the coupled convolution model, and combined with gamma distribution processing and U-Net model, the accuracy of image recognition is improved.
It significantly improves the accuracy of breast cancer image segmentation and classification, reduces the error of manual detection, and improves the work efficiency of doctors.
Smart Images
Figure CN119625311A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image recognition, and more particularly to an image segmentation and recognition method and system based on a coupling model. Background Art
[0002] Cancer remains a major challenge to global health, killing more people than any other disease. Breast cancer is a disease caused by the uncontrolled proliferation of breast cells, leading to the formation of tumors. The disease poses a major threat to the lives of women around the world. The latest statistics show that the incidence of breast cancer has surpassed lung cancer to become the most common type of cancer, and the health threat to women is particularly significant. By 2023, breast cancer, colorectal cancer and lung cancer are predicted to become the most commonly diagnosed cancers in women in the United States. The incidence of breast cancer may reach alarming levels, with some cases as high as 31%. Despite the rising trend in breast cancer incidence, the mortality rate from breast cancer has dropped by 43% since its peak in 1989. This change is mainly attributed to mammography screening, increased public awareness and advances in treatment technology. Although the overall trend is positive, the mortality rate of breast cancer has slightly rebounded in recent years, highlighting the importance of continued screening and early intervention. The prognosis of breast cancer is closely related to the stage at diagnosis. The cure rate in the middle stage is between 50% and 80%, while the cure rate in the early stage can be as high as 90%. However, the treatment of advanced breast cancer remains extremely challenging, which further highlights the importance of early detection and intervention.
[0003] Breast tissue can be classified as benign, invasive, or normal based on its characteristics. Benign tissue appears as non-malignant growth with clear borders and irregular structures that pose no significant threat to health. If non-invasive breast cancer is detected early, it stays within certain structures (such as ducts or lobules) and does not spread to other tissues, which means a better prognosis. Early detection of breast cancer is expected to lead to the possibility of cure, while invasive breast cancer refers to the spread of cancer cells from the breast to other organs. The most effective early detection and treatment methods currently use a combination of ultrasound, thermal imaging, and mammography. However, clinicians face the problem of excessive workload as they need to review medical images every few seconds during long working hours, which can lead to errors and inaccuracies in analysis. Advances in artificial intelligence (AI) have revolutionized several fields, including machine vision, image recognition, and speech recognition. These technologies have great potential in medical image analysis, aiming to simplify doctors' work and improve the accuracy of diagnosis.
[0004] Therefore, how to propose an image segmentation and recognition method and system based on a coupling model to improve image recognition capabilities and perform breast cancer image segmentation and classification is an urgent problem that technicians in this field need to solve. Summary of the invention
[0005] In view of this, the present invention provides an image segmentation and recognition method and system based on a coupling model to improve image recognition capabilities and perform breast cancer ultrasound image segmentation and classification. In order to achieve the above purpose, the present invention adopts the following technical solutions:
[0006] An image segmentation and recognition method based on a coupling model, comprising:
[0007] Collect the data to be tested, preprocess the data to be tested, and train the YOLOv10 model with the preprocessed data;
[0008] Perform gamma distribution processing on the data to be tested, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature;
[0009] Combine the YOLOv10 model with the coupled convolutional model, and train and optimize the combined image segmentation and recognition model;
[0010] Image recognition is performed based on the trained and optimized image segmentation and recognition model.
[0011] Optionally, collecting the data to be tested includes: collecting a CBIS-DDSM data set and a MIAS data set from a Kaggle data repository.
[0012] Optionally, the preprocessing includes: amplifying the data, performing random scaling, brightness adjustment, random mirroring, random rotation and noise interference on the test data, parsing the annotated xml file, modifying the corresponding anchor frame position, ensuring that the labels in the augmented data are not offset, and then converting the xml file into a corresponding txt label file.
[0013] Optionally, the training of the YOLOv10 model using the preprocessed data includes:
[0014] Input the preprocessed data into the YOLO model with set hyperparameters for training;
[0015] Use the advanced classifier XGBoost to classify the bounding boxes and obtain the probability of defect type corresponding to each bounding box;
[0016] During the model training process, iterative training is performed, and the model is verified through the validation set to evaluate the model's accuracy and recall. The learning rate is then adjusted dynamically based on the validation results. The training process continues until the preset number of iterations is reached or the validation set performance no longer improves, and the network model training is completed.
[0017] Optionally, before performing gamma distribution processing on the data to be measured, the method further includes:
[0018] Convert DICOM images to PNG format using DICOM Converter;
[0019] Data resampling, data normalization, data enhancement, and data splitting were performed on the dataset;
[0020] The CLAHE method is used to optimize local contrast and enhance the distinguishability of small features in the image.
[0021] Artifacts were removed by Weiner filtering.
[0022] Optionally, the coupled convolution model includes:
[0023] A SPP layer is connected before the fully connected layer of the original MobileNetV2 model; the output of the modified MobileNetV2 model is used as the input of the densely connected convolutional network DenseNet; the densely connected convolutional network DenseNet passes the output of each layer to all subsequent layers, enhances the information flow through the skip connection mechanism, and finally outputs the segmented image result.
[0024] Optionally, after the modification, the MobileNetV2 model structure is: a sequentially connected convolutional layer, seven linear bottleneck layers, a convolutional layer, an average pooling layer, a convolutional layer, an SPP layer, and a fully connected layer.
[0025] Optionally, the splicing of the YOLOv10 model with the coupled convolutional model includes: connecting the YOLOv10 model and the coupled convolutional model using a picture cropping layer, the picture cropping layer performing numerical conversion with the input of the anchor frame, obtaining the coordinates of the anchor frame, and cropping the original picture to obtain a picture without environmental information; using a cropping layer between the spliced YOLOv10 model and the coupled convolutional model to connect the two models, and cropping the input picture according to the anchor frame output by the YOLOv10 model as the input of the coupled convolutional model.
[0026] Optionally, the training and optimization of the spliced image segmentation recognition model includes: using four indicators: true positive examples TP, true negative examples TN, false positive examples FP and false negative examples FN to measure performance indicators, and evaluating the performance of the image segmentation recognition model through accuracy, sensitivity, precision and F1 score performance indicators.
[0027] Optionally, an image segmentation and recognition system based on a coupling model includes:
[0028] Acquisition module: used to collect the data to be tested, pre-process the data to be tested, and train the YOLOv10 model with the pre-processed data;
[0029] Training module: used to process the test data with gamma distribution, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature;
[0030] Synthesis module: used to splice the YOLOv10 model with the coupled convolutional model, and train and optimize the spliced image segmentation and recognition model;
[0031] Recognition module: used to perform image recognition based on the trained and optimized image segmentation recognition model.
[0032] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses an image segmentation and recognition method and system based on a coupling model, which has the following beneficial effects:
[0033] The present invention proposes an image segmentation and recognition method based on a coupled model, comprising: collecting data to be tested, preprocessing the data to be tested, and training a YOLOv10 model through the preprocessed data; performing gamma distribution processing on the data to be tested, coupling a U-Net model to output a first feature, and training a coupled convolution model through the first feature; splicing the YOLOv10 model with the coupled convolution model, and training and optimizing the spliced image segmentation and recognition model; and performing image recognition according to the trained and optimized image segmentation and recognition model. The present invention applies a variety of data preprocessing technologies, such as CLAHE technology and Weiner filtering method, to eliminate unnecessary artifacts and significantly improve image quality. First, the input breast X-ray image is segmented, and then the segmented breast X-ray image is feature extracted, and finally the breast X-ray image is classified as benign or malignant. This method uses an enhanced U-Net architecture combined with gamma distribution for image segmentation; uses modified MobileNetV2 and DenseNet, splices the YOLOv10 model with the coupled convolutional model, trains and optimizes the spliced image segmentation and recognition model to achieve image classification, improves the detection and classification of breast lesions, and avoids errors in manual detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0035] Figure 1 A schematic flow chart of an image segmentation and recognition method based on a coupling model provided by the present invention.
[0036] Figure 2 This is a structural framework diagram of an image segmentation and recognition system based on a coupling model provided by the present invention. DETAILED DESCRIPTION
[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0038] The embodiment of the present invention discloses an image segmentation and recognition method based on a coupling model, such as Figure 1 As shown, including:
[0039] Collect the data to be tested, preprocess the data to be tested, and train the YOLOv10 model with the preprocessed data;
[0040] Perform gamma distribution processing on the data to be tested, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature;
[0041] Combine the YOLOv10 model with the coupled convolutional model, and train and optimize the combined image segmentation and recognition model;
[0042] Image recognition is performed based on the trained and optimized image segmentation and recognition model.
[0043] Furthermore, the data collection includes: collecting the CBIS-DDSM data set and the MIAS data set from the Kaggle data repository. The data set is labeled using the labelimg tool to obtain a standard file in XML format, and then each XML file is converted into a txt file; the converted txt file content includes: category, anchor box center point coordinate x, anchor box center point coordinate y, anchor box width and anchor box height; the actual value x of the anchor box position recorded in the XML file is min 、x max ,y min and max , get the center point coordinates and anchor box size, normalize the actual values according to the length and width of the image, save the normalized result to a txt file, and store it in proportional form in the interval [0, 1].
[0044] Furthermore, the preprocessing includes: amplifying the data, performing random scaling, brightness adjustment, random mirroring, random rotation and noise interference on the test data, parsing the annotated xml file, modifying the corresponding anchor frame position, ensuring that the labels in the augmented data are not offset, and then converting the xml file into the corresponding txt label file.
[0045] Furthermore, the training of the YOLOv10 model using the preprocessed data includes:
[0046] Input the preprocessed data into the YOLO model with set hyperparameters for training;
[0047] Use the advanced classifier XGBoost to classify the bounding boxes and obtain the probability of defect type corresponding to each bounding box;
[0048] During the model training process, iterative training is performed, and the model is verified through the validation set to evaluate the model's accuracy and recall. The learning rate is then adjusted dynamically based on the validation results. The training process continues until the preset number of iterations is reached or the validation set performance no longer improves, and the network model training is completed.
[0049] In a specific implementation manner, the YOLO model with the hyperparameters set is an improved YOLOv10 model, and the improvement process includes: introducing an attention module EMA;
[0050] (1) Module design: The attention module EMA divides the input feature map into multiple subgroups along the channel direction; each subgroup is processed in a parallel subnetwork, including a 1×1 convolution branch and a 3×3 convolution branch;
[0051] (2) Run a 1×1 convolution branch and a 3×3 convolution branch, including:
[0052] 1) Parallel subnetwork: The 1×1 convolution branch is responsible for processing global information to recalibrate the channel weights in each parallel branch; the specific implementation steps are as follows:
[0053] 11) 1×1 convolution operation, the expression is:
[0054] F global =Conn 1×1 (F input );
[0055] In the formula, F global is the extracted global feature map, Conv 1×1 is a 1×1 convolution operation, F input is the input feature map;
[0056] 12) Global average pooling, the expression is:
[0057] G pool =Global Average Pooling(F global );
[0058] In the formula, G pool It is the feature after global average pooling, which represents the global feature of each channel;
[0059] 13) Full connection generates weights, the expression is:
[0060] W calibrated =Fully Connected(G pool );
[0061] Where W calibrated The weight for each channel is generated through a fully connected layer;
[0062] 14) Channel weighted calibration, the expression is:
[0063] F calibrated =F input ×W chanel ;
[0064] In the formula, F calibrated is the recalibrated feature map, which is calibrated by multiplying the channel weight with the input feature map; W channel The weight for each channel is generated through a fully connected layer;
[0065] 2) The 3×3 convolution branch further aggregates the output features of the two parallel branches through cross-dimensional interaction to capture the pairwise relationship at the pixel level. The specific implementation steps are as follows:
[0066] 21) 3×3 convolution operation, the expression is:
[0067] F 3×3 =Conv 3×3 (F inpu t);
[0068] In the formula, F 3×3 is the feature map after 3×3 convolution, Conv 3×3 It is a 3×3 convolution operation;
[0069] 22) Cross-dimensional interaction and feature aggregation, the expression is:
[0070] F aggregated =F 1×1 +F 3×3 ;
[0071] In the formula, F aggregated is the feature map after aggregation, F 1×1 It is the feature map after convolution;
[0072] 23) Capture the pairwise relationship at the pixel level, expressed as:
[0073]
[0074] In the formula, A pixel is the attention weight, which is generated by performing element-by-element dot product operation on the aggregated feature map; sigmoid() is the activation function, and T is the exponent;
[0075] 24) Apply the weight, the expression is:
[0076] F final =F aggregated ×A pixel ;
[0077] In the formula, F final For the final feature map, the pixel-level pairwise relationship is enhanced by applying attention weights to the aggregated feature map;
[0078] (3) Cross-space learning: The output feature maps of the two parallel sub-networks are fused together and the pairwise relationships at the pixel level are captured through matrix dot product operations. The 1×1 convolution branch is responsible for processing global information and recalibrating the channel weights in each parallel branch, while the 3×3 convolution branch further aggregates the output features of the two parallel branches through cross-dimensional interactions. Feature map fusion is achieved by element-by-element addition or concatenation. The correlation matrix is generated through matrix dot product operations to capture the pairwise relationships at the pixel level, and the attention weights are generated through activation functions and applied to the feature map to enhance pixel pairs with strong correlation and suppress pixel pairs with weak correlation.
[0079] In a specific embodiment, convolutional neural networks (CNNs), especially U-Net, have become an important tool for medical image segmentation. Probability distributions such as the gamma distribution excel in simplifying and reducing the variability of breast tissue features. The U-Net architecture has advantages in segmenting breast masses, demonstrating its ability to extract detailed spatial information from medical images. The encoder-decoder architecture of U-Net is able to collect and replicate contextual data, which is critical in accurately detecting breast tumors.
[0080] Using probability distribution to show the variation of image density, the effectiveness of U-Net in medical image segmentation has been demonstrated. The flexibility and accuracy of the gamma distribution make it a popular choice in medical image analysis. An architecture called Gamma-Net integrates the gamma distribution into the U-Net design to improve the discrimination of brain tumors and is beneficial for segmenting breast masses to more accurately represent tissue characteristics. The potential of combining U-Net with the gamma distribution to improve the accuracy of breast cancer segmentation for detection and diagnosis, especially between groups with different shapes and textures.
[0081] Gamma distribution is combined with U-Net model to achieve improved segmentation results. In the mathematical model integrated with Gamma distribution and U-Net model, pixel intensity is used to measure the amount of data in each region of the image. The variation of pixel intensity in different regions can reveal unique features of the image. The current demonstration model uses Gamma distribution to simulate the pixel brightness of each region of the image.
[0082] Furthermore, before performing gamma distribution processing on the data to be measured, the method further includes:
[0083] Convert DICOM images to PNG format using DICOM Converter;
[0084] Data resampling, data normalization, data enhancement, and data splitting were performed on the dataset;
[0085] The CLAHE method is used to optimize local contrast and enhance the distinguishability of small features in the image.
[0086] Artifacts were removed by Weiner filtering.
[0087] Furthermore, the coupled convolution model includes:
[0088] A SPP layer is connected before the fully connected layer of the original MobileNetV2 model; the output of the modified MobileNetV2 model is used as the input of the densely connected convolutional network DenseNet; the densely connected convolutional network DenseNet passes the output of each layer to all subsequent layers, enhances the information flow through the skip connection mechanism, and finally outputs the segmented image result.
[0089] In a specific implementation, the MobileNetV2 model is a lightweight convolutional neural network (CNN) architecture designed for embedded and mobile vision applications. Its lightweight design and high efficiency make it suitable for tasks with limited computing resources. Several advantages of MobileNetV2 in image classification include: its lightweight architecture enables efficient deployment on resource-constrained devices; in addition, its architecture is comparable to larger models without sacrificing accuracy. Furthermore, an SPP layer is connected before the fully connected layer of the original MobileNetV2 model, which is suitable for image inputs of different sizes;
[0090] DenseNet is an architecture called "Densely Connected Convolutional Network". Through its innovative dense connection mode, DenseNet performs well on multiple tasks. DenseNet overcomes the gradient vanishing problem in deep neural networks by passing the output of each layer to all subsequent layers. Compared with ResNet, DenseNet enhances the flow of information through its "skip connection" mechanism;
[0091] The gamma distribution is integrated with the U-Net model to segment the input breast X-ray images. The coupled convolutional model combines the modified MobileNetV2 and DenseNet to extract features for segmented images and classify the images as normal, benign, or malignant.
[0092] Furthermore, after the modification, the MobileNetV2 model structure is: a sequentially connected convolutional layer, seven linear bottleneck layers, a convolutional layer, an average pooling layer, a convolutional layer, an SPP layer, and a fully connected layer.
[0093] Furthermore, the splicing of the YOLOv10 model with the coupled convolutional model includes: connecting the YOLOv10 model with the coupled convolutional model using a picture cropping layer, the picture cropping layer performing numerical conversion with the input of the anchor frame, obtaining the coordinates of the anchor frame, and cropping the original picture to obtain a picture without environmental information; a cropping layer is used between the spliced YOLOv10 model and the coupled convolutional model to connect the two models, and the output of the YOLOv10 model is used as the input of the coupled convolutional model.
[0094] In a specific implementation, because the proportions of the cropped images may vary greatly, in order to meet the requirements of image inputs of different sizes, an SPP layer is connected before the fully connected layer of the original MobileNetV2 model. The SPP (spatial pyramid pooling) layer can handle the situation of different input data sizes well. Then, the modified MobileNetV2 is trained using the cropped image dataset, and the optimal model parameters are repeatedly optimized;
[0095] The training steps of MobileNetV2 are as follows: input the image into the model to obtain the calculation result, perform the cost function operation on the result and the label to obtain the cost that can represent the prediction error, and then use the optimization algorithm to optimize the model parameters according to the cost so that the cost obtained in the next round of calculation is reduced, and the training is completed after multiple rounds;
[0096] Concatenate the trained YOLOv10 model and the modified MobileNetV2 model, use a crop layer to connect the two models, and use the output of the YOLOv10 model as the input of the MobileNetV2 model;
[0097] The structure of the spliced model is data input, data processing, YOLOv10, image cropping layer, the part before the MobileNetV2 fully connected layer, SPP layer, MobileNetV2 fully connected layer, and output.
[0098] The output of the modified MobileNetV2 model is used as the input of the densely connected convolutional network DenseNet; the densely connected convolutional network DenseNet passes the output of each layer to all subsequent layers, enhances the flow of information through the skip connection mechanism, and finally outputs the segmented image result.
[0099] In a specific implementation, the output result of the coupled convolutional model and the label are subjected to a cost function operation to obtain a cost that can represent the prediction error, and then the model parameters are optimized using an optimization algorithm based on the cost so that the cost obtained in the next round of calculation is reduced, and the training is completed after multiple rounds.
[0100] Furthermore, the training and optimization of the spliced image segmentation recognition model includes: using four indicators: true positive examples TP, true negative examples TN, false positive examples FP and false negative examples FN to measure performance indicators, and evaluating the performance of the image segmentation recognition model through accuracy, sensitivity, precision and F1 score performance indicators.
[0101] In a specific embodiment, an image segmentation and recognition system based on a coupling model, such as Figure 2 As shown, including:
[0102] Acquisition module: used to collect the data to be tested, pre-process the data to be tested, and train the YOLOv10 model with the pre-processed data;
[0103] Training module: used to process the test data with gamma distribution, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature;
[0104] Synthesis module: used to splice the YOLOv10 model with the coupled convolutional model, and train and optimize the spliced image segmentation and recognition model;
[0105] Recognition module: used to perform image recognition based on the trained and optimized image segmentation recognition model.
[0106] In a specific implementation, it also includes constructing an experimental environment:
[0107] Each experiment was conducted on a device with the following specifications: 16GB RAM memory module, 2.80GHz Intel(R) Core(TM) i7-7700 CPU, and NVIDIA GTX 1050Ti graphics card. Training and testing the segmentation model took 46 minutes and 38 minutes on the CBIS-DDSM dataset, respectively. This example uses the Python programming language, specifically the Image Data Generator module in the Keras framework, to generate a dataset of mammographic images, and data preparation is performed in a real-time environment.
[0108] The model parameter configuration is set as follows: the split ratio of the test data set is 1:4 to the split ratio of the training data set, the activation function is the ReLU activation function, the batch size is 8, the number of training rounds is 50, and the learning rate is 0.001. The super performance indicators of the proposed model include: training time, test time, usage threshold, DSC value (Dice similarity coefficient), IOU value (intersection over union ratio), recall rate, precision, global accuracy and AUC ROC value.
[0109] The model was evaluated using 1569 training images and 522 validation images. The training phase started by training each of the proposed architectures, setting the learning rate to 0.001 and the momentum constant to 0.6. L2 regularization was applied during training and the optimization was performed in 50 epochs.
[0110] Experimental results show that the proposed method outperforms existing models in multiple performance indicators. A SoftMax layer is added to the dataset in each cycle. The proposed model exhibits fast processing time and efficient segmentation ability, with a global accuracy of 97.62% and an AUC-ROC value of 98.32%.
[0111] The present invention proposes an image segmentation and recognition method based on a coupled model to segment different parts in a breast X-ray image. The model uses a GU-Net neural network to perform image segmentation, which helps to capture global information features from the data set, thereby enhancing the image analysis capabilities of the model. In addition, the architecture combines a synthetic model of DenseNet and a modified MobileNetV2 for feature extraction, splices the YOLOv10 model with the coupled convolutional model, trains and optimizes the spliced image segmentation and recognition model, and realizes image classification. The present invention focuses on outlining the contours and boundaries of breast lesions in breast X-ray images, and the resulting image contains multiple components, covering dynamic breast imaging data. Experimental results show that the method has a significant effect in accurately identifying and analyzing MRI images of breast lesions, providing valuable insights for medical workers and facilitating diagnostic evaluation.
[0112] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0113] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An image segmentation and recognition method based on a coupling model, characterized in that: include: Collect the data to be tested, preprocess the data to be tested, and train the YOLOv10 model with the preprocessed data; Perform gamma distribution processing on the data to be tested, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature; Combine the YOLOv10 model with the coupled convolutional model, and train and optimize the combined image segmentation and recognition model; Image recognition is performed based on the trained and optimized image segmentation and recognition model.
2. The image segmentation and recognition method based on the coupling model according to claim 1 is characterized in that: The collecting of the test data includes: collecting the CBIS-DDSM data set and the MIAS data set from the Kaggle data repository.
3. The image segmentation and recognition method based on the coupling model according to claim 1 is characterized in that: The preprocessing includes: amplifying the data, performing random scaling, brightness adjustment, random mirroring, random rotation and noise interference on the data to be tested, parsing the annotated xml file, modifying the corresponding anchor frame position, ensuring that the labels in the augmented data are not offset, and then converting the xml file into the corresponding txt label file.
4. The image segmentation and recognition method based on the coupling model according to claim 1 is characterized in that: The training of the YOLOv10 model using the preprocessed data includes: Input the preprocessed data into the YOLO model with set hyperparameters for training; Use the advanced classifier XGBoost to classify the bounding boxes and obtain the probability of defect type corresponding to each bounding box; During the model training process, iterative training is performed, and the model is verified through the validation set to evaluate the model's accuracy and recall. The learning rate is then adjusted dynamically based on the validation results. The training process continues until the preset number of iterations is reached or the validation set performance no longer improves, and the network model training is completed.
5. The image segmentation and recognition method based on the coupling model according to claim 1, characterized in that: Before the gamma distribution processing is performed on the data to be measured, the method further includes: Convert DICOM images to PNG format using DICOM Converter; Data resampling, data normalization, data enhancement, and data splitting were performed on the dataset; The CLAHE method is used to optimize local contrast and enhance the distinguishability of small features in the image. Artifacts were removed by Weiner filtering.
6. The image segmentation and recognition method based on the coupling model according to claim 1, characterized in that: The coupled convolution model includes: A SPP layer is connected before the fully connected layer of the original MobileNetV2 model; the output of the modified MobileNetV2 model is used as the input of the densely connected convolutional network DenseNet; the densely connected convolutional network DenseNet passes the output of each layer to all subsequent layers, enhances the information flow through the skip connection mechanism, and finally outputs the segmented image result.
7. The image segmentation and recognition method based on the coupling model according to claim 6 is characterized in that: After the modification, the MobileNetV2 model structure is: sequentially connected convolutional layers, seven linear bottleneck layers, convolutional layers, average pooling layers, convolutional layers, SPP layers, and fully connected layers.
8. The image segmentation and recognition method based on the coupling model according to claim 1, characterized in that: The splicing of the YOLOv10 model and the coupled convolution model includes: connecting the YOLOv10 model and the coupled convolution model using a picture cropping layer, the picture cropping layer performs numerical conversion on the input of the anchor frame, obtains the coordinates of the anchor frame, and crops the original picture to obtain a picture without environmental information; a cropping layer is used between the spliced YOLOv10 model and the coupled convolution model to connect the two models, and the input picture is cropped according to the anchor frame output by the YOLOv10 model as the input of the coupled convolution model.
9. The image segmentation and recognition method based on the coupling model according to claim 1, characterized in that: The training and optimization of the spliced image segmentation recognition model includes: using four indicators: true positive examples TP, true negative examples TN, false positive examples FP and false negative examples FN to measure performance indicators, and evaluating the performance of the image segmentation recognition model through accuracy, sensitivity, precision and F1 score performance indicators.
10. An image segmentation and recognition system based on a coupling model, characterized in that: include: Acquisition module: used to collect the data to be tested, pre-process the data to be tested, and train the YOLOv10 model with the pre-processed data; Training module: used to process the test data with gamma distribution, and couple the U-Net model to output the first feature, and train the coupled convolution model through the first feature; Synthesis module: used to splice the YOLOv10 model with the coupled convolutional model, and train and optimize the spliced image segmentation and recognition model; Recognition module: used to perform image recognition based on the trained and optimized image segmentation recognition model.