Brain tumor image segmentation method and system

Abstract

The invention discloses a brain tumor image segmentation method and system, and relates to the technical field of brain tumor image segmentation. The segmentation efficiency of the brain tumor image is improved. Comprising the following steps: data preprocessing: carrying out denoising, image registration and intensity standardization processing on an original image; feature extraction: combining morphological features, texture features and histogram features; neural network model construction: constructing and optimizing a neural network model based on deep learning for a brain tumor segmentation task; and model training and verification: carrying out model training by using the preprocessed MRI image data and corresponding labels, and evaluating model performance. In the preprocessing stage, a series of noise removal technologies, such as mean filtering and median filtering, are adopted to eliminate noise interference in the image, and the MRI image of the patient is aligned with a standard brain template space by using an image registration method, so that the difference between the images is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of brain tumor image segmentation, and in particular, to a brain tumor image segmentation method and system. Background Art

[0002] Brain tumors, as a type of serious neurological diseases, early diagnosis and treatment thereof are crucial for the survival rate and quality of life of patients. With the rapid development of medical imaging technology, especially the wide application of magnetic resonance imaging (MRI) technology, doctors can obtain high-resolution, multi-modal brain image data, which provides an important basis for the accurate diagnosis of brain tumors. However, how to accurately segment the tumor region from a large amount of MRI image data has always been a major challenge in the field of medical image analysis. Traditional brain tumor image segmentation methods mainly rely on manual segmentation or rule-based semi-automatic segmentation. These methods are not only time-consuming and laborious, but also the accuracy and consistency of the segmentation results are often affected by the operator's experience and skills. Therefore, developing an efficient and accurate automated brain tumor image segmentation method has important clinical significance and application value. In recent years, the rise of deep learning technology has provided a new solution for brain tumor image segmentation. Neural network models based on deep learning, such as convolutional neural networks (CNNs), generative adversarial networks (GANs), etc., can achieve automatic segmentation of brain tumor images by learning feature representations in a large amount of labeled data. Especially the U-Net architecture has shown excellent performance in medical image segmentation tasks due to its unique encoder-decoder structure and skip connections.

[0003] Directly applying deep learning models to brain tumor image segmentation still faces many challenges. First of all, brain tumor images usually have a high degree of complexity and diversity, including different tumor types, sizes, shapes, and positions, etc., which makes it difficult for the model to learn robust feature representations. Secondly, MRI image data often has problems such as noise, artifacts, and image registration, which will all affect the accuracy of the segmentation results.

[0004] To address these challenges, researchers have proposed various improvement strategies. On the one hand, through data preprocessing techniques, such as noise removal, image registration, and intensity normalization, etc., the image quality is improved and the error of subsequent processing is reduced. On the other hand, multi-modal feature extraction methods, such as morphological features, texture features, and histogram features, etc., are combined to comprehensively describe the geometric complexity and irregularity of tumors, providing richer information for the model. The present invention proposes a new brain tumor image segmentation method and system, aiming to further improve the accuracy and efficiency of segmentation. By introducing regularization strategies, optimizing data augmentation methods, designing efficient loss functions and training algorithms and other innovative means, more accurate and reliable segmentation results are provided, thus assisting in the early diagnosis and treatment of brain tumors. Summary of the Invention

[0005] The object of the present invention is to solve the deficiencies existing in the prior art, and a method and system for brain tumor image segmentation are proposed.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for brain tumor image segmentation includes the following steps:

[0008] Data preprocessing, performing denoising, image registration, and intensity normalization on the original image;

[0009] Feature extraction, combining morphological features, texture features, and histogram features;

[0010] Neural network model construction, constructing and optimizing a neural network model based on deep learning for brain tumor segmentation tasks;

[0011] Model training and verification, using the preprocessed MRI image data and corresponding labels for model training and evaluating the model performance.

[0012] Preferably: In the data preprocessing step, the denoising process includes hardware methods and software methods. The hardware method eliminates noise by improving the performance of the MR scanner, and the software methods include mean, median, Wiener, and diffusion filters.

[0013] Furthermore: In the feature extraction step, the morphological features, texture features, and histogram features are extracted through specific algorithms and calculation methods, including but not limited to the box dimension method, Hurst exponent method, gray-level co-occurrence matrix, LBP method, and pixel value distribution calculation.

[0014] Furthermore: In the neural network model construction step, the model architecture selection includes but not limited to U-Net, GAN, and the model is optimized. The optimization strategies include attention mechanism, multi-scale feature fusion, and cross-level feature fusion.

[0015] As a preferred solution of the present invention: In the model training and verification step, the cross-validation method is used to evaluate the model performance, and the Dice loss function is used as the optimization target.

[0016] As a further solution of the present invention: In the tumor segmentation and post-processing step, the specific methods of denoising, hole filling, and image enhancement processing are adjusted and optimized according to actual needs.

[0017] A brain tumor image segmentation system includes a data preprocessing module, a feature extraction module, a neural network model training module, and a post-processing module. Each module works together to achieve the segmentation and processing of brain tumor images.

[0018] Based on the foregoing solution: The data preprocessing module includes a data input interface, a preprocessing configuration interface, a preprocessing execution interface, and a result output interface, which are used to receive MRI image data, set preprocessing parameters, execute the preprocessing process, and output the preprocessed image data.

[0019] Based on the foregoing solution: The feature extraction module uses edge detection algorithms to extract tumor boundaries, calculates area, perimeter, roundness morphological features, calculates gray-level co-occurrence matrices to extract contrast, energy, homogeneity texture features, and statistically counts the number of gray-level pixels to extract gray mean, variance, range histogram features.

[0020] Based on the foregoing solution: The post-processing module uses appropriate denoising methods to remove noise in the segmentation results, detects holes in the segmentation results and selects appropriate filling methods for filling, and performs contrast stretching and sharpening enhancement processing on the filled segmentation results.

[0021] The beneficial effects of the present invention are as follows:

[0022] 1. A brain tumor image segmentation method and system adopts a series of noise removal techniques, such as mean filtering, median filtering, etc., in the preprocessing stage to eliminate noise interference in the image, and uses image registration methods to align the patient's MRI image with the standard brain template space, thereby reducing the differences between images. In the feature extraction stage, the method adopts a multi-modal fusion-based technique to effectively integrate MRI image information of different modalities, thereby improving the segmentation accuracy and Dice score.

[0023] 2. A brain tumor image segmentation method and system utilizes the information of multi-modal MRI image data, combines morphological features, texture features, and histogram features in the feature extraction stage, and comprehensively and accurately extracts the feature information of the tumor region. Morphological features can capture the shape and size of the tumor, texture features reflect the internal texture changes of the tumor, and histogram features provide the distribution of gray values in the tumor region.

[0024] 3. A brain tumor image segmentation method and system introduces an attention mechanism to enable the model to pay more attention to the important features of the tumor region, thereby improving the segmentation accuracy. Secondly, a multi-scale feature fusion strategy is adopted to integrate feature information of different scales to capture the detailed information of the tumor at different scales. In addition, data augmentation processing such as cropping, rotation, and mirroring is also performed on the model to increase the diversity of training samples.

[0025] 4. A method and system for brain tumor image segmentation adopt methods such as cross-validation to evaluate the performance of the model. Cross-validation divides the dataset into multiple subsets and alternately uses them as the training set and the validation set, so as to more comprehensively evaluate the stability and reliability of the model, reduce the overfitting phenomenon of the model on a specific dataset, and improve the generalization ability of the model in practical applications.

[0026] 5. A method and system for brain tumor image segmentation. Denoising processing can remove noise interference in the segmentation result, making the tumor area clearer. Filling hole processing can solve the hole problem in the segmentation result, making the tumor area more complete. Through enhancement processing such as contrast stretching and sharpening, the visual effect of the segmentation result is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a schematic flowchart of a method for brain tumor image segmentation proposed by the present invention;

[0028] Figure 2 is a system block diagram of a system for brain tumor image segmentation proposed by the present invention;

[0029] Figure 3 is a data preprocessing flowchart of a method for brain tumor image segmentation proposed by the present invention.

[0030] Figure 4 is a schematic flowchart of a method for brain tumor image segmentation (regularization strategy) of a method for brain tumor image segmentation proposed by the present invention;

[0031] Figure 5 is a system flowchart of a system for brain tumor image segmentation proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The technical solutions of this patent will be further described in detail below in conjunction with the specific embodiments.

[0033] The embodiments of this patent are described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain this patent and should not be construed as a limitation of this patent.

[0034] Embodiment 1:

[0035] A method for brain tumor image segmentation, as Figure 1 shown, includes the following steps:

[0036] S1: Data preprocessing

[0037] Obtain brain image data from the patient's MRI scan, preprocess the original image, remove noise, perform image registration and intensity normalization in the original image to improve the image quality and reduce errors in subsequent processing;

[0038] S2: Feature extraction

[0039] Combine morphological features, texture features and histogram features to extract the features of the tumor region, combine multi-modal MRI image data, and according to the features of the tumor in different modalities;

[0040] S3: Neural network model construction

[0041] Construct a neural network model based on deep learning to adapt to the brain tumor segmentation task, improve and optimize the model, introduce strategies such as attention mechanism and multi-scale feature fusion. During the model training process, use data augmentation techniques and loss functions to optimize the performance of the model;

[0042] S4: Model training and validation

[0043] Use the preprocessed MRI image data and corresponding labels for model training. During the training process, use methods such as cross-validation to evaluate the performance of the model to ensure the stability and reliability of the model;

[0044] S5: Tumor segmentation and post-processing

[0045] Apply the trained model to new MRI image data for tumor segmentation, and perform noise removal and hole filling on the segmentation results to improve the quality and readability of the segmentation results;

[0046] In step S1, during the preprocessing of the data, when removing noise from the original image, the hardware method eliminates noise by improving the performance of the MR scanner, and the software methods include mean, median, Wiener, and diffusion filters;

[0047] Among them, during the process of eliminating noise using the hardware method, it is necessary to improve the signal receiving coil, and its improvement process includes the following steps:

[0048] 1: Optimize the coil design:

[0049] Design a suitable coil shape, size and number of turns, and use high-performance wire materials such as copper or silver to reduce resistance and loss. Consider the shielding and grounding design of the coil to reduce external interference;

[0050] 2: Enhance the coil performance:

[0051] By increasing the sensitivity of the coil, it can capture weak signals more effectively; optimize the impedance matching of the coil to ensure that signals are not lost due to impedance mismatch during transmission; introduce resonance techniques, such as inductance-capacitance resonance arrays, to enhance the signal reception ability of the coil at specific frequencies;

[0052] At the same time, it is also necessary to improve the signal-to-noise ratio of the MR logic sequence, and its operation process includes the following steps:

[0053] 1: Optimize scanning parameters:

[0054] According to the imaging target and requirements, select appropriate scanning parameters, such as repetition time (TR), echo time (TE), flip angle, etc.; by adjusting these parameters, optimize the signal intensity and contrast, thereby improving the signal-to-noise ratio;

[0055] 2: Adopt imaging techniques:

[0056] Introduce parallel imaging techniques, such as SENSE (Sensitivity Encoding), to accelerate the imaging process and reduce noise; use high-resolution imaging techniques, such as 3D volume acquisition, to improve the details and clarity of the image;

[0057] 3: Signal processing techniques:

[0058] Apply filtering and noise reduction algorithms, such as wavelet transform, non-local mean filtering, etc., to reduce the noise in the image; utilize image reconstruction techniques, such as iterative reconstruction algorithms, to improve the resolution and signal-to-noise ratio of the image;

[0059] 4: Hardware upgrade:

[0060] Upgrade the hardware components of the magnetic resonance equipment, such as gradient coils, radio frequency coils, etc., to improve the performance and stability of the equipment; introduce superconducting technology or cryogenic cooling systems to reduce the noise level of the equipment;

[0061] 5: Data enhancement and post-processing:

[0062] In the data acquisition stage, adopt data enhancement strategies, such as multiple acquisitions, acquisitions from different angles, etc., to improve the reliability and signal-to-noise ratio of the data; in the data post-processing stage, apply techniques such as image registration and fusion to further improve the quality and signal-to-noise ratio of the image;

[0063] When performing image registration, the registration process includes the following steps:

[0064] Use feature point detection algorithms (such as SIFT, SURF, ORB, etc.) to detect feature points in the image;

[0065] Generate descriptors for each feature point to describe the surrounding image information;

[0066] Use feature matching algorithms (such as BFMatcher, FLANN, etc.) to find pairs of matching feature points between two images;

[0067] Calculate the transformation matrix (such as affine transformation matrix, perspective transformation matrix, etc.) based on the pairs of matching feature points;

[0068] Apply the transformation matrix to transform one image to a position aligned with the other image;

[0069] Define an objective function and then use an optimization algorithm to find the optimal spatial transformation parameters; when performing intensity normalization, register the images in the preprocessing process to the standard brain template space MNI to unify the coordinate spaces of all images, and the algorithm used for normalization is a non-rigid registration algorithm including affine transformation and non-linear transformation;

[0070] Combining morphological features, texture features, and histogram features in the step S2, extract the features of the tumor region, and its specific operation process includes the following steps:

[0071] S21: According to the characteristics and analysis requirements of the tumor image, select a suitable fractal dimension calculation method. Commonly used methods include the box dimension method, Hurst exponent method, and box counting method. Apply the selected calculation method to the preprocessed tumor image to extract its fractal dimension. According to the selected calculation method, obtain the fractal dimension value of the tumor image, and compare the calculated fractal dimension with the fractal dimension range of known tumor types to assist in diagnosis and classification;

[0072] S22: When extracting texture features, first calculate the gray-level co-occurrence situation of adjacent pixels in the image to obtain a gray-level co-occurrence matrix, and extract texture features from this matrix, such as contrast, energy, and homogeneity. Statistically analyze the number and distribution of consecutive pixels with the same gray value in the image, and extract texture features such as run length and gray-level distribution from this matrix. Calculate the gray-level difference between each pixel in the image and its neighboring pixels, and statistically analyze the distribution of these differences;

[0073] S23: When extracting histogram features, statistically analyze the number of pixels at each gray level in the image to obtain a gray-level histogram, and extract features from this histogram, such as gray-level mean, gray-level variance, and gray-level range;

[0074] During data preparation and preprocessing for morphological feature extraction, deep learning models (such as CNN, RNN, or Transformer) are used to train and optimize the preprocessed images, and high-dimensional features including local and global features such as edges, textures, and shapes are extracted from them; traditional image processing techniques are combined to extract geometric, texture, and other features, such as fractal dimension and Hurst exponent, to comprehensively describe the geometric complexity and irregularity of tumors; in the feature fusion stage, high-dimensional features and traditional features are fused by methods such as concatenation, weighted summation, or PCA, and dimensionality reduction techniques such as PCA and t-SNE are used as needed to reduce the feature dimension; the fused features are used to construct a classification or regression model, and the model performance is evaluated through training and validation, and the model is optimized according to the validation results;

[0075] In the process of constructing the neural network model in step S3, the model architecture selects U-Net and GAN. When optimizing the performance of the model, the optimization process includes the following steps:

[0076] Customize the U-Net structure and enhance it with specific data. The operation process includes the following steps:

[0077] Design and implement a custom U-Net model, including:

[0078] Encoder: Usually composed of multiple convolutional blocks, each block contains a convolutional layer, an activation function (such as ReLU), and a pooling layer (such as max pooling);

[0079] Decoder: Similar to the encoder structure, but in the opposite direction;

[0080] Skip connection: Directly transfer the feature maps of the encoder to the corresponding layers of the decoder to retain detailed information;

[0081] Collect and prepare the dataset, including images and corresponding labels, and ensure that the format and size of the dataset match the input requirements of the U-Net model;

[0082] Implement data augmentation strategies according to requirements, including:

[0083] Rotation: Randomly rotate the images and labels;

[0084] Flip: Horizontally or vertically flip the images and labels;

[0085] Scaling: Randomly scale the images and labels;

[0086] Translation: Randomly move the images and labels on the image plane;

[0087] Cropping: Randomly crop a region from the image;

[0088] Add noise: Add Gaussian noise or other types of noise to the image;

[0089] Create a data generator to dynamically provide augmented data;

[0090] Use the data generator to train the U-Net model, ensure that data augmentation is correctly applied in the training loop, and monitor the performance and loss of the model;

[0091] Evaluate the model performance on the validation set or test set, calculate evaluation metrics such as IoU (Intersection over Union), precision, recall, etc.;

[0092] Adjust the U-Net architecture, data augmentation strategy, and training parameters according to the evaluation results. The U-Net architecture adopts an attention mechanism, residual connection modules, and a topological feature fusion strategy;

[0093] The attention mechanism weights the information of different parts by learning a set of weights, thereby highlighting the response of important information. When processing sequence data or images, the attention mechanism dynamically adjusts the degree of attention to different positions or regions;

[0094] The topological feature fusion strategy includes the following steps:

[0095] 1: Multi-scale feature fusion: Extract and fuse features at different scales to capture information at different scales; for example, methods such as pyramid pooling and pyramid convolution;

[0096] 2: Cross-level feature fusion: Extract and fuse features at different network levels to combine high-level semantic information in the deep layer and low-level detail information in the shallow layer, such as methods like -Net, FPN (Feature Pyramid Network), etc.;

[0097] 3: Attention mechanism-assisted feature fusion: Introduce the attention mechanism to weight the importance of different features to achieve more effective feature fusion; for example, in the image caption generation task, the attention mechanism helps the model dynamically select key regions in the image for feature fusion;

[0098] In step S4, model training and validation are carried out. The specific training and validation processes include the following steps:

[0099] 1: Collect MRI image data, perform preprocessing operations such as scaling, cropping, and grayscaling on the MRI images, and divide the dataset into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to adjust the model parameters and select the best model, and the test set is used to evaluate the final performance of the model;

[0100] 2: Select an appropriate deep learning model architecture according to the task requirements to ensure that the model can process 3D MRI image data and has sufficient capacity to capture tumor features. Train the model using the preprocessed training set data, and set hyperparameters such as appropriate batch size, learning rate, and optimizer.

[0101] 3: Divide the training set into K subsets (usually K is taken as 5 or 10). For each cross - validation iteration, use K - 1 subsets as training data and the remaining one subset as validation data.

[0102] The tumor segmentation and post - processing process in step S5 includes the following steps:

[0103] Perform binarization on the denoised segmentation result to separate the tumor region from the background. Use the connected component labeling algorithm to detect holes in the tumor region; according to the size, shape, and position of the holes, select an appropriate filling method.

[0104] Adjust the parameters of the filling method to ensure that the holes are effectively filled while avoiding over - filling that may cause the expansion of the tumor region.

[0105] Visually inspect the filled segmentation result to ensure that the holes are effectively filled while maintaining the integrity and accuracy of the tumor region. Use quantitative metrics to evaluate the filling effect.

[0106] Perform further enhancement processing on the filled segmentation result, such as contrast stretching, sharpening, etc.

[0107] Use visualization tools (such as MATLAB, matplotlib in Python, etc.) to display the segmentation result in the form of an image.

[0108] Example 2:

[0109] A brain tumor image segmentation system includes a data pre - processing module, a feature extraction module, a neural network model training module, and a post - processing module.

[0110] The design of the data preprocessing module follows the principles of low coupling, high cohesion and information hiding, ensuring that the interaction between modules is clear and easy to maintain; the module provides a data input interface to receive the patient's MRI scan brain image data, and supports multiple data format inputs; the preprocessing configuration interface allows external setting of preprocessing parameters, such as noise removal intensity, image registration algorithm selection, etc.; the preprocessing execution interface triggers the preprocessing process, performs noise removal, image registration and intensity standardization, etc.; the result output interface outputs the preprocessed image data, providing a unified data format and coordinate space; in terms of input, the module receives the original MRI image data and preprocessing parameters; in terms of output, it provides preprocessed MRI image data and preprocessing reports; in terms of core function implementation, the module uses a combination of hardware filters and software filtering algorithms to remove noise, and realizes image registration by extracting feature points and matching these feature points, ensuring that the coordinate space of all images is unified, and using intensity standardization processing;

[0111] The feature extraction module is the core component in the image processing system, responsible for extracting key information from the image data for subsequent processing; its input is the generated image, which is the original data or pre-processed data; the output includes the morphological features, texture features and histogram features of the image, which are extracted by edge detection, shape analysis and other algorithms, gray level co-occurrence matrix, LBP and other methods, as well as pixel value distribution calculation;

[0112] In the collaborative workflow between modules, the feature extraction module works closely with the image preprocessing module and the classification and recognition module; the image preprocessing module performs a series of operations on the original image, such as denoising, enhancement and standardization, aiming to optimize image quality, reduce noise interference, and unify image features, laying a solid foundation for subsequent feature extraction; the preprocessed image is passed to the feature extraction module, which uses advanced algorithms such as morphological analysis, texture analysis and histogram calculation to deeply explore key information in the image, such as edges, shapes, texture patterns and color distribution features. The features are then input into the classification and recognition module, which uses machine learning or deep learning technology to analyze these features to achieve accurate classification or recognition of the image, thereby completing the entire image processing process;

[0113] In terms of system integration, the feature extraction module demonstrates high flexibility and adaptability. It relies on the read and write functions of the file system in the local environment and exchanges data with other modules such as the preprocessing module and the classification and recognition module. This way of transferring data through files is not only simple and feasible but also can effectively ensure data security. In a distributed or cloud computing environment, the feature extraction module becomes an independent Web service and opens a REST API interface for other modules to make remote calls through HTTP requests. The preprocessing module sends the processed image data to the feature extraction module in the form of an HTTP POST request. The latter receives the request, extracts features, and then encapsulates the results into formats such as JSON or XML and returns them to the caller through an HTTP response.

[0114] The neural network model training module is mainly responsible for constructing deep learning-based neural network models such as U-Net and GAN, and training and optimizing these models. During the training process, various data augmentation techniques such as cropping, rotation, and mirroring are used to increase the diversity of training samples. At the same time, in order to measure the similarity between the segmentation result and the ground truth label, the module uses the Dice loss function as the optimization objective.

[0115] The post-processing module plays a role in further optimizing and improving the segmentation result in the system. It first uses the trained neural network model to perform tumor segmentation on new MRI image data. Then, it performs denoising processing on the segmentation result to remove existing noise interference. The module will detect the hole areas in the segmentation result and select appropriate filling methods for filling to ensure the integrity and accuracy of the segmentation result. It performs enhancement processing such as contrast stretching and sharpening on the filled segmentation result to improve the visual effect and readability of the image. With these post-processing steps, the system can generate more accurate and reliable tumor segmentation results, providing strong assistance for doctors' diagnosis and treatment.

[0116] The system architecture design covers multiple optimization levels from hardware acceleration to intelligent modules. The hardware acceleration layer selects high-performance GPUs / TPUs, combines TensorFlow / PyTorch frameworks and CUDA / cuDNN libraries to maximize hardware performance for accelerating deep learning model inference. The model optimization layer adopts lightweight network architectures and applies model pruning and quantization techniques to reduce model size and computational complexity. The streaming processing layer realizes streaming data reading and processing, introduces asynchronous I / O mechanisms to improve processing efficiency and control memory occupancy. The power efficiency optimization layer dynamically adjusts resource allocation, uses Docker containerization technology, and combines intelligent power management to reduce energy consumption. The algorithm optimization layer optimizes data augmentation methods, designs efficient loss functions and training algorithms to improve model performance. The intelligent module optimization layer introduces an online learning mechanism, utilizes pre-trained model migration to achieve automatic fusion and feature extraction of MRI images, develops interactive segmentation tools for intelligent tumor prediction, and provides visualization tools and report generation functions, comprehensively enhancing the system's intelligence level and processing capabilities.

[0117] Example 3

[0118] A method for brain tumor image segmentation. To further improve the segmentation effect of brain tumor images, in this example, a regularization strategy is introduced based on the U-Net and GAN architectures to handle the brain tumor image segmentation task. The specific operation process includes the following steps:

[0119] 1: Weight decay:

[0120] Add a penalty term proportional to the square of the model weights to the loss function;

[0121] Set the weight decay coefficient through an optimizer (such as Adam, SGD, etc.);

[0122] 2: Dropout:

[0123] Add Dropout layers after certain layers of the model (such as after fully connected layers, convolutional layers);

[0124] Set an appropriate dropout rate, usually between 0.2 and 0.5;

[0125] 3: Batch Normalization:

[0126] Add Batch Normalization layers after convolutional layers or fully connected layers;

[0127] Restore the data representation ability by learning scaling factors and offsets;

[0128] 4: Data augmentation regularization:

[0129] During the data augmentation process, more transformations and perturbations are introduced to increase the diversity of the data.

[0130] Example 4

[0131] A brain tumor image segmentation method. This example demonstrates the complete working principle from data segmentation to result visualization, and the process includes:

[0132] Removing image noise, image registration, and intensity normalization to reduce errors in subsequent processing and improve image quality;

[0133] Enter the feature extraction stage. This stage combines morphological features, texture features, and histogram features to comprehensively extract the feature information of the tumor region; through methods such as fractal dimension calculation, gray-level co-occurrence matrix analysis, and gray-level histogram statistics, the internal characteristics of the tumor image are deeply explored to provide rich feature inputs for the subsequent neural network model training;

[0134] In the model construction stage, a neural network architecture based on deep learning, such as U-Net or GAN, is selected and customized for the brain tumor segmentation task; by introducing optimization means such as attention mechanism, residual connection, and topological feature fusion strategy, the segmentation accuracy and generalization ability of the model are improved; during the model training process, data augmentation techniques (such as rotation, flipping, scaling, etc.) are used to increase the diversity of training samples, and at the same time, the Dice loss function is used to measure the similarity between the segmentation result and the true label, and a suitable optimizer (such as Adam, SGD) is selected to update the model parameters; the performance of the model is evaluated through methods such as cross-validation to ensure the stability and reliability of the model;

[0135] In the tumor segmentation stage, the trained neural network model is applied to new MRI image data for accurate segmentation of the tumor region; the segmentation result is denoised to remove existing noise interference; then, the connected component labeling algorithm is used to detect the holes in the tumor region, and a suitable filling method is selected according to the size, shape, and position of the holes for effective filling; to ensure the filling effect, the parameters of the filling method need to be adjusted and visual inspection is carried out; finally, enhancement processing such as contrast stretching and sharpening is performed on the filled segmentation result to improve the readability and visual effect of the image; the segmentation result is displayed in the form of an image using a visualization tool, and the complete process from data segmentation to result visualization is clearly presented;

[0136] This application is described based on MRI scans, but its core steps are applicable to other types of imaging data, such as CT or PET; when applied to CT or PET, the following adjustments need to be considered in specific implementation:

[0137] For CT images, metal artifacts need to be removed;

[0138] PET images need to address the issues of low resolution and noise;

[0139] CT images need to provide structural information, while PET images provide functional information;

[0140] When optimizing for specific tumor types (such as pediatric and adult brain tumors), for brain tumors of different age groups, the following aspects of optimization need to be considered:

[0141] Pediatric brain tumor images contain more motion artifacts and require more complex denoising and registration algorithms;

[0142] Adult brain tumor images contain more degenerative lesions and require more refined preprocessing steps to distinguish tumors from normal aging;

[0143] Brain tumors of different age groups have different morphological and texture characteristics;

[0144] The optimization of feature extraction methods needs to be carried out according to the characteristics of specific tumor types;

[0145] Regarding future improvement directions:

[0146] Combining multimodal data (MRI + CT) to improve segmentation accuracy. Multimodal data fusion makes full use of the complementary information provided by different imaging modalities, thereby improving segmentation accuracy. Use multimodal fusion techniques in deep learning, such as feature-level fusion or decision-level fusion.

[0147] As described above, this is a preferred specific embodiment of the present invention. The protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made by those skilled in the art within the technical scope disclosed by the present invention in combination with the prior art or common knowledge, within the spirit and principles of the present invention, shall be covered by the protection scope of the present invention.

Claims

1. A brain tumor image segmentation method, characterized in that: The following steps are involved: Data preprocessing: denoising, image registration and intensity normalization of the original image; Feature extraction, combining morphological features, texture features and histogram features; Neural network model construction: build and optimize deep learning-based neural network models for brain tumor segmentation tasks; Model training and validation,The preprocessed MRI image data and corresponding labels are used to train the model and evaluate the model performance.

2. A brain tumor image segmentation method according to claim 1, characterized in that: In the data preprocessing step, denoising processing includes hardware methods and software methods. The hardware method eliminates noise by improving the performance of the MR scanner, and the software method includes mean, median, Wiener, and diffusion filters.

3. The brain tumor image segmentation method according to claim 2, characterized in that: In the feature extraction step, morphological features, texture features and histogram features are extracted through specific algorithms and calculation methods, including but not limited to box dimension method, Hurst index method, gray level co-occurrence matrix, LBP method and pixel value distribution calculation.

4. The brain tumor image segmentation method according to claim 3, characterized in that: In the neural network model construction step, model architecture selection includes but is not limited to U-Net and GAN, and the model is optimized. The optimization strategies include attention mechanism, multi-scale feature fusion, and cross-level feature fusion.

5. The brain tumor image segmentation method according to claim 4, characterized in that: In the model training and validation steps, the cross-validation method is used to evaluate the model performance, and the Dice loss function is used as the optimization objective.

6. The brain tumor image segmentation method according to claim 5, characterized in that: In the tumor segmentation and post-processing steps, the specific methods of denoising, hole filling and image enhancement are adjusted and optimized according to actual needs.

7. A brain tumor image segmentation system, using the segmentation method according to any one of claims 1 to 6 for segmentation, characterized in that: It includes a data preprocessing module, a feature extraction module, a neural network model training module and a post-processing module, and each module works together to achieve the segmentation and processing of brain tumor images.

8. The brain tumor image segmentation system according to claim 7, characterized in that: The data preprocessing module includes a data input interface, a preprocessing configuration interface, a preprocessing execution interface and a result output interface, which are used to receive MRI image data, set preprocessing parameters, execute preprocessing procedures and output preprocessed image data.

9. The brain tumor image segmentation system according to claim 8, characterized in that: The feature extraction module can extract morphological features, texture features and histogram features from the preprocessed image and output these features for subsequent processing and analysis.

10. The brain tumor image segmentation system according to claim 9, characterized in that: The neural network model training module can build and optimize the neural network model based on deep learning, and the post-processing module can denoise, fill holes and enhance the image of the segmentation results to generate more accurate and reliable tumor segmentation results.

Images