Medical image segmentation precision improvement method and system

Abstract

The invention discloses a medical image segmentation precision improvement method and system. The method comprises the steps that an input medical image is preprocessed, denoising is conducted firstly, parameters are adjusted according to the local noise condition through an adaptive filtering algorithm, and then the image contrast is enhanced through histogram equalization; image texture and shape features are extracted, texture features are extracted by using a gray level co-occurrence matrix, edges are extracted by using a Canny edge detection algorithm, and shape features are calculated; constructing a segmentation model, selecting a U-Net network, training with a large amount of labeled data, and optimizing with a cross entropy loss function and stochastic gradient descent; a segmentation result is optimized, denoising is performed by setting an area threshold, holes are filled by morphological closed operation, multi-model fusion is adopted, and a weight weighted average fusion result is determined according to performance indexes of each model. The system is used for executing the method. Through multi-step processing, the medical image segmentation precision is effectively improved, the adaptability to different patient images is enhanced, a more accurate and reliable segmentation result is provided for medical diagnosis and treatment, and the method has important clinical application value.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for improving the accuracy of medical image segmentation. The invention aims to improve the accuracy of segmentation of target regions (organs, lesions) in medical images (X-ray, CT, MRI), providing more reliable and accurate image segmentation results for precise disease diagnosis, treatment planning, and subsequent medical research. Background Technology

[0002] (I) Current Technological Status

[0003] Medical image segmentation is a crucial technology in medical diagnosis and treatment. By segmenting medical images, doctors can clearly identify key information such as lesion areas and organ boundaries, thereby more accurately assessing the condition and developing personalized treatment plans. Currently, common medical image segmentation methods include threshold-based segmentation, edge detection-based segmentation, region-based segmentation, and deep learning-based segmentation methods.

[0004] (II) Problems with Existing Technologies

[0005] Noise interference: Medical images are often affected by various noises during acquisition, such as electronic noise and motion artifacts. These noises blur the boundaries of the image, making it difficult for segmentation algorithms to accurately identify the boundaries of the target region, resulting in deviations in the segmentation results.

[0006] Complexity of tissue structure: The internal tissue structure of the human body is complex and diverse, and the boundaries between different tissues may be blurred, with overlaps and occlusions occurring between them. For example, in lung CT images, the blood vessels, bronchi, and other structures of the lungs are intertwined with the lung tissue, increasing the difficulty of segmentation.

[0007] Individual differences: Factors such as the patient's physical condition, age, and gender can lead to significant variations in medical images. For example, the bone density of an elderly person may differ from that of a younger person. These individual differences need to be considered when segmenting bone images. However, existing segmentation methods often lack adaptability to individual differences, making it difficult to guarantee segmentation accuracy across images from different patients.

[0008] Limited segmentation accuracy: Existing segmentation methods often fall short of clinical requirements when processing complex medical images. For example, threshold-based segmentation methods are sensitive to noise and easily misclassify noise as target regions; edge detection-based segmentation methods struggle to accurately extract edge information when edges are blurred; and region-based segmentation methods perform poorly when processing target regions with complex textures and shapes. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for improving the accuracy of medical image segmentation, so as to solve the problems of low accuracy, poor adaptability to noise and individual differences in existing medical image segmentation methods, improve the accuracy and reliability of medical image segmentation, and provide better image segmentation results for medical diagnosis and treatment.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for improving the accuracy of medical image segmentation includes the following steps:

[0012] Image preprocessing steps:

[0013] The input medical image is subjected to denoising and image enhancement processing. The denoising processing adopts an adaptive filtering algorithm, which automatically adjusts the filtering parameters according to the noise statistical characteristics of the local area of ​​the image, removing image noise while retaining image edge and detail information. The image enhancement processing adopts histogram equalization technology, which improves the contrast of the image by adjusting the gray-level distribution of the image, making the difference between the target area and the background area more significant.

[0014] Feature extraction steps:

[0015] Texture and shape features are extracted from the preprocessed image. The texture features are calculated using the gray-level co-occurrence matrix and specifically include energy, entropy, and contrast parameters, which are used to describe the texture characteristics of different regions in the image. The shape features are calculated based on the edge information extracted by the edge detection algorithm and include area, perimeter, and roundness parameters, which are used to characterize the morphological features of the target region.

[0016] Segmentation model training steps:

[0017] A training dataset and a test dataset containing a large number of medical images and their corresponding annotations were constructed. A convolutional neural network was selected as the segmentation model. The segmentation model was trained using the training dataset. The model parameters were continuously adjusted through the backpropagation algorithm so that the model could learn the mapping relationship between image features and segmentation results. During the training process, the cross-entropy loss function was used to measure the difference between the model's prediction results and the real annotations. The stochastic gradient descent optimization algorithm was used to minimize the loss function until the model's performance on the test dataset reached the preset requirements.

[0018] Segmentation result optimization steps:

[0019] The initial segmentation results output by the segmentation model are post-processed and multi-model fusion is performed. The post-processing includes removing small noise regions and filling holes to improve the accuracy and completeness of the segmentation results. The multi-model fusion adopts a weighted average method to merge the segmentation results of multiple different segmentation models, comprehensively consider the advantages of each model, and further improve the segmentation accuracy.

[0020] Furthermore, in the denoising process, corresponding adaptive filtering algorithms are used for different types of noise; when Gaussian noise is present in the image, Gaussian filtering is used; when salt-and-pepper noise is present in the image, median filtering is used.

[0021] Furthermore, in the edge detection algorithm, the Canny edge detection algorithm is selected to extract the edge information of the image. This algorithm accurately extracts the edges in the image through Gaussian filtering to smooth the image, calculating the image gradient magnitude and direction, non-maximum suppression, and double threshold detection steps.

[0022] Furthermore, in the segmentation model training step, the segmentation model is a U-Net network with an encoder-decoder structure. The encoder part extracts high-level features of the image through convolution and pooling operations, and the decoder part restores the spatial resolution of the image through upsampling and convolution operations. It also fuses the features of the corresponding layers of the encoder and decoder through skip connections to retain more image detail information.

[0023] Furthermore, in the multi-model fusion step, the multiple different segmentation models are models trained on different network structures, different hyperparameter settings, or different training datasets; during weighted average fusion, the weights are determined based on the performance metrics of each model on the test dataset, with models having higher performance metrics having higher weights.

[0024] A medical image segmentation accuracy improvement system includes:

[0025] Image preprocessing module: used to perform denoising and image enhancement processing on the input medical images; the denoising module is implemented using an adaptive filtering algorithm, which automatically adjusts the filtering parameters according to the noise characteristics of local areas of the image; the image enhancement module is implemented using histogram equalization technology to improve the contrast of the image;

[0026] Feature extraction module: used to extract texture features and shape features from the preprocessed image; the texture feature extraction module calculates the texture feature parameters of the image using the gray-level co-occurrence matrix; the shape feature extraction module calculates the shape feature parameters of the target region based on the edge information extracted by the edge detection algorithm;

[0027] Segmentation model training module: used to build training and test datasets, select and train segmentation models; the training dataset construction module is responsible for collecting a large amount of medical image data and annotating it by professional doctors; the model training module uses the training dataset to train the segmentation model, adjusts the model parameters through the backpropagation algorithm, and optimizes the model using the cross-entropy loss function and stochastic gradient descent optimization algorithm.

[0028] The segmentation result optimization module is used to post-process and fuse the preliminary segmentation results output by the CNN model. The post-processing module includes a sub-module for removing small noise regions and a sub-module for filling holes. The multi-model fusion module uses a weighted average method to fuse the segmentation results of multiple different CNN models.

[0029] Furthermore, in the image preprocessing module, the denoising module calls the Gaussian filtering function when Gaussian noise is detected, and calls the median filtering function when salt-and-pepper noise is detected.

[0030] Furthermore, in the feature extraction module, the shape feature extraction module uses the Canny edge detection algorithm to extract the edge information of the image, and calculates the shape feature parameters of the target region based on the edge information, including area, perimeter, and roundness.

[0031] Furthermore, in the segmentation model training module, the segmentation model is a U-Net network, which is implemented through a deep learning framework and has an encoder-decoder structure and a skip connection mechanism.

[0032] Furthermore, in the segmentation result optimization module, the multi-model fusion module determines the weights of each segmentation model based on the performance indicators of each segmentation model on the test dataset, performs a weighted average fusion of the segmentation results of multiple models, and outputs the final segmentation result.

[0033] The medical image segmentation accuracy improvement method and system of the present invention have the following beneficial effects:

[0034] Improving segmentation accuracy: Through a series of steps including image preprocessing, feature extraction, segmentation model training, and segmentation result optimization, this invention can effectively improve the accuracy of medical image segmentation, reduce segmentation errors, provide doctors with more accurate information on lesion areas, and help improve the accuracy of disease diagnosis.

[0035] Enhanced adaptability: By employing adaptive filtering algorithms and deep learning models, this invention is better adapted to different types of medical images with different noise levels and individual differences among patients, and can achieve high segmentation accuracy under various complex conditions.

[0036] Improved efficiency: Deep learning models have powerful learning capabilities, enabling them to process large amounts of medical image data in a short time. Furthermore, the post-processing and multi-model fusion operations in the segmentation result optimization module can improve segmentation efficiency while ensuring accuracy, thus meeting actual clinical needs. Attached Figure Description

[0037] Figure 1 This is a schematic diagram illustrating the implementation process of the method for improving the accuracy of medical image segmentation in this invention.

[0038] Figure 2 This is a schematic diagram of the system architecture of the medical image segmentation accuracy improvement system of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0040] The present invention proposes a method for improving the accuracy of medical image segmentation, comprising the following steps:

[0041] Image preprocessing

[0042] Denoising: An adaptive filtering algorithm is used to denoise the input medical images. This algorithm first performs local region analysis on the image, calculating the noise statistics of each local region, such as the noise variance. Then, it automatically adjusts the filtering parameters based on the noise statistics, using a higher filtering intensity for areas with more noise and a lower filtering intensity for areas with less noise. For example, when Gaussian noise is detected in the image, Gaussian filtering is used; when salt-and-pepper noise is detected, median filtering is used. In this way, noise is effectively removed while preserving as much edge and detail information as possible.

[0043] Image enhancement: Histogram equalization is used to enhance the denoised image. Histogram equalization expands the grayscale range of the image by adjusting the grayscale distribution, thereby improving the image contrast. After enhancement, the difference between the target area and the background area is more obvious, which facilitates subsequent segmentation operations.

[0044] Feature extraction

[0045] Texture feature extraction: The texture features of the image are extracted using the gray-level co-occurrence matrix (GLCM). The gray-level co-occurrence matrix is ​​a statistical method that describes the spatial dependence of gray levels in an image. By calculating the probability of different gray level pairs appearing in the image, various texture feature parameters can be obtained, such as energy, entropy, and contrast. These texture features can reflect the texture differences in different regions of the image and help to distinguish between target areas and background areas.

[0046] Shape feature extraction: The edge information of the image is extracted using the edge detection algorithm (Canny edge detection algorithm). The Canny edge detection algorithm can accurately extract the edges in the image by smoothing the image with Gaussian filtering, calculating the image gradient magnitude and direction, non-maximum suppression, and double threshold detection steps. Then, the shape features of the target region are calculated based on the edge information, including area, perimeter, and roundness. The shape features can provide important information about the shape of the target region for the segmentation algorithm.

[0047] Segmentation model training

[0048] Dataset Construction: Collect a large amount of medical image data and have it labeled by professional doctors to build training and test datasets containing images and corresponding labeling results. The datasets should cover different types of medical images, including X-rays, CT scans, and MRI scans, and include images of various lesions and normal tissues to ensure the generalization ability of the model.

[0049] Model Selection and Training: Convolutional Neural Networks (CNNs) from deep learning were selected as the segmentation model, including the U-Net network. The U-Net network has an encoder-decoder structure. The encoder part extracts high-level features of the image through convolution and pooling operations, while the decoder part restores the spatial resolution of the image through upsampling and convolution operations. Skip connections are used to fuse features from corresponding layers of the encoder and decoder to retain more image detail. The CNN model was trained using a training dataset, and the model parameters were continuously adjusted through backpropagation to enable the model to learn the mapping relationship between image features and segmentation results. During training, the cross-entropy loss function was used to measure the difference between the model's prediction results and the ground truth labels, and the stochastic gradient descent (SGD) optimization algorithm was used to minimize the loss function.

[0050] Segmentation result optimization

[0051] Post-processing operations: Post-processing is performed on the initial segmentation results output by the CNN model, including removing small noise regions and filling holes. Removing small noise regions can be achieved by setting an area threshold, and regions with an area smaller than the threshold are identified as noise and removed. Filling holes can be achieved using morphological operations (closing operations). Through post-processing operations, the accuracy and completeness of the segmentation results can be improved.

[0052] Multi-model fusion: This involves using multiple different CNN models for segmentation and then fusing their segmentation results. For example, three different U-Net network models can be selected to segment the image separately. Then, weights are determined based on the performance metrics (accuracy, recall, F1 score) of each model on the test dataset. The segmentation results of the three models are then weighted and averaged. Multi-model fusion can comprehensively consider the advantages of different models and further improve segmentation accuracy.

[0053] The present invention also provides a medical image segmentation accuracy improvement system, the system comprising:

[0054] Image preprocessing module: This module is used to denoise and enhance the input medical images. It automatically selects an appropriate filtering algorithm for denoising based on the noise characteristics of the image and uses histogram equalization technology for image enhancement.

[0055] Feature extraction module: This module is used to extract texture and shape features from the preprocessed image. It uses the gray-level co-occurrence matrix to calculate texture feature parameters and uses an edge detection algorithm to extract edge information, and then calculates shape feature parameters.

[0056] Segmentation Model Training Module: This module is used to build training and testing datasets, select and train CNN segmentation models. It is responsible for collecting and labeling medical image data, building the CNN model architecture, and training the model using the training dataset.

[0057] The segmentation result optimization module is used to perform post-processing and multi-model fusion on the preliminary segmentation results output by the CNN model. This module performs post-processing operations such as removing noisy regions and filling holes on the preliminary segmentation results, and performs weighted average fusion of the segmentation results of multiple different CNN models to output the final segmentation result.

[0058] An application example of a method and system for improving the accuracy of medical image segmentation:

[0059] I. Overview of Application Scenarios

[0060] In the radiology department of a large tertiary hospital, doctors process a large number of lung CT images daily to diagnose lung diseases such as lung cancer and pneumonia. Accurate lung region segmentation is crucial for early disease detection, disease assessment, and treatment planning. However, existing medical image segmentation methods have some problems when processing lung CT images, such as unclear segmentation boundaries due to noise interference and unstable segmentation accuracy due to significant differences in lung structure among different patients. To solve these problems, the hospital introduced the medical image segmentation accuracy improvement method and system of this invention.

[0061] II. Specific Application Process

[0062] (I) Image Acquisition and Input

[0063] The hospital uses CT equipment to scan the patient's lungs and obtain lung CT images. These images are stored in DICOM format and transmitted to the medical image segmentation accuracy improvement system of this invention through the hospital's PACS (Picture Archiving and Communication System). After receiving the images, the system first performs format conversion and preprocessing.

[0064] (II) Image Preprocessing

[0065] Noise reduction processing

[0066] The system automatically performs noise detection on the input lung CT images. By calculating the gray-level variance of local regions of the image, it detects the presence of Gaussian noise. Therefore, the system employs an adaptive Gaussian filtering algorithm for noise reduction. Based on the noise variance of different local regions, the kernel size and standard deviation of the Gaussian filter are automatically adjusted. For example, in areas with higher noise, a larger kernel (5×5) and a higher standard deviation (1.5) are used for filtering; in areas with lower noise, a smaller kernel (3×3) and a lower standard deviation (1.0) are used for filtering. This effectively removes noise while preserving as much edge and detail information as possible in the image.

[0067] After denoising, the noise in the image is significantly reduced, and the outline of the lung tissue is clearer.

[0068] Image enhancement

[0069] The system uses histogram equalization to enhance the denoised image. By adjusting the grayscale distribution of the image, the pixel values ​​that were originally concentrated in a small grayscale range are distributed more evenly, thus improving the image contrast. For example, in the original image, the grayscale difference between lung tissue and surrounding tissue is small. After histogram equalization, this difference becomes more obvious, which facilitates subsequent feature extraction and segmentation.

[0070] (III) Feature Extraction

[0071] Texture feature extraction

[0072] The system extracts texture features from the enhanced image using the Gray-Level Co-occurrence Matrix (GLCM). When calculating the GLCM, a distance of 1 and angles of 0°, 45°, 90°, and 135° are selected. Texture feature parameters such as energy, entropy, and contrast are calculated through the GLCM. For example, the texture of normal lung tissue is relatively uniform, with high energy and low entropy values. In contrast, the texture of lesion areas (lung cancer) is usually more complex, with lower energy and higher entropy values. These texture features can provide important information about different regions of the lung for the segmentation algorithm.

[0073] Shape feature extraction

[0074] The system employs the Canny edge detection algorithm to extract edge information from images. First, a Gaussian filter is used to smooth the image, reducing the impact of noise. Then, the gradient magnitude and direction of the image are calculated. Next, non-maximum suppression is performed to preserve local maxima along the gradient direction. Finally, a dual-threshold detection method is used: pixels with gradient magnitudes greater than the higher threshold are identified as edge points; pixels with gradient magnitudes between the higher and lower thresholds are identified as edge points if they are connected to edge points, otherwise they are identified as non-edge points. Based on the extracted edge information, shape feature parameters such as area, perimeter, and roundness of the lung region are calculated. For example, normal lungs have relatively regular shapes and high roundness, while lung regions affected by lesions may have irregular shapes and lower roundness.

[0075] (IV) Segmentation Model Training and Segmentation

[0076] Model training

[0077] The hospital collected a large amount of lung CT image data, which was annotated by experienced radiologists. Training and test datasets were constructed, which included images and corresponding annotation results. The training dataset contained 1,000 lung CT images and their annotation results, and the test dataset contained 200 lung CT images and their annotation results.

[0078] The system selected the U-Net network as the segmentation model, built the model architecture using Python and the TensorFlow deep learning framework, set the learning rate to 0.001 and the batch size to 32, and trained the model using the training dataset. The model parameters were continuously adjusted using the backpropagation algorithm, enabling the model to learn the mapping relationship between image features and segmentation results. During training, the cross-entropy loss function was used to measure the difference between the model's predictions and the ground truth labels, and the stochastic gradient descent (SGD) optimization algorithm was used to minimize the loss function. After multiple iterations of training, the model achieved an accuracy of over 90%, a recall of over 85%, and an F1 score of over 87% on the test dataset.

[0079] Image segmentation

[0080] For new lung CT images, the system uses a trained U-Net model for segmentation. The model outputs preliminary lung region segmentation results, which are presented in the form of a binary image, where white areas represent lung regions and black areas represent background regions.

[0081] (V) Optimization of Segmentation Results

[0082] Post-processing

[0083] The system performs post-processing on the initial segmentation results. First, an area threshold (500 pixels) is set, and regions with an area smaller than this threshold are identified as noise regions and removed. This avoids misidentifying small noise spots as lung regions. Then, morphological closing operations are used to fill the cavities. The closing operation fills the small cavities inside the lung regions by first dilating and then eroding, making the segmentation results more complete.

[0084] Multi-model fusion

[0085] To further improve segmentation accuracy, the system employs multiple different U-Net models for segmentation and fuses their results. The hospital also trained two additional U-Net models on different subsets of training data with varying hyperparameter settings. Weights were determined based on each model's performance metrics (accuracy, recall, and F1 score) on the test dataset. For example, Model 1 had an accuracy of 92%, a recall of 88%, and an F1 score of 90%; Model 2 had an accuracy of 91%, a recall of 87%, and an F1 score of 89%; and Model 3 had an accuracy of 90%, a recall of 86%, and an F1 score of 88%. Based on these metrics, Model 1 was assigned a weight of 0.4, Model 2 a weight of 0.35, and Model 3 a weight of 0.25. The segmentation results from the three models were then weighted and averaged to obtain the final segmentation result.

[0086] III. Evaluation of Application Effectiveness

[0087] (I) Improved segmentation accuracy

[0088] Compared with traditional segmentation methods (threshold-based and edge-detection-based segmentation methods), the medical image segmentation accuracy improvement method and system of this invention have achieved significant results in lung CT image segmentation. Traditional threshold-based segmentation methods suffer from low segmentation accuracy (accuracy of only about 70%) and recall (only about 65%) when processing lung CT images due to noise and uneven gray-scale distribution in lung tissue. Edge-detection-based segmentation methods struggle to accurately extract edge information when edges are blurred, resulting in unsatisfactory segmentation accuracy (accuracy of approximately 75%) and recall (approximately 70%). In contrast, the system of this invention achieves a segmentation accuracy of over 95%, a recall of over 92%, and an F1 score of over 93%, significantly improving the segmentation accuracy of the lung region.

[0089] (II) Adaptability to different patients

[0090] The lung structure varies among patients, including shape, size, and lesion status. The system of this invention, through adaptive filtering algorithms and deep learning models, can effectively adapt to these individual differences, maintaining high segmentation accuracy when segmenting lung CT images of patients of different ages, genders, and disease conditions. For example, in elderly patients, the image features are more complex due to potential degenerative changes in the lungs, such as emphysema, yet the system of this invention can still accurately segment the lung region.

[0091] (III) Clinical Application Value

[0092] Accurate lung region segmentation provides crucial information for doctors' diagnoses. Based on the segmentation results, doctors can more clearly observe the morphology and structure of lung tissue, identifying the location, size, and shape of lesions. For example, in the early diagnosis of lung cancer, the lung regions segmented by the system of this invention allow doctors to more accurately determine the nature of lung nodules, improving the early detection rate of lung cancer. Furthermore, the segmentation results can be used for subsequent medical research, such as studying the pathogenesis of lung diseases and evaluating the effectiveness of treatment regimens.

[0093] IV. Conclusion

[0094] The medical image segmentation accuracy improvement method and system of this invention has significant application effects in lung CT image segmentation. Through a series of steps including image preprocessing, feature extraction, segmentation model training, and segmentation result optimization, it effectively improves segmentation accuracy, enhances adaptability to different patients, and provides more accurate and reliable image segmentation results for medical diagnosis and treatment, thus having important clinical application value.

[0095] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for improving the accuracy of medical image segmentation, characterized in that, Includes the following steps: Image preprocessing steps: The input medical image is subjected to denoising and image enhancement processing. The denoising processing adopts an adaptive filtering algorithm, which automatically adjusts the filtering parameters according to the noise statistical characteristics of the local area of ​​the image, removing image noise while retaining image edge and detail information. The image enhancement processing adopts histogram equalization technology, which improves the contrast of the image by adjusting the gray-level distribution of the image, making the difference between the target area and the background area more significant. Feature extraction steps: Texture and shape features are extracted from the preprocessed image. The texture features are calculated using the gray-level co-occurrence matrix and specifically include energy, entropy, and contrast parameters, which are used to describe the texture characteristics of different regions in the image. The shape features are calculated based on the edge information extracted by the edge detection algorithm and include area, perimeter, and roundness parameters, which are used to characterize the morphological features of the target region. Segmentation model training steps: A training dataset and a test dataset containing a large number of medical images and their corresponding annotations were constructed. A convolutional neural network was selected as the segmentation model. The segmentation model was trained using the training dataset. The model parameters were continuously adjusted through the backpropagation algorithm so that the model could learn the mapping relationship between image features and segmentation results. During the training process, the cross-entropy loss function was used to measure the difference between the model's prediction results and the real annotations. The stochastic gradient descent optimization algorithm was used to minimize the loss function until the model's performance on the test dataset reached the preset requirements. Segmentation result optimization steps: The initial segmentation results output by the segmentation model are post-processed and multi-model fusion is performed. The post-processing includes removing small noise regions and filling holes to improve the accuracy and completeness of the segmentation results. The multi-model fusion adopts a weighted average method to merge the segmentation results of multiple different segmentation models, comprehensively consider the advantages of each model, and further improve the segmentation accuracy.

2. The method for improving the accuracy of medical image segmentation according to claim 1, characterized in that: In the denoising process, appropriate adaptive filtering algorithms are used for different types of noise; when Gaussian noise is present in the image, Gaussian filtering is used; when salt-and-pepper noise is present in the image, median filtering is used.

3. The method for improving the accuracy of medical image segmentation according to claim 1, characterized in that: In the edge detection algorithm, the Canny edge detection algorithm is selected to extract the edge information of the image. This algorithm accurately extracts the edges in the image by smoothing the image with Gaussian filtering, calculating the image gradient magnitude and direction, non-maximum suppression, and double threshold detection steps.

4. The method for improving the accuracy of medical image segmentation according to claim 1, characterized in that: In the segmentation model training step, the segmentation model is a U-Net network with an encoder-decoder structure. The encoder part extracts high-level features of the image through convolution and pooling operations, and the decoder part restores the spatial resolution of the image through upsampling and convolution operations. The features of the corresponding layers of the encoder and decoder are fused through skip connections to retain more image detail information.

5. The method for improving the accuracy of medical image segmentation according to claim 1, characterized in that: In the multi-model fusion step, the multiple different segmentation models are models trained on different network structures, different hyperparameter settings, or different training datasets; during weighted average fusion, the weights are determined based on the performance metrics of each model on the test dataset, with models having higher performance metrics having higher weights.

6. A medical image segmentation accuracy improvement system, characterized in that, include: Image preprocessing module: used to perform noise reduction and image enhancement processing on the input medical images; The denoising module employs an adaptive filtering algorithm, automatically adjusting filtering parameters based on the noise characteristics of local image regions; the image enhancement module utilizes histogram equalization technology to improve image contrast. Feature extraction module: used to extract texture features and shape features from the preprocessed image; the texture feature extraction module calculates the texture feature parameters of the image using the gray-level co-occurrence matrix; the shape feature extraction module calculates the shape feature parameters of the target region based on the edge information extracted by the edge detection algorithm; Segmentation model training module: used to build training and test datasets, select and train segmentation models; the training dataset construction module is responsible for collecting a large amount of medical image data and annotating it by professional doctors; the model training module uses the training dataset to train the segmentation model, adjusts the model parameters through the backpropagation algorithm, and optimizes the model using the cross-entropy loss function and stochastic gradient descent optimization algorithm. Segmentation result optimization module: used for post-processing and multi-model fusion of the preliminary segmentation results output by the CNN model; the post-processing module includes a sub-module for removing small noise regions and a sub-module for filling holes; The multi-model fusion module uses a weighted average method to fuse the segmentation results of multiple different CNN models.

7. The medical image segmentation accuracy improvement system according to claim 6, characterized in that: In the image preprocessing module, the denoising module calls the Gaussian filtering function when Gaussian noise is detected, and calls the median filtering function when salt-and-pepper noise is detected.

8. The medical image segmentation accuracy improvement system according to claim 6, characterized in that: In the feature extraction module, the shape feature extraction module uses the Canny edge detection algorithm to extract the edge information of the image, and calculates the shape feature parameters of the target region based on the edge information, including area, perimeter, and circularity.

9. A medical image segmentation accuracy improvement system according to claim 6, characterized in that: In the segmentation model training module, the segmentation model is a U-Net network, which is implemented through a deep learning framework and has an encoder-decoder structure and a skip connection mechanism.

10. A medical image segmentation accuracy improvement system according to claim 6, characterized in that: In the segmentation result optimization module, the multi-model fusion module determines the weights of each segmentation model based on the performance indicators of each segmentation model on the test dataset, performs a weighted average fusion of the segmentation results of multiple models, and outputs the final segmentation result.

Images