An adaptive image classification method and system for medical image datasets

Abstract

The present invention relates to the field of image recognition technology, and in particular to an adaptive image classification method and system for medical image data sets. The method comprises the following steps: obtaining a medical image data set to be classified, and extracting a key interest region of each image, wherein the medical image data set to be classified includes a number of X-ray medical images and CT medical images; performing frequency domain conversion on the medical image data set to be classified according to the key interest region, and analyzing the texture fine-grainedness of each image in the medical image data set to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; integrating high-frequency texture fine-grainedness and low-frequency texture fine-grainedness as the first classification feature point. The present invention improves the accuracy and adaptability of classification of different types of medical images through adaptive image enhancement, texture fine-grainedness analysis and contrast difference calculation.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to an adaptive image classification method and system for medical image data sets. Background Art

[0002] In the early days, the classification methods of medical images mainly relied on traditional computer vision techniques, such as hand-crafted algorithms based on feature extraction. These methods rely on features designed by experts, such as texture, shape, and edge information, but due to the complexity and diversity of images, the effect of hand-crafted features is limited and it is difficult to adapt to different types of image data sets. With the rise of deep learning, especially the application of convolutional neural networks (CNN), image classification technology has made revolutionary progress. Deep learning can automatically learn more complex and abstract features from large-scale image data, thereby improving classification accuracy. In particular, the emergence of image classification networks such as Ale×Net, VGG, and ResNet has significantly improved the performance of medical image classification tasks. In order to solve the adaptability problem of different types of images, researchers have gradually proposed adaptive classification methods, using technologies such as transfer learning, data augmentation, and generative adversarial networks (GAN) to enhance the generalization ability of the model. However, traditional methods often lack effective classification adaptability between similar types of medical images (such as X-ray and CT images), resulting in unstable classification results. At the same time, traditional classification methods have the problem of insufficient image feature extraction, resulting in poor classification results. Summary of the invention

[0003] Based on this, it is necessary to provide an adaptive image classification method and system for medical image datasets to solve at least one of the above technical problems.

[0004] To achieve the above object, an adaptive image classification method for a medical image dataset is provided, the method comprising the following steps:

[0005] Step S1: obtaining a medical image dataset to be classified and extracting a key region of interest of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0006] Step S2: performing frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyzing the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; integrating the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point;

[0007] Step S3: performing local and global contrast difference calculations on the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data;

[0008] Step S4: designing an adaptive image enhancement function using the local contrast difference data and the global contrast difference data, and enhancing the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features;

[0009] Step S5: Analyze the texture fold density of the key interest area after the enhanced features based on the first classification feature points, and use the calculated texture fold density of the key interest area as the second classification feature points;

[0010] Step S6: Classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain classification results of X-ray medical images and CT medical images.

[0011] The present invention can focus on the most important part of the image by extracting the key interest area (such as tumor, lesion area, etc.) of each image, reduce the complexity of data processing, and improve the accuracy of classification. Frequency domain conversion helps to capture subtle texture differences in the image, which is particularly important for the classification of medical images. High-frequency and low-frequency texture information can reveal different pathological characteristics and help distinguish normal and abnormal areas. Contrast differences can help establish clear boundaries between different pathological structures. Local contrast differences can reveal details, while global contrast differences help identify large-scale features of the image. Such contrast analysis can enhance the visibility of lesions and contribute to more accurate classification. Image enhancement improves the recognizability and detail expression of the image, especially in low-contrast areas, which is crucial to improving the accuracy of the classification model in complex environments. Texture fold density can reflect the lesion area (such as cancer, mass, etc.) in the image. As an additional feature, it helps to improve the fine granularity of classification. Combining multiple features for classification (such as frequency domain texture fine granularity and texture fold density) can improve the robustness and accuracy of the classification model, help to distinguish the types of lesions in X-ray and CT images, and improve the accuracy of medical diagnosis. By extracting multi-dimensional features (frequency domain texture, contrast, wrinkle density, etc.), the details of medical images can be described more accurately and the effect of image classification can be improved. Adaptive image enhancement optimizes image quality based on local and global contrast differences, which is particularly helpful for processing low-contrast or blurred medical images. By combining multiple feature points for comprehensive classification, not only can the accuracy of classification be improved, but also the misclassification can be reduced, especially in complex or marginal medical image areas. Therefore, the present invention improves the accuracy and adaptability of classification of different types of medical images through adaptive image enhancement, fine-grained texture analysis and contrast difference calculation.

[0012] Preferably, step S1 comprises the following steps:

[0013] Step S11: obtaining a medical image dataset to be classified, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0014] Step S12: setting a Gaussian filter kernel size of 5×5 or 7×7, performing Gaussian blur on each image in the medical image dataset to be classified, and obtaining a denoised medical image dataset to be classified;

[0015] Step S13: performing edge detection on the denoised medical image data set to be classified to obtain a medical classification edge image;

[0016] Step S14: extracting edge response values ​​of medical classification edge images and performing edge image weighted fusion on the denoised medical image dataset to be classified, thereby obtaining edge image intensity weights;

[0017] Step S15: perform edge difference calibration on each medical classification edge image according to the edge image intensity weight to obtain the key interest region of each image.

[0018] The present invention can ensure the generalization ability of the model and adapt to different types of image data by constructing data sets from various types of medical images (X-rays, CT), thereby improving the accuracy and robustness of classification. Gaussian blur helps to remove noise in images, especially low-quality images or random noise commonly seen during scanning. Denoising can reduce unnecessary interference, retain key features of the image, and lay the foundation for subsequent edge detection and feature extraction. Edge detection can highlight significant structures in images, especially lesion areas or important anatomical features. In medical images, edge detection can reveal the outline of lesions, making key features in the image more obvious and convenient for subsequent analysis. Edge response values ​​and weighted fusion help to enhance the significance of edge information. In image processing, edge information is often the most reflective feature of image structure and morphology. After weighted fusion, edge information can more accurately guide subsequent regional analysis and classification. Through edge difference calibration, key interest areas (such as lesions, masses, etc.) in the image can be identified. This method ensures accurate processing of the most important parts of the image and avoids excessive processing of irrelevant areas, thereby improving the efficiency and accuracy of image classification.

[0019] Preferably, step S2 comprises the following steps:

[0020] Step S21: Perform Fourier transform on the key interest region of the medical image data set to be classified to obtain a frequency domain conversion map of the key interest region;

[0021] Step S22: dividing the frequency domain conversion map of the key interest area based on the threshold value to obtain a high frequency part and a low frequency part of the frequency domain conversion map;

[0022] Step S23: performing fine-grained texture analysis on the high-frequency part and the low-frequency part of the frequency domain conversion image respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness;

[0023] Step S24: performing difference value calculation based on the fine-grained texture data of the high-frequency part and the fine-grained texture data of the low-frequency part, and selecting texture feature difference points of the key interest area as the first classification feature points through the calculated difference values.

[0024] The present invention can transform an image from the spatial domain to the frequency domain through Fourier transform, revealing the periodicity and frequency characteristics implicit in the image. In medical images, frequency domain transformation helps to capture some details that are difficult to observe through direct spatial domain analysis, especially the microscopic features of lesions. This provides a more comprehensive perspective for subsequent feature extraction and analysis. The high-frequency part of the frequency domain image usually reflects the details in the image (such as edges and textures), while the low-frequency part contains the basic structure of the image (such as large-scale shapes and textures). By dividing the frequency domain image into high and low frequencies, the details and overall structure of the image can be analyzed separately, helping to highlight the characteristics of different lesions or abnormal areas. High-frequency texture analysis can reveal the details, changes and tiny structures in the image, while low-frequency texture analysis helps to describe the large-scale changes and overall characteristics in the image. By analyzing the texture of the high-frequency and low-frequency parts separately, different levels of features in the image can be more comprehensively captured, especially pathological structures, masses, tissue damage, etc. in medical images. By calculating the difference between the high-frequency and low-frequency texture parts, the area where significant changes occur in the image can be highlighted. These texture difference points often represent the lesion area or the most diagnostically significant part of the image. Using these difference points as classification feature points can help improve the classification model's ability to recognize key pathological information, thereby improving image classification accuracy.

[0025] Preferably, step S22 includes the following steps:

[0026] A 2D directional filter is used to calculate the angular distribution of the high-frequency part and the low-frequency part of the frequency domain conversion diagram in the frequency domain direction, and the frequency domain angular distribution of the high-frequency part and the frequency domain angular distribution of the low-frequency part are obtained. The calculation formula is as follows:

[0027]

[0028] in, Represents the direction angle of each point in the frequency domain, is the complex frequency component in the frequency domain, is the frequency component of the image in the horizontal direction, is the frequency component of the image in the vertical direction;

[0029] The Gabor filter is used to analyze the texture information of the high-frequency part and the low-frequency part of the frequency domain conversion image, and the texture information of the high-frequency part and the texture information of the low-frequency part are obtained. The calculation formula is as follows:

[0030]

[0031] in, Represents texture information, is the number of sampling points in the region, is the regional mean, For the The complex frequency components in the frequency domain, For the The frequency component of an image in the horizontal direction, For the The frequency component of an image in the vertical direction;

[0032] The frequency domain angle distribution of the high frequency part and the texture information of the high frequency part are integrated into the high frequency texture fine granularity, and the frequency domain angle distribution of the low frequency part and the texture information of the low frequency part are integrated into the low frequency texture fine granularity.

[0033] The present invention can reveal the directional characteristics of different frequency components in the image by calculating the frequency domain angle distribution. In particular, in medical images, the lesion area often has specific directional characteristics (such as the shape of the tumor, the direction of the blood vessels, etc.). The frequency domain angle distribution can help capture these directional information, enhance the model's understanding of the image texture and structure, and thus improve the accuracy and efficiency of classification. Gabor filters are widely used in texture analysis because they can effectively extract local structural features in images. Using Gabor filters in frequency domain analysis, texture features can be extracted for different frequency components, which is particularly important for medical images because many lesions (such as tumors or vascular abnormalities) show specific texture patterns. By calculating texture information, the texture differences in different regions can be more accurately identified, further improving the accuracy of image analysis. Combining frequency domain angle distribution and texture information can provide more comprehensive features for image classification. These features not only reflect the directional and structural characteristics of the image, but also contain detailed information of the texture, thereby helping the classification model to better identify lesions and healthy areas. The integration of high-frequency and low-frequency parts can reveal the multi-level and multi-dimensional characteristics of the image, allowing the classification model to process different image features more finely and improve the overall classification accuracy.

[0034] Preferably, step S3 comprises the following steps:

[0035] Step S31: Based on the preset medical pathology image library, the key interest area of ​​each image is screened for abnormal areas, and the contrast difference between the lesion area and the normal tissue is analyzed to obtain local contrast difference data, wherein the calculation formula of the contrast difference is as follows:

[0036]

[0037] in, For images at position The local area difference value, For images at position The gradient at For images at position The gray value of is the image horizontal coordinate, is the image ordinate;

[0038] Step S32: Calculate the global grayscale mean and standard deviation of the medical image data set to be classified to obtain the global grayscale mean and the global grayscale standard deviation of the image;

[0039] Step S33: Analyze the overall brightness, contrast and grayscale distribution characteristics of the medical image data set to be classified using the global grayscale mean and the global grayscale standard deviation of the image to obtain global contrast difference data.

[0040] The present invention can reflect the grayscale changes between different areas in the image through local contrast differences. By calculating the gradient and local contrast of each point, the area with significant contrast difference can be identified, which is crucial for the identification of the lesion area. The lesion area usually has a large difference in grayscale value from the normal tissue, so the analysis of contrast difference can help to accurately screen out the abnormal area in the image. The method is particularly suitable for detecting local lesions such as tumors, nodules or other tissue abnormalities. Calculating the global grayscale mean and standard deviation can help capture the overall brightness and grayscale distribution characteristics of the image. These global features are very useful for global adjustment and classification of images. By comparing the grayscale mean and standard deviation of different images, the changes in brightness and contrast of the image in the global range can be identified, which helps to understand the overall structure of the image and optimize the subsequent analysis and classification. The global contrast difference can effectively reveal the differences in the overall grayscale distribution of different images. By combining the global grayscale mean and standard deviation, the brightness, contrast and grayscale distribution characteristics of the image can be further analyzed to identify the overall changes in the image. For normal and abnormal regions in medical images, the analysis of global contrast differences helps to identify lesion regions with abnormal global brightness and contrast (such as the edge of a tumor), providing support for image classification and lesion detection.

[0041] Preferably, designing an adaptive image enhancement function using local contrast difference data and global contrast difference data in step S4 comprises the following steps:

[0042] Set the initial image enhancement function, where the initial image enhancement function is as follows:

[0043]

[0044] in, is the initial image enhancement function, is the original image, To control the weighting coefficient of local difference on the enhancement effect, To control the weighting coefficient of the global difference on the enhancement effect, is the local contrast, is the global contrast;

[0045] Introducing adaptive coefficients The initial image enhancement function is adjusted for low-contrast region intensity, and the adaptive coefficient is set to 1.2~2.0 to obtain the low-contrast region intensity adjustment parameter, where the formula for low-contrast region intensity adjustment is as follows:

[0046]

[0047] in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect;

[0048] Introducing adaptive coefficients The initial image enhancement function is adjusted for high contrast region intensity, and the adaptive coefficient is set to 0.5-1.5 to obtain the high contrast region intensity adjustment parameter, where the formula for high contrast region intensity adjustment is as follows:

[0049]

[0050] in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect;

[0051] The initial image enhancement function is adaptively selected using the low-contrast region intensity adjustment parameter and the high-contrast region intensity adjustment parameter to obtain an adaptive image enhancement function.

[0052] The present invention can enhance the original image according to the local contrast ( ) and global contrast ( ) data, adjust the brightness and contrast of the image, so that the image has a more balanced enhancement effect before processing. This is to improve the visual effect of the overall image, especially the low-contrast area, which can help the feature extraction and lesion area identification in the subsequent image analysis process. Low-contrast areas usually appear in areas with small grayscale differences in the image, such as early lesion areas or background noise. Introducing adaptive coefficients It can dynamically adjust the enhancement intensity of low-contrast areas, making these areas clearer and improving the visibility of image details. By optimizing low-contrast areas, it can enhance the details that are difficult to distinguish in the image, especially the detection of early lesions in medical images, helping to improve diagnostic accuracy. High-contrast areas usually contain more detailed information, especially in already clearly visible lesions or contrasting tissues. Use appropriate adaptive coefficients Enhancement and adjustment of these areas can help avoid over-enhancement, which can lead to loss of image details or artifacts. By balancing the enhancement strength of these areas, it is possible to ensure that the lesion area presents the best contrast in the image, avoiding overexposure or blurring, thereby optimizing the doctor's observation and analysis of the lesion area. Finally, by adaptively selecting the enhancement functions of low-contrast and high-contrast areas, different enhancement strategies can be applied to different areas according to the specific needs of the image. This adaptive mechanism enables image processing to more accurately optimize the local features of the image, avoiding over-enhancement or information loss caused by a unified enhancement strategy. This method is particularly effective in medical imaging, and can refine image enhancement and improve the visibility and diagnostic quality of lesion areas.

[0053] Preferably, the adaptive selection of the initial image enhancement function using the low contrast region intensity adjustment parameter and the high contrast region intensity adjustment parameter comprises:

[0054] If the local contrast If it is less than 0.5, the low contrast area intensity adjustment function is used. ;

[0055] If the local contrast If it is greater than 1.5, the low contrast area intensity adjustment function is used. ;

[0056] If the local contrast In the range of 0.5 ~ 1.5, the initial image enhancement function is used .

[0057] The present invention helps to improve the visibility of these areas by selecting the low-contrast area intensity adjustment function in areas with low local contrast (such as early lesion areas or tiny detail areas). By adjusting the enhancement intensity of low-contrast areas, details can be effectively enhanced to help doctors identify potential lesions or abnormal areas. Enhancing low-contrast areas makes the details in the image clearer, especially in areas with small grayscale value changes or unclear structures, which helps early diagnosis. In areas with high local contrast (such as areas with obvious lesions or places with strong contrast between the background and the target area), excessive enhancement leads to image artifacts or loss of details. Selecting the intensity adjustment function of high-contrast areas can avoid this problem, moderately adjust the contrast of these areas, retain important details and prevent excessive enhancement. Maintain the clarity of details in high-contrast areas, avoid artifacts or overexposure, and ensure optimal image quality, especially at significant lesions or tissue boundaries. For areas with medium local contrast, there is usually no strong enhancement demand. Using the initial image enhancement function can balance global and local differences, avoid unnecessary overprocessing, and retain the natural details of the area. Balanced enhancement processing is performed on the image to retain the natural contrast of the overall image, so that the image is balanced in details and quality.

[0058] Preferably, step S5 comprises the following steps:

[0059] Step S51: constructing a gray level co-occurrence matrix of the key interest region after the enhanced features based on the first classification feature points, to obtain a gray level co-occurrence matrix of the key interest region after the enhanced features;

[0060] Step S52: Analyze the texture distribution of the key interest region after the enhanced features by using the gray level co-occurrence matrix of the key interest region after the enhanced features to obtain the texture distribution;

[0061] Step S53: Calculate the pixel complexity of the key interest area after the enhanced features according to the texture distribution, and construct a texture change gradient image of the key interest area after the enhanced features through the pixel complexity;

[0062] Step S54: For each region of interest, the frequency of the change of the texture change gradient image is calculated to obtain the texture wrinkle density of the key region of interest and use the texture wrinkle density as the second classification feature point.

[0063] The present invention can effectively capture the spatial relationship of image texture and the statistical characteristics of grayscale change by constructing a grayscale co-occurrence matrix for the key interest region after strengthening the features based on the first classification feature points. Grayscale co-occurrence matrix is ​​a common method for texture analysis. It can reveal the grayscale distribution of pixels in the region and their mutual relationship, and help extract the detailed texture information of the image. It provides high-dimensional statistical features for subsequent texture analysis, reflects the spatial consistency and local structure of the image texture, and improves the detail perception ability of image classification. By analyzing the grayscale co-occurrence matrix of the key interest region after strengthening the features, the texture characteristics of the region can be deeply understood, including the uniformity, directionality, roughness, etc. of the texture. Texture distribution analysis helps to reveal the complexity inside the region and further distinguish different types of regions (such as the difference between lesions and normal tissues). By analyzing the distribution of image texture, a more accurate basis can be provided for subsequent classification, so that the system can accurately classify based on texture information and enhance the accuracy of image processing and analysis. By calculating pixel complexity and constructing a texture change gradient image, the complexity and change trend of the key interest region can be revealed. Pixel complexity reflects the degree of detail in the image region, while the texture change gradient image can reflect the degree of change of the texture in the region and help identify the change details in the image. The dynamic changes of texture information in the area are enhanced, and complex texture features related to the lesions can be extracted, thereby improving the image's ability to identify the lesion area and assisting doctors in making better diagnoses. By calculating the change frequency of the texture change gradient image, the texture fold density is obtained and used as the second classification feature point, which helps capture subtle image texture changes and complex structures. This density value is very important for detailed identification of lesion areas, especially when the texture of lesions or abnormal areas changes significantly. Texture fold density can provide key feature information for image classification, help accurately distinguish lesions from normal areas, and further improve the accuracy and reliability of medical image analysis.

[0064] Preferably, step S6 comprises the following steps:

[0065] Step S61: constructing a feature vector for an image in the medical image dataset to be classified by using the first classification feature point and the second classification feature point to obtain a classification feature vector;

[0066] Step S62: training a classification model based on the classification feature vector, and performing image classification on the medical image data set to be classified according to the trained classification model, thereby obtaining classification results of X-ray medical images and CT medical images.

[0067] The present invention combines the first classification feature point with the second classification feature point to generate a feature vector containing rich information. The feature vector of each image is a comprehensive representation of the image in different dimensions, covering information such as local contrast, global contrast, texture distribution, and texture change. This high-dimensional feature vector provides a multi-dimensional data input for the classification model, which helps the model to capture the details of the image more accurately. The feature vector provides comprehensive image feature data, so that the classification model can efficiently identify medical images of different categories, thereby improving the performance of the classification task. The classification model is trained based on the classification feature vector, and the relationship between various features in the image and the target category can be learned. Through the machine learning algorithm, the model can effectively identify the different features of X-ray medical images and CT medical images, thereby completing the image classification task. This process makes full use of the multi-dimensional information of the feature vector to ensure the high accuracy of the classification. The trained classification model can efficiently classify the medical images to be classified, classify the X-ray images and CT images separately, and help doctors or medical systems quickly and accurately identify different types of image data, providing support for subsequent diagnosis and analysis.

[0068] In this specification, an adaptive image classification system for a medical image dataset is provided, which is used to execute the adaptive image classification method for a medical image dataset. The adaptive image classification system for a medical image dataset includes:

[0069] An image acquisition module is used to obtain a medical image dataset to be classified and extract key interest regions of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0070] A fine-grained extraction module is used to perform frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyze the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; and integrate the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point;

[0071] A difference analysis module is used to calculate the local and global contrast differences of the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data;

[0072] A feature enhancement module is used to design an adaptive image enhancement function using local contrast difference data and global contrast difference data, and enhance the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features;

[0073] A wrinkle extraction module is used to analyze the texture wrinkle density of the key interest area after the enhanced feature based on the first classification feature point, and use the calculated texture wrinkle density of the key interest area as the second classification feature point;

[0074] The classification module is used to classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain the classification results of the X-ray medical image and the CT medical image.

[0075] The beneficial effect of the present invention is that the key interest region of each image can be extracted from the medical image data set to be classified through the image acquisition module, which lays the foundation for subsequent fine-grained extraction, difference analysis and other processing. Through accurate interest region extraction, it is ensured that subsequent analysis and feature extraction only focus on the most representative part of the image, thereby improving overall efficiency. Through frequency domain conversion and texture fine-grained analysis, the module can accurately capture the texture features of each medical image, especially the detailed information of the high-frequency and low-frequency parts. This helps to capture subtle changes in the image, especially the lesion area in medical images. Local and global contrast difference calculation helps the system to deeply understand the difference between the lesion area and the normal area in the image. Contrast difference is one of the key features for identifying lesions and healthy areas, and can effectively reveal the contrast changes of the image. Through the adaptive image enhancement function, the image features are dynamically adjusted according to the local and global contrast differences of the image, so that the important areas (such as lesion areas) in the image are more prominent after enhancement. In this way, the system can extract key image features more accurately. Texture wrinkle density analysis can reveal subtle texture changes in the image, which are usually related to the presence of lesion areas. By calculating the density of texture wrinkles, the system can further refine the image features and help the classification model better distinguish different types of medical images. Based on the extracted first classification feature points (high-frequency and low-frequency texture fine-grainedness) and second classification feature points (texture wrinkle density), the classification module can efficiently and accurately classify X-ray images and CT images. Through machine learning or deep learning models, the system can identify type differences in images to ensure the efficiency and accuracy of classification results. Therefore, the present invention improves the accuracy and adaptability of classification of different types of medical images through adaptive image enhancement, texture fine-grained analysis and contrast difference calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 A schematic diagram of the steps of an adaptive image classification method for a medical image dataset;

[0077] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0078] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0079] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0080] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0081] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0082] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0083] To achieve this, please refer to Figures 1 to 3 , an adaptive image classification method for a medical image dataset, the method comprising the following steps:

[0084] Step S1: obtaining a medical image dataset to be classified and extracting a key region of interest of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0085] Step S2: performing frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyzing the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; integrating the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point;

[0086] Step S3: performing local and global contrast difference calculations on the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data;

[0087] Step S4: designing an adaptive image enhancement function using the local contrast difference data and the global contrast difference data, and enhancing the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features;

[0088] Step S5: Analyze the texture fold density of the key interest area after the enhanced features based on the first classification feature points, and use the calculated texture fold density of the key interest area as the second classification feature points;

[0089] Step S6: Classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain classification results of X-ray medical images and CT medical images.

[0090] The present invention can focus on the most important part of the image by extracting the key interest area (such as tumor, lesion area, etc.) of each image, reduce the complexity of data processing, and improve the accuracy of classification. Frequency domain conversion helps to capture subtle texture differences in the image, which is particularly important for the classification of medical images. High-frequency and low-frequency texture information can reveal different pathological characteristics and help distinguish normal and abnormal areas. Contrast differences can help establish clear boundaries between different pathological structures. Local contrast differences can reveal details, while global contrast differences help identify large-scale features of the image. Such contrast analysis can enhance the visibility of lesions and contribute to more accurate classification. Image enhancement improves the recognizability and detail expression of the image, especially in low-contrast areas, which is crucial to improving the accuracy of the classification model in complex environments. Texture fold density can reflect the lesion area (such as cancer, mass, etc.) in the image. As an additional feature, it helps to improve the fine granularity of classification. Combining multiple features for classification (such as frequency domain texture fine granularity and texture fold density) can improve the robustness and accuracy of the classification model, help to distinguish the types of lesions in X-ray and CT images, and improve the accuracy of medical diagnosis. By extracting multi-dimensional features (frequency domain texture, contrast, wrinkle density, etc.), the details of medical images can be described more accurately and the effect of image classification can be improved. Adaptive image enhancement optimizes image quality based on local and global contrast differences, which is particularly helpful for processing low-contrast or blurred medical images. By combining multiple feature points for comprehensive classification, not only can the accuracy of classification be improved, but also the misclassification can be reduced, especially in complex or marginal medical image areas. Therefore, the present invention improves the accuracy and adaptability of classification of different types of medical images through adaptive image enhancement, fine-grained texture analysis and contrast difference calculation.

[0091] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a schematic diagram of a process flow of an adaptive image classification method for a medical image dataset of the present invention. In this example, the adaptive image classification method for a medical image dataset includes the following steps:

[0092] Step S1: obtaining a medical image dataset to be classified and extracting a key region of interest of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0093] In an embodiment of the present invention, a medical image data set to be classified is collected, and the data set should include multiple X-ray medical images and CT medical images. The data source can be a hospital, a medical imaging device, or a public medical imaging database (such as NIH Chest ×-ray Dataset, LIDC-IDRI, etc.). The collected medical image data is preprocessed to ensure the quality and consistency of the data. The preprocessing steps include image size unification, image denoising, gray value standardization, etc., to ensure that images from different sources have the same processing standards. A deep learning-based target detection model (such as Faster R-CNN, YOLO, etc.) is used to automatically identify key regions of interest (ROIs) in each image. These regions of interest are usually lesion areas of concern to doctors (for example, tumors, nodules, fractures, etc.). The lesion areas in X-ray images and CT images are labeled and detected by training the target detection model. In this process, a labeled training data set (with expert-annotated lesion areas) is required to train the model. In order to cope with different types of images (X-ray, CT), you can choose to train specific detection models for different image modes, or choose a general model for multimodal image detection. Once the model detects the region of interest, an image cutting algorithm (such as rectangular or irregular region cropping) is used to extract the parts of interest from the image. These regions can be used for subsequent classification and analysis. The extracted regions of interest are ensured to be accurate and have no deviations in boundaries through calibration systems or manual annotation. This usually requires verification in combination with the doctor's clinical experience. Due to the differences in imaging methods between X-ray and CT images, image alignment and registration need to be considered. Common methods include registration methods based on feature point matching or deep learning models to automatically align multimodal images. If it is necessary to further combine the information in the X-ray and CT images, feature fusion techniques (such as convolutional neural networks) can be used to fuse the information of different modalities into a unified feature representation to improve the accuracy of subsequent classification. The extracted regions of interest and the original images are stored, usually using standardized data formats (such as DICOM, NIfTI, etc.) for subsequent storage, retrieval and processing.

[0094] Step S2: performing frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyzing the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; integrating the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point;

[0095] In an embodiment of the present invention, each image in the medical image data set to be classified is normalized and denoised to reduce the influence of noise on the frequency domain conversion. Fourier transform is performed on each extracted key interest region to obtain the frequency domain representation of the image. In this way, the amplitude spectrum and phase spectrum of the image can be obtained. Usually, the amplitude spectrum is mainly concerned, which reflects the intensity of the frequency components in the image. The high-frequency part represents the details and texture information in the image, usually including edges, noise, and high-frequency fluctuations. Through the result of Fourier transform, the part with higher frequency is selected. The specific method is to set a threshold to extract the high-frequency area (such as the part greater than a certain frequency) in the image frequency domain. The high-frequency part contains fine-grained texture information, such as edges, texture details, etc. The extracted high-frequency area is analyzed, and common texture feature extraction methods (such as gray level co-occurrence matrix (GLCM), local binary pattern (LBP), etc.) can be used to obtain texture detail features. For example, GLCM is used to extract texture features such as contrast, homogeneity, and entropy. The low-frequency part usually represents the smooth area and rough structural information in the image. Select the lower frequency part from the amplitude spectrum of the Fourier transform, usually the central part of the spectrum. These areas contain low-frequency information of the image, such as background, brightness changes, etc. Similarly, perform texture analysis on the low-frequency part and use techniques such as GLCM or LBP to extract texture features in the low-frequency area. These features usually include larger structures or regional textures. Integrate the texture features extracted from the high-frequency and low-frequency areas to obtain a comprehensive texture fine-grained representation as the first classification feature point for subsequent classification. The high-frequency and low-frequency texture features can be fused into a high-dimensional feature vector by concatenation or weighted averaging. Specifically, the texture features extracted from the high-frequency and low-frequency areas (such as contrast, entropy, homogeneity, etc.) can be sequentially merged into a unified feature vector. In the fused feature space, feature selection methods (such as principal component analysis PCA, mutual information selection, etc.) can be used to reduce redundant features and improve classification performance.

[0096] Step S3: performing local and global contrast difference calculations on the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data;

[0097] In an embodiment of the present invention, a reference library containing multiple types of medical pathology images is prepared. Generally, these images should have labels to describe their characteristics or pathological states. The images in the library should cover different types of lesions (such as tumors, nodules, normal tissues, etc.), and each image should contain annotated lesion areas (i.e., key interest areas). Image samples or lesion types similar to the image to be classified are selected as references for subsequent contrast difference calculations. Local contrast reflects the degree of change in pixel brightness values ​​in a small area (usually a key interest area) in the image. Areas with higher local contrast usually contain more details and texture information, while areas with lower contrast represent smooth or uniform areas. The key interest area is selected for local contrast calculation. The method can be based on the local contrast formula, compare the local contrast of the image to be classified with the corresponding image in the preset medical pathology image library, and calculate the contrast difference of each key interest area. Methods such as mean square error (MSE) or structural similarity index (SSIM) can be used to quantitatively evaluate the contrast difference. Global contrast reflects the overall brightness difference of the entire image or a larger area, which is usually related to the brightness and global structure of the image. Images with high global contrast usually have obvious light and dark changes and clear structural boundaries. Compare the global contrast of the image to be classified with the global contrast in the preset pathology image library, and calculate the global contrast difference of each image. Compare the global contrast of the image to be classified with the global contrast in the preset pathology image library, and calculate the global contrast difference of each image. The local and global contrast difference data can be combined by simple splicing or weighted averaging to obtain a unified feature representation.

[0098] Step S4: designing an adaptive image enhancement function using the local contrast difference data and the global contrast difference data, and enhancing the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features;

[0099] In an embodiment of the present invention, for each image, a sliding window (such as a 3×3 or 5×5 window) is used to calculate the contrast difference of the local area around each pixel. The local contrast can be obtained by calculating the difference between the maximum and minimum values ​​in the window. This calculation can reflect the degree of change in brightness or color in the local area of ​​the image. The global contrast is obtained by calculating the difference between the maximum and minimum values ​​of the entire image. This step represents the overall contrast of the brightness or color of the entire image. The global contrast can measure the overall contrast intensity of the image as a global enhancement reference. Based on the local contrast difference and the global contrast difference, an adaptive enhancement function is designed. The core idea of ​​the function is to adjust the brightness or color of each pixel according to the contrast difference between the local area and the global area. The function enhances the local area with large contrast by adaptively adjusting the enhancement value of each pixel, while considering the influence of global contrast. Use image segmentation or target detection algorithms to identify key regions of interest (ROIs) in the image. These regions are usually regions containing important features, such as edges, texture-rich regions, or regions that need to be focused on in image analysis. The designed adaptive image enhancement function is applied to each key region of interest. Through this function, the brightness or color of the pixels in the area is adaptively adjusted according to the local contrast difference and global contrast difference of the area. The details are enhanced in the area with strong local contrast, and the overall perception effect is enhanced when the global contrast is weak, so as to effectively highlight the features of the area of ​​interest. According to the pixel values ​​adjusted by the adaptive enhancement function, the key area of ​​interest in the image is updated, and the image features of the area are strengthened to make it more visually prominent. In this way, the key details of the image are more obvious, which is helpful for subsequent image analysis or processing tasks.

[0100] Step S5: Analyze the texture fold density of the key interest area after the enhanced feature based on the first classification feature point, and use the calculated texture fold density of the key interest area as the second classification feature point;

[0101] In an embodiment of the present invention, the first classification feature points generally refer to feature points identified by preprocessing or preliminary analysis in an image, and these feature points are extracted based on information such as brightness, color, texture, etc., such as edge points, corner points or important texture areas. Using these first classification feature points, the position of the key region of interest (ROI) is determined by image segmentation, feature point detection or region extraction methods (such as SIFT, SURF, Harris corner point detection, etc.). The image has been enhanced by an adaptive image enhancement function, and the key regions of interest in the image are highlighted. Ensure that these regions are clearly visible in the image and contain rich texture features. Through the extracted key regions of interest, further texture analysis can be performed in subsequent steps. Texture wrinkle density is a measure of the details of the texture structure in the image, reflecting the complexity of the details in the image area. Generally, areas with higher texture wrinkle density have more details and complex texture structures. Gray level co-occurrence matrix (GLCM) can be used to extract texture features and calculate its parameters such as energy, contrast, homogeneity, etc., which reflect different aspects of texture in the image. Areas with greater contrast generally have higher texture wrinkle density. Another method is to extract local texture features of the image using methods such as local binary patterns (LBP) or Gabor filters. Texture analysis is performed on each key region of interest in the image to calculate the texture complexity, roughness, and wrinkle density of the region. After the texture features are extracted, the texture wrinkle density of each key region of interest is obtained by statistical or mathematical calculations. The texture wrinkle density can be obtained by calculating the frequency of texture changes in the region, which is usually reflected as the degree of change in the grayscale value within the image. The calculated texture wrinkle density is used as a new feature point, and these density values ​​will be used as the second classification feature points, combined with the first classification feature points for further image analysis or classification. These second classification feature points can be used for subsequent image analysis tasks, such as texture classification, defect detection, quality assessment, etc.

[0102] Step S6: Classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain classification results of X-ray medical images and CT medical images.

[0103] In an embodiment of the present invention, a medical image data set to be classified is collected, including X-ray images and CT images, to ensure that the annotation information of each image is accurate and contains known categories (such as X-ray and CT). The images in the data set can be various types of medical imaging data, including different pathological conditions, image quality, and shooting conditions. Feature extraction is performed on each image in the image to be classified, and the first classification feature points (such as edges, corners, textures, etc.) in the image are identified. The first classification feature points are usually obtained by traditional feature extraction methods (such as SIFT, SURF, Harris corner detection, etc.) or deep learning-based methods (such as features extracted by convolutional neural networks). These feature points can reflect important structures or key information in the image and help the classification model distinguish different categories. Texture analysis is performed on each image, and the texture wrinkle density of each key interest area is calculated (such as by GLCM, LBP, etc.). This will be used as the second classification feature point to further enhance the feature representation of the image. Texture wrinkle density is the texture complexity in the local area of ​​the image, and usually has different manifestations for different types of images in medical imaging (such as X-ray and CT images). Combine the first classification feature points (such as edges, corners, etc.) and the second classification feature points (such as texture wrinkle density) extracted from each image to form a feature vector containing rich features. This feature vector will represent the overall characteristics of the image, which is convenient for subsequent classification tasks. The combination method can be a simple splicing, combining the two feature sets into a new feature set in parallel; or weighted fusion according to the importance of different features. According to the characteristics of the feature vector, select an appropriate classification model for image classification. Common classification models include: Support Vector Machine (SVM): Suitable for small sample high-dimensional data, can effectively separate X-ray and CT images. Random Forest (RF): Suitable for processing complex nonlinear features, can automatically select features. K Nearest Neighbor (KNN): Suitable for classification tasks with relatively uniform distribution of data points. Neural Network: If the amount of data is large enough, you can use Multilayer Perceptron (MLP) or Convolutional Neural Network (CNN) for feature learning and classification. Use the training set data containing the first classification feature points and the second classification feature points to train the selected classification model. Through back propagation (for neural networks), SVM kernel function (for SVM) or other methods, the model can learn the optimal features to distinguish between X-ray and CT images. During the training process, cross-validation can be used to evaluate the performance of the model, prevent overfitting, and select the optimal hyperparameters. The trained model is applied to the medical image dataset to be classified to obtain the classification results for each image. The classification result will assign a label to each image, indicating whether it is an X-ray image or a CT image. When outputting the classification results, evaluation indicators such as accuracy, recall, and F1-score can be calculated to verify the performance of the model. Post-process the classification results to optimize the accuracy of image classification.For example, the classification results can be further adjusted through rule sets or subsequent correction steps, especially in the field of medical images, which requires the combination of expert knowledge for final verification.

[0104] Preferably, step S1 comprises the following steps:

[0105] Step S11: obtaining a medical image dataset to be classified, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0106] Step S12: setting a Gaussian filter kernel size of 5×5 or 7×7, performing Gaussian blur on each image in the medical image dataset to be classified, and obtaining a denoised medical image dataset to be classified;

[0107] Step S13: performing edge detection on the denoised medical image data set to be classified to obtain a medical classification edge image;

[0108] Step S14: extracting edge response values ​​of medical classification edge images and performing edge image weighted fusion on the denoised medical image dataset to be classified, thereby obtaining edge image intensity weights;

[0109] Step S15: perform edge difference calibration on each medical classification edge image according to the edge image intensity weight to obtain the key interest region of each image.

[0110] In an embodiment of the present invention, a set of medical images to be classified is collected, and the data set includes X-ray images and CT images, so as to ensure that the annotation information of these images is accurate and has a certain diversity, covering different pathological conditions and shooting conditions. Each image is appropriately preprocessed, including resizing, image format conversion, etc., to ensure the uniformity and availability of the data set. A 5×5 or 7×7 Gaussian filter is used to denoise each medical image to be classified. The Gaussian filter can effectively smooth the image and reduce the noise in the image, making the subsequent edge detection and feature extraction more stable and accurate. After Gaussian blurring, the noise of the image is effectively removed, making the details (such as edges, textures, etc.) more prominent, and the subsequent analysis will be more accurate. The denoised image is subjected to edge detection, and commonly used methods include Canny edge detection, Sobel operator, etc. Edge detection can extract areas with large changes in the image (i.e., significant changes in object boundaries or textures), providing key information for subsequent feature extraction. Canny edge detection process: Use Gaussian filtering to denoise the image (if not completed in S12), calculate the image gradient, obtain the intensity change of each pixel, perform non-maximum suppression on the gradient image, remove non-edge points, and generate an edge image through double threshold segmentation. The generated edge image clearly shows the object boundaries and key information in the image, making the subsequent edge weighted fusion and key interest area extraction more effective. Extract the edge response value of each pixel from the edge image. The edge response value is usually expressed as the gradient value or edge intensity of the image at that position, indicating the area with the largest change in the image (i.e., the edge). For each denoised image, weighted fusion is performed according to the edge response value obtained by edge detection. The purpose of the weighted process is to give higher weights to the edge area of ​​the image, thereby highlighting the area with important structural features in the image. Use the edge response value as the weight coefficient and combine it with the pixel value of the original image for weighted processing. The fused image will show higher contrast in the edge area. Through weighted fusion, the area with important structural information in the image can be enhanced while retaining the detailed features of the original image. This method can effectively highlight the edges in the image and improve the accuracy of subsequent analysis. Through weighted fusion, the areas with important structural information in the image can be enhanced while retaining the detailed features of the original image. This method can effectively highlight the edges in the image and improve the accuracy of subsequent analysis. The edge image intensity of each image is thresholded and the areas with higher edge intensity are screened out. Combined with the edge image intensity weight, these high-intensity areas are finely calibrated, and the calibrated areas are the key areas of interest of the image. The calibrated key areas of interest usually contain important structures or lesion areas in the image. Subsequent classification or analysis will focus more on these key areas, improving classification accuracy and diagnostic capabilities.

[0111] As an example of the present invention, refer to Figure 2As shown, in this example, step S2 includes:

[0112] Step S21: Perform Fourier transform on the key interest region of the medical image data set to be classified to obtain a frequency domain conversion map of the key interest region;

[0113] Step S22: dividing the frequency domain conversion map of the key interest area based on the threshold value to obtain a high frequency part and a low frequency part of the frequency domain conversion map;

[0114] Step S23: performing fine-grained texture analysis on the high-frequency part and the low-frequency part of the frequency domain conversion image respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness;

[0115] Step S24: performing difference value calculation based on the fine-grained texture data of the high-frequency part and the fine-grained texture data of the low-frequency part, and selecting texture feature difference points of the key interest area as the first classification feature points through the calculated difference values.

[0116] In an embodiment of the present invention, the key interest area of ​​each image is Fourier transformed and converted to the frequency domain. Fourier transform can convert the image from the time domain to the frequency domain, thereby effectively separating the low-frequency and high-frequency information in the image and helping to understand the global structure and detail features of the image. After Fourier transform, the frequency domain conversion map of the image is presented in a complex form, and the high-frequency and low-frequency parts can be extracted, corresponding to the detail information and global information of the image respectively. The frequency domain conversion map is divided based on the threshold, and the low-frequency part and the high-frequency part in the frequency domain conversion map are separated. The low-frequency part contains the smooth information and large-scale structure of the image, while the high-frequency part contains details, textures and edge information. The amplitude spectrum of the frequency domain conversion map is calculated, and a suitable threshold is selected according to the amplitude of the spectrum to separate the low-frequency part (usually concentrated in the center of the spectrum) from the high-frequency part (periphery of the spectrum). Through this step, the frequency domain conversion map is effectively decomposed into high-frequency and low-frequency parts, which provides a basis for subsequent fine-grained texture analysis and difference value calculation. Fine-grained texture analysis is performed on the high-frequency and low-frequency parts in the frequency domain, respectively, and the main purpose is to analyze the texture structure and characteristics of these areas. Commonly used texture analysis methods include gray level co-occurrence matrix (GLCM), local binary pattern (LBP), texture spectrum analysis, etc. The high-frequency part usually contains the details and texture information of the image. Analyzing the texture of these areas can help identify small lesions or texture changes in the image. Use texture feature extraction methods (such as GLCM and LBP) to extract fine-grained texture features in the high-frequency area. The low-frequency part usually describes the large-scale structural features of the image. Analyzing the texture of these areas can help identify the basic structure and main contours in the image. Through texture analysis, the texture fine-grained data of the high-frequency and low-frequency parts are obtained, which respectively describe the texture features of the details and the overall structure in the image, and provide texture information for subsequent difference calculation and feature screening. According to the high-frequency texture fine-grained data and the low-frequency texture fine-grained data obtained in step S23, the difference value between them is calculated. The difference value can be calculated in a variety of ways, such as Euclidean distance, Pearson correlation coefficient or other appropriate difference measurement methods. According to the calculated difference value, the area with large difference is screened as the texture feature difference point. These difference points usually correspond to key areas or abnormal areas in the image, which contain important lesion information. Sort the difference values ​​and select the top N points with the largest differences as texture feature difference points. Use these texture feature difference points as the first classification feature points and as input features for subsequent classification analysis. By calculating the texture differences between high-frequency and low-frequency parts, the most recognizable texture feature difference points in the image can be screened out. These points can help the model accurately identify important features in the image and provide key information for subsequent image classification.

[0117] Preferably, step S22 includes the following steps:

[0118] A 2D directional filter is used to calculate the angular distribution of the high-frequency part and the low-frequency part of the frequency domain conversion diagram in the frequency domain direction, and the frequency domain angular distribution of the high-frequency part and the frequency domain angular distribution of the low-frequency part are obtained. The calculation formula is as follows:

[0119]

[0120] in, Represents the direction angle of each point in the frequency domain, is the complex frequency component in the frequency domain, is the frequency component of the image in the horizontal direction, is the frequency component of the image in the vertical direction;

[0121] The Gabor filter is used to analyze the texture information of the high-frequency part and the low-frequency part of the frequency domain conversion image, and the texture information of the high-frequency part and the texture information of the low-frequency part are obtained. The calculation formula is as follows:

[0122]

[0123] in, Represents texture information, is the number of sampling points in the region, is the regional mean, For the The complex frequency components in the frequency domain, For the The frequency component of an image in the horizontal direction, For the The frequency component of an image in the vertical direction;

[0124] The frequency domain angle distribution of the high frequency part and the texture information of the high frequency part are integrated into the high frequency texture fine granularity, and the frequency domain angle distribution of the low frequency part and the texture information of the low frequency part are integrated into the low frequency texture fine granularity.

[0125] In the embodiment of the present invention, the angular distribution of the high frequency part and the low frequency part of the frequency domain conversion image is calculated by using a 2D directional filter. The specific steps are as follows: Fourier transform the frequency domain conversion image to obtain the complex frequency component , calculate the direction angle of each frequency domain point Use the formula: ;in, Represents the direction angle of each point in the frequency domain, is the complex frequency component in the frequency domain, is the frequency component of the image in the horizontal direction, is the frequency component of the image in the vertical direction. Use Gabor filter to analyze the texture information of the high-frequency and low-frequency parts of the frequency domain conversion image. Gabor filter can extract texture features of different scales and directions in the frequency domain. The calculation formula is: in, Represents texture information, is the number of sampling points in the region, is the regional mean, For the The complex frequency components in the frequency domain, For the The frequency component of an image in the horizontal direction, For the The frequency components of an image in the vertical direction. The frequency domain angle distribution of the high-frequency part is integrated with the texture information of the high-frequency part to generate high-frequency texture fine-grainedness. Similarly, the frequency domain angle distribution of the low-frequency part is integrated with the texture information of the low-frequency part to generate low-frequency texture fine-grainedness. This can be achieved through a specific fusion algorithm (such as weighted averaging or feature fusion). Finally, by combining the angle distribution and texture information of the high-frequency and low-frequency parts, a set of high-frequency texture fine-grainedness and low-frequency texture fine-grainedness can be obtained for subsequent analysis. These fine-grained features can effectively describe the different frequency domain texture characteristics of the image.

[0126] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0127] Step S31: Based on the preset medical pathology image library, the key interest area of ​​each image is screened for abnormal areas, and the contrast difference between the lesion area and the normal tissue is analyzed to obtain local contrast difference data, wherein the calculation formula of the contrast difference is as follows:

[0128]

[0129] in, For images at position The local area difference value, For images at position The gradient at For images at position The gray value of is the horizontal coordinate of the image, is the image ordinate;

[0130] Step S32: Calculate the global grayscale mean and standard deviation of the medical image data set to be classified to obtain the global grayscale mean and the global grayscale standard deviation of the image;

[0131] Step S33: Analyze the overall brightness, contrast and grayscale distribution characteristics of the medical image data set to be classified using the global grayscale mean and the global grayscale standard deviation of the image to obtain global contrast difference data.

[0132] In the embodiment of the present invention, each image is segmented according to predefined key regions of interest (such as tumors, lesions, etc.) to determine the contrast difference between these regions and the surrounding normal tissues. The given formula is used to calculate the contrast difference between each position The local contrast difference at: ;in, For images at position The local area difference value, For images at position The gradient at For images at position The gray value of is the horizontal coordinate of the image, is the image ordinate; the calculation of contrast difference data can help identify the obvious difference between the lesion area and normal tissue, and assist the classification model in identifying the lesion area. The global grayscale value is calculated for the entire image data set. The grayscale mean and standard deviation can reveal the changes in the overall brightness. The grayscale mean is the average of the grayscale values ​​of all pixels in the image, and the grayscale standard deviation measures the discrete degree of the image grayscale value, reflecting the overall contrast of the image. For each image, its global grayscale mean and standard deviation are calculated to obtain the global brightness and contrast features of the image. Using the global grayscale mean and standard deviation, the overall brightness, contrast and grayscale distribution features of each image are analyzed. Through these features, the global contrast difference data of the image can be extracted to further help distinguish between lesions and normal tissue areas. This global feature can be combined with local features to improve the model performance of image classification, segmentation and abnormal area detection.

[0133] Preferably, designing an adaptive image enhancement function using local contrast difference data and global contrast difference data in step S4 comprises the following steps:

[0134] Set the initial image enhancement function, where the initial image enhancement function is as follows:

[0135]

[0136] in, is the initial image enhancement function, is the original image, To control the weighting coefficient of local difference on the enhancement effect, To control the weighting coefficient of the global difference on the enhancement effect, is the local contrast, is the global contrast;

[0137] Introducing adaptive coefficients The initial image enhancement function is adjusted for low-contrast region intensity, and the adaptive coefficient is set to 1.2~2.0 to obtain the low-contrast region intensity adjustment parameter, where the formula for low-contrast region intensity adjustment is as follows:

[0138]

[0139] in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect;

[0140] Introducing adaptive coefficients The initial image enhancement function is adjusted for high contrast region intensity, and the adaptive coefficient is set to 0.5-1.5 to obtain the high contrast region intensity adjustment parameter, where the formula for high contrast region intensity adjustment is as follows:

[0141]

[0142] in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect;

[0143] The initial image enhancement function is adaptively selected using the low-contrast region intensity adjustment parameter and the high-contrast region intensity adjustment parameter to obtain an adaptive image enhancement function.

[0144] In the embodiment of the present invention, a basic image enhancement function is defined to adjust the local and global contrast of the image. The formula is as follows: ;in, is the initial image enhancement function, is the original image, To control the weighting coefficient of local difference on the enhancement effect, To control the weighting coefficient of the global difference on the enhancement effect, is the local contrast, is the global contrast. This initial enhancement function adjusts the enhancement effect based on the local and global contrast of the image. In order to enhance the performance of low-contrast areas, an adaptive coefficient is introduced. , which ranges from 1.2 to 2.0. This coefficient is used to adjust the enhancement strength according to the characteristics of low-contrast areas. The formula is: in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, It is a weighting factor to control the global difference on the enhancement effect. Similarly, in order to enhance the high contrast area more appropriately, an adaptive coefficient is defined. , whose value range is 0.5 to 1.5. This coefficient adjusts the enhancement strength of high contrast areas in the image. The formula is: ;in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, It is the weighted coefficient that controls the global difference on the enhancement effect. Finally, through the low contrast area intensity adjustment parameters and the high contrast area intensity adjustment parameters, the appropriate enhancement function is automatically selected according to the characteristics of different areas in the image. This step makes the enhancement effect of different areas more accurate through the adaptive adjustment of local contrast and global contrast. Finally, an adaptive image enhancement function is obtained.

[0145] Preferably, the adaptive selection of the initial image enhancement function using the low contrast region intensity adjustment parameter and the high contrast region intensity adjustment parameter comprises:

[0146] If the local contrast If it is less than 0.5, the low contrast area intensity adjustment function is used. ;

[0147] If the local contrast If it is greater than 1.5, the low contrast area intensity adjustment function is used. ;

[0148] If the local contrast In the range of 0.5 ~ 1.5, the initial image enhancement function is used .

[0149] In the embodiment of the present invention, by Different values ​​of use different image enhancement functions to optimize each area of ​​the image: when the local contrast When it is less than 0.5, the image belongs to the low contrast area. In this case, the intensity adjustment function of the low contrast area is used. To enhance the image. This function uses the adaptive coefficient Adjusts the enhancement effect of low contrast areas. When it is greater than 1.5, the image belongs to the high contrast area. In this case, the intensity adjustment function of the high contrast area is used. To enhance the image. This function uses the adaptive coefficient Adjusts the enhancement effect of high contrast areas. When the value is between 0.5 and 1.5, the image belongs to the normal contrast area. At this time, the initial image enhancement function is used. , no special adjustments are made, and the original enhancement effect is maintained. The initial image enhancement function provides standard enhancement effects based on local and global contrast.

[0150] Preferably, step S5 comprises the following steps:

[0151] Step S51: constructing a gray level co-occurrence matrix of the key interest region after the enhanced features based on the first classification feature points, to obtain a gray level co-occurrence matrix of the key interest region after the enhanced features;

[0152] Step S52: Analyze the texture distribution of the key interest region after the enhanced features by using the gray level co-occurrence matrix of the key interest region after the enhanced features to obtain the texture distribution;

[0153] Step S53: Calculate the pixel complexity of the key interest area after the enhanced features according to the texture distribution, and construct a texture change gradient image of the key interest area after the enhanced features through the pixel complexity;

[0154] Step S54: For each region of interest, the frequency of the change of the texture change gradient image is calculated to obtain the texture wrinkle density of the key region of interest and use the texture wrinkle density as the second classification feature point.

[0155] In the embodiment of the present invention, it is particularly important that step S52 specifically further includes the following steps: step S521: using the grayscale co-occurrence matrix of the key interest region after the enhanced feature to perform grayscale pair co-occurrence probability analysis on the key interest region after the enhanced feature, and generate grayscale pair contribution probability data; step S522: extracting advanced texture features from the key interest region after the enhanced feature according to the grayscale pair contribution probability data, extracting contrast, homogeneity, energy, correlation, entropy, local uniformity and inverse difference matrix, and constructing a high-dimensional texture feature vector; step S523: quantifying the complexity and self-similarity of the regional texture of the high-dimensional texture feature vector, and obtaining the complexity measurement data of the regional texture; step S524: using the complexity measurement data of the regional texture to perform energy distribution modeling on the key interest region after the enhanced feature, and generate the distribution feature data of the local and global texture; step S525: performing regional spatial frequency domain mapping on the key interest region after the enhanced feature based on the distribution feature data of the local and global texture, and obtaining the texture distribution. Specifically, the image processing technology (such as corner detection, edge detection, regional growth, etc.) first determines the key interest region in the image. These regions usually contain rich textures and important visual information. Select the enhanced feature regions of the image (such as the enhanced image parts) and calculate the gray level co-occurrence matrix (GLCM) in these regions. The gray level co-occurrence matrix reflects the texture information of the image by calculating the gray level combination of pixel pairs in the image at a specific direction and distance.

[0156] Set the direction (such as horizontal, vertical, diagonal direction) and distance (usually choose 1 pixel unit or greater distance) to construct the co-occurrence matrix. Through the above steps, the gray level co-occurrence matrix of each key interest area is finally obtained, reflecting the texture characteristics of the area. A series of texture features are extracted from the constructed gray level co-occurrence matrix, such as: features that measure the gray level difference of the image, features that reflect the uniformity of the texture, features that reflect the uniformity of the texture, which is usually related to the smoothness of the image, and the complexity of the image texture. Through the analysis of the above features, the distribution characteristics of the texture in the area can be understood, and different texture patterns (such as regular texture, random texture, etc.) can be identified. Based on the texture distribution (such as contrast, energy, etc.), the pixel complexity of the area is calculated. Pixel complexity reflects the complexity of the details in the image area. Generally speaking, areas with high contrast and high entropy values ​​indicate higher complexity. Based on the calculated pixel complexity, a texture change gradient image is generated. This image reflects the rate and direction of texture change in the image area. By calculating the gradient of each pixel, the texture change in different areas of the image can be obtained. A common method is to use the Sobel operator or other edge detection algorithms to extract the gradient of the texture. Analyze the texture change gradient image and calculate the frequency of texture changes in the image. Areas with high frequencies usually indicate that the texture changes in the area are more drastic and contain more complex textures or object structures. The frequency of texture changes can be analyzed by methods such as Fourier transform, or calculated by counting the frequency of changes in gradient values ​​in the image. Texture wrinkle density reflects the relationship between the complexity of details in the image and the frequency of changes. High-frequency change areas usually correspond to areas with rapid changes or wrinkles in the texture. The calculated texture wrinkle density is used as the second classification feature point for subsequent image classification or analysis. This feature point can help distinguish different types of texture patterns or be used for further texture recognition tasks.

[0157] Preferably, step S6 comprises the following steps:

[0158] Step S61: constructing a feature vector for an image in the medical image dataset to be classified by using the first classification feature point and the second classification feature point to obtain a classification feature vector;

[0159] Step S62: training a classification model based on the classification feature vector, and performing image classification on the medical image data set to be classified according to the trained classification model, thereby obtaining classification results of X-ray medical images and CT medical images.

[0160] In an embodiment of the present invention, by selecting the first classification feature point and the second classification feature point, for each medical image, the required feature points are first extracted by applying edge detection, texture analysis, image segmentation and other algorithms. A deep learning model such as a convolutional neural network (CNN) is used to extract feature points from the image and convert them into feature vectors. The extracted feature points are combined into a high-dimensional feature vector to represent the classification information of the image. Each feature vector is normalized to ensure that each feature dimension has a similar scale to prevent certain features from having too much influence on the classification model. According to the complexity of the task, a suitable classification algorithm is selected, such as a support vector machine (SVM), a random forest (RF), or a deep learning model (such as a convolutional neural network CNN or a transfer learning model). The constructed feature vector is input into the classification model for training. During the training process, the model maximizes the classification accuracy by adjusting parameters multiple times, and learns how to distinguish between X-ray images and CT images. A cross-validation method can be used to evaluate the performance of the model to avoid overfitting. The trained classification model is used to classify the medical image data set to be classified. For each medical image to be classified, its category (X-ray image or CT image) is predicted based on the extracted classification feature vector. The classification results are output into two categories: X-ray medical images and CT medical images. An image classification report can be generated to display evaluation indicators such as classification accuracy and recall rate.

[0161] In this specification, an adaptive image classification system for a medical image dataset is provided, which is used to execute the adaptive image classification method for a medical image dataset. The adaptive image classification system for a medical image dataset includes:

[0162] An image acquisition module is used to obtain a medical image dataset to be classified and extract key interest regions of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images;

[0163] A fine-grained extraction module is used to perform frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyze the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; and integrate the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point;

[0164] A difference analysis module is used to calculate the local and global contrast differences of the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data;

[0165] A feature enhancement module is used to design an adaptive image enhancement function using local contrast difference data and global contrast difference data, and enhance the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features;

[0166] A wrinkle extraction module is used to analyze the texture wrinkle density of the key interest area after the enhanced feature based on the first classification feature point, and use the calculated texture wrinkle density of the key interest area as the second classification feature point;

[0167] The classification module is used to classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain the classification results of the X-ray medical image and the CT medical image.

[0168] The beneficial effect of the present invention is that the key interest region of each image can be extracted from the medical image data set to be classified through the image acquisition module, which lays the foundation for subsequent fine-grained extraction, difference analysis and other processing. Through accurate interest region extraction, it is ensured that subsequent analysis and feature extraction only focus on the most representative part of the image, thereby improving overall efficiency. Through frequency domain conversion and texture fine-grained analysis, the module can accurately capture the texture features of each medical image, especially the detailed information of the high-frequency and low-frequency parts. This helps to capture subtle changes in the image, especially the lesion area in medical images. Local and global contrast difference calculation helps the system to deeply understand the difference between the lesion area and the normal area in the image. Contrast difference is one of the key features for identifying lesions and healthy areas, and can effectively reveal the contrast changes of the image. Through the adaptive image enhancement function, the image features are dynamically adjusted according to the local and global contrast differences of the image, so that the important areas (such as lesion areas) in the image are more prominent after enhancement. In this way, the system can extract key image features more accurately. Texture wrinkle density analysis can reveal subtle texture changes in the image, which are usually related to the presence of lesion areas. By calculating the density of texture wrinkles, the system can further refine the image features and help the classification model better distinguish different types of medical images. Based on the extracted first classification feature points (high-frequency and low-frequency texture fine-grainedness) and second classification feature points (texture wrinkle density), the classification module can efficiently and accurately classify X-ray images and CT images. Through machine learning or deep learning models, the system can identify type differences in images to ensure the efficiency and accuracy of classification results. Therefore, the present invention improves the accuracy and adaptability of classification of different types of medical images through adaptive image enhancement, texture fine-grained analysis and contrast difference calculation.

[0169] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0170] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those 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 should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. An adaptive image classification method for medical image datasets, characterized in that: The following steps are involved: Step S1: obtaining a medical image dataset to be classified and extracting a key region of interest of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images; Step S2: performing frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyzing the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; integrating the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point; Step S3: Based on the preset medical pathology image library, local and global contrast difference calculations are performed on the key interest area of ​​each image to obtain local contrast difference data and global contrast difference data; Step S2 includes the following steps: Step S21: Perform Fourier transform on the key interest region of the medical image data set to be classified to obtain a frequency domain conversion map of the key interest region; Step S22: dividing the frequency domain conversion map of the key interest area based on the threshold value to obtain the high frequency part and the low frequency part of the frequency domain conversion map; Step S22 includes the following steps: A 2D directional filter is used to calculate the angular distribution of the high-frequency part and the low-frequency part of the frequency domain conversion diagram in the frequency domain direction, and the frequency domain angular distribution of the high-frequency part and the frequency domain angular distribution of the low-frequency part are obtained. The calculation formula is as follows: in, Represents the direction angle of each point in the frequency domain, is the complex frequency component in the frequency domain, is the frequency component of the image in the horizontal direction, is the frequency component of the image in the vertical direction; The Gabor filter is used to analyze the texture information of the high-frequency part and the low-frequency part of the frequency domain conversion image, and the texture information of the high-frequency part and the texture information of the low-frequency part are obtained. The calculation formula is as follows: in, Represents texture information, is the number of sampling points in the region, is the regional mean, For the The complex frequency components in the frequency domain, For the The frequency component of an image in the horizontal direction, For the The frequency component of an image in the vertical direction; Integrate the frequency domain angle distribution of the high frequency part and the texture information of the high frequency part into high frequency texture fine granularity, and integrate the frequency domain angle distribution of the low frequency part and the texture information of the low frequency part into low frequency texture fine granularity; Step S23: performing fine-grained texture analysis on the high-frequency part and the low-frequency part of the frequency domain conversion image respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; Step S24: Calculating difference values ​​based on the fine-grained texture data of the high-frequency part and the fine-grained texture data of the low-frequency part, and selecting texture feature difference points of the key interest area as first classification feature points through the calculated difference values; Step S4: designing an adaptive image enhancement function using the local contrast difference data and the global contrast difference data, and enhancing the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features; Step S5: Analyze the texture fold density of the key interest area after the enhanced features based on the first classification feature points, and use the calculated texture fold density of the key interest area as the second classification feature points; Step S6: Classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain classification results of X-ray medical images and CT medical images.

2. The method for adaptive image classification of medical image datasets according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining a medical image dataset to be classified, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images; Step S12: setting a Gaussian filter kernel size of 5×5 or 7×7, performing Gaussian blur on each image in the medical image dataset to be classified, and obtaining a denoised medical image dataset to be classified; Step S13: performing edge detection on the denoised medical image data set to be classified to obtain a medical classification edge image; Step S14: extracting edge response values ​​of medical classification edge images and performing edge image weighted fusion on the denoised medical image dataset to be classified, thereby obtaining edge image intensity weights; Step S15: perform edge difference calibration on each medical classification edge image according to the edge image intensity weight to obtain the key interest region of each image.

3. The adaptive image classification method for medical image datasets according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Based on the preset medical pathology image library, the key interest area of ​​each image is screened for abnormal areas, and the contrast difference between the lesion area and the normal tissue is analyzed to obtain local contrast difference data, wherein the calculation formula of the contrast difference is as follows: in, For images at position The local area difference value, For images at position The gradient at For images at position The gray value of is the image horizontal coordinate, is the image ordinate; Step S32: Calculate the global grayscale mean and standard deviation of the medical image data set to be classified to obtain the global grayscale mean and the global grayscale standard deviation of the image; Step S33: Analyze the overall brightness, contrast and grayscale distribution characteristics of the medical image data set to be classified using the global grayscale mean and the global grayscale standard deviation of the image to obtain global contrast difference data.

4. The adaptive image classification method for medical image datasets according to claim 1, characterized in that: Designing an adaptive image enhancement function using local contrast difference data and global contrast difference data in step S4 includes the following steps: Set the initial image enhancement function, where the initial image enhancement function is as follows: in, is the initial image enhancement function, is the original image, To control the weighting coefficient of local difference on the enhancement effect, To control the weighting coefficient of the global difference on the enhancement effect, is the local contrast, is the global contrast; Introducing adaptive coefficients The initial image enhancement function is adjusted for low-contrast region intensity, and the adaptive coefficient is set to 1.2~2.0 to obtain the low-contrast region intensity adjustment parameter, where the formula for low-contrast region intensity adjustment is as follows: in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect; Introducing adaptive coefficients The initial image enhancement function is adjusted for high contrast region intensity, and the adaptive coefficient is set to 0.5-1.5 to obtain the high contrast region intensity adjustment parameter, where the formula for high contrast region intensity adjustment is as follows: in, Adjust the parameters for low contrast area intensity, is the adaptive coefficient, is the local contrast, is the global contrast, To control the weighting coefficient of global difference on the enhancement effect; The initial image enhancement function is adaptively selected using the low-contrast region intensity adjustment parameter and the high-contrast region intensity adjustment parameter to obtain an adaptive image enhancement function.

5. The method for adaptive image classification of medical image datasets according to claim 1, characterized in that: Adaptively selecting the initial image enhancement function using the low contrast region intensity adjustment parameter and the high contrast region intensity adjustment parameter includes: If the local contrast If it is less than 0.5, the low contrast area intensity adjustment function is used. ; If the local contrast If it is greater than 1.5, the low contrast area intensity adjustment function is used. ; If the local contrast In the range of 0.5 ~ 1.5, the initial image enhancement function is used .

6. The method for adaptive image classification of medical image datasets according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: constructing a gray level co-occurrence matrix of the key interest region after the enhanced features based on the first classification feature points, to obtain a gray level co-occurrence matrix of the key interest region after the enhanced features; Step S52: Analyze the texture distribution of the key interest region after the enhanced features by using the gray level co-occurrence matrix of the key interest region after the enhanced features to obtain the texture distribution; Step S53: Calculate the pixel complexity of the key interest area after the enhanced features according to the texture distribution, and construct a texture change gradient image of the key interest area after the enhanced features through the pixel complexity; Step S54: For each region of interest, the frequency of the change of the texture change gradient image is calculated to obtain the texture wrinkle density of the key region of interest and use the texture wrinkle density as the second classification feature point.

7. The method for adaptive image classification of medical image datasets according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: constructing a feature vector for an image in the medical image dataset to be classified by using the first classification feature point and the second classification feature point to obtain a classification feature vector; Step S62: training a classification model based on the classification feature vector, and performing image classification on the medical image data set to be classified according to the trained classification model, thereby obtaining classification results of X-ray medical images and CT medical images.

8. An adaptive image classification system for a medical image dataset, characterized in that: The method for performing the adaptive image classification of a medical image dataset according to claim 1, wherein the adaptive image classification system of the medical image dataset comprises: An image acquisition module is used to obtain a medical image dataset to be classified and extract key interest regions of each image, wherein the medical image dataset to be classified includes a number of X-ray medical images and CT medical images; A fine-grained extraction module is used to perform frequency domain conversion on the medical image dataset to be classified according to the key interest region, and analyze the texture fine-grainedness of each image in the medical image dataset to be classified respectively to obtain high-frequency texture fine-grainedness and low-frequency texture fine-grainedness; and integrate the high-frequency texture fine-grainedness and the low-frequency texture fine-grainedness as the first classification feature point; A difference analysis module is used to calculate the local and global contrast differences of the key interest area of ​​each image based on a preset medical pathology image library to obtain local contrast difference data and global contrast difference data; A feature enhancement module is used to design an adaptive image enhancement function using local contrast difference data and global contrast difference data, and enhance the image features of the key interest region of each image according to the adaptive image enhancement function to generate the key interest region after the enhanced features; A wrinkle extraction module is used to analyze the texture wrinkle density of the key interest area after the enhanced feature based on the first classification feature point, and use the calculated texture wrinkle density of the key interest area as the second classification feature point; The classification module is used to classify the images in the medical image data set to be classified by using the first classification feature point and the second classification feature point, so as to obtain the classification results of the X-ray medical image and the CT medical image.

Images