An automatic classification method for lymphoma pathological sections based on convolutional neural network

Abstract

The present invention relates to the field of medical image processing, and in particular to a method for automatic classification of lymphoma pathological sections based on a convolutional neural network. The content includes: collecting original slice images and preprocessing them to obtain slice images, and calculating image enhancement factors; calculating local contrast based on slice image pixel values; calculating adaptive enhancement factors based on image enhancement factors and local contrast; enhancing the slice images based on the adaptive enhancement factors to obtain enhanced images; inputting the enhanced images into a feature screening convolutional neural network for classification to obtain the final classification results. The method solves the problem that traditional image enhancement and classification technologies are difficult to make fine adjustments based on local features of images, and cannot effectively improve local contrast and details of images during the enhancement process; the traditional convolutional neural network has redundant extracted features during the training process, high computational complexity, and difficulty in improving classification accuracy.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a method for automatically classifying lymphoma pathological sections based on a convolutional neural network. Background Art

[0002] With the rapid development of computer vision and deep learning technologies, medical image analysis has gradually become an important auxiliary tool for modern medical diagnosis. Especially in the field of pathological images, convolutional neural networks (CNNs) have made significant progress in image classification and feature extraction, and are widely used in the automatic identification and diagnosis of various diseases. Lymphoma is a common malignant tumor, and early diagnosis is crucial to the patient's treatment plan. However, due to the complexity and diversity of lymphoma pathological section images, manual analysis is not only labor-intensive, but also has a high risk of misdiagnosis. With the development of deep learning algorithms, automated lymphoma pathological section classification methods have emerged, which has promoted the intelligent and precise direction of pathological image analysis.

[0003] However, traditional image enhancement and classification technologies still face some challenges: first, image enhancement methods often have too uniform enhancement effects on different regions, making it difficult to make fine adjustments based on the local features of the image, resulting in loss of image details or uneven enhancement effects; second, local contrast and details of the image are often not effectively improved during the enhancement process, especially in low-contrast areas, where the enhancement effect is not obvious, affecting the accuracy of subsequent classification tasks; in addition, during the training process of traditional convolutional neural networks, the extracted features are redundant, resulting in high computational complexity and difficulty in further improving classification accuracy. Therefore, it is urgent to improve it through a new algorithm and processing technology. Summary of the invention

[0004] The present invention provides a method for automatically classifying lymphoma pathological sections based on a convolutional neural network, so as to solve the problems that the image enhancement method in the traditional image enhancement and classification technology often has too uniform enhancement effects on different regions, and it is difficult to make fine adjustments according to the local features of the image, resulting in loss of image details or uneven enhancement effects; the local contrast and details of the image are often unable to be effectively improved during the enhancement process, especially in low-contrast areas, the enhancement effect is not obvious, affecting the accuracy of subsequent classification tasks; the traditional convolutional neural network has redundant features extracted during the training process, resulting in high computational complexity and difficulty in further improving the classification accuracy.

[0005] The present invention provides a method for automatically classifying lymphoma pathological sections based on a convolutional neural network, which specifically includes the following technical solutions:

[0006] A method for automatic classification of lymphoma pathological sections based on convolutional neural network, comprising the following steps:

[0007] S1: Collect the original slice image and preprocess it to obtain the slice image; based on the slice image pixel value, introduce the adaptive enhancement algorithm to calculate the image enhancement factor;

[0008] S2: Calculate the local contrast based on the slice image pixel value; Calculate the adaptive enhancement factor based on the image enhancement factor and the local contrast; Enhance the slice image based on the adaptive enhancement factor to obtain an enhanced image;

[0009] S3: Input the enhanced image into the feature screening convolutional neural network for classification to obtain the final classification result.

[0010] Preferably, the S1 specifically includes:

[0011] The slice image is divided into regions, and the pixel value of the image is dynamically adjusted in an adaptive manner through an adaptive enhancement algorithm combined with the gradient amplitude, local texture characteristics and local contrast of the slice image.

[0012] Preferably, the S1 specifically includes:

[0013] In the implementation process of the adaptive enhancement algorithm, the image enhancement factor is calculated based on the gradient amplitude and local texture features of the slice image. The specific formula is:

[0014] ,

[0015] in, is in position Image enhancement factor at ; are the pixel coordinates of the slice image; is the image area The total number of pixels in ; Represents the image area The pixel position in ; Represents the image area; are the pixel coordinates of the image region; is the slice image at position The pixel value at ; is the slice image at position The gradient along the horizontal direction; is the slice image at position The gradient in the vertical direction; is the slice image at position The gradient magnitude at ; is the slice image at position Local texture features at is the image area The maximum value of all texture features in the slice image.

[0016] Preferably, the S2 specifically includes:

[0017] The local contrast is calculated based on the difference between each pixel in the slice image and its adjacent pixels, combined with the pixel standard deviation of the image region.

[0018] Preferably, the S2 specifically includes:

[0019] The local contrast weighting method is introduced to adaptively adjust the image enhancement factor to obtain the adaptive enhancement factor.

[0020] Preferably, the S2 specifically includes:

[0021] The adaptive enhancement factor is applied to the slice image pixel value after normalization, and the slice image is enhanced in combination with the local contrast to obtain an enhanced image. The enhancement adjustment formula is as follows:

[0022] ,

[0023] in, Indicates at location The pixel value of the enhanced image; Indicates that the slice image is at position The pixel value at ; is the adaptive enhancement factor; It is the sum of the adaptive enhancement factors of all pixels; is the adjustment factor; is the slice image at position The local contrast at .

[0024] Preferably, the S3 specifically includes:

[0025] In the feature screening convolutional neural network, the enhanced image is convolved to obtain a preliminary feature matrix; the preliminary feature matrix is ​​divided to generate sub-matrices; a feature screening mechanism is introduced to calculate the score value of each sub-matrix based on the matrix determinant, condition number and Frobenius norm, and the score values ​​of all sub-matrices are sorted in descending order to screen out optimized features.

[0026] Preferably, the S3 specifically includes:

[0027] The optimized features are used as new inputs, and features are extracted through convolution operations to obtain high-level features. The high-level features are processed through a fully connected layer, and the probability value of each category is output through a normalized exponential function. The category with the highest probability is selected as the final classification result.

[0028] The beneficial effects of the technical solution of the present invention are:

[0029] 1. The present invention uses an adaptive enhancement algorithm to dynamically adjust the image enhancement factor of each area based on the gradient amplitude, local texture features and local contrast of the slice image, thereby specifically improving the details and contrast of each area of ​​the slice image, avoiding the blind enhancement problem of traditional methods, and significantly improving the detail performance of the slice image.

[0030] 2. The present invention introduces a method of combining local contrast calculation with image enhancement factors to ensure that the slice image can be effectively enhanced in low-contrast areas, so that the details in the slice image can be better preserved, thereby improving the accuracy of subsequent classification tasks and providing clearer feature support for small lesions and details that are difficult to identify in pathological slice images.

[0031] 3. The present invention introduces a feature screening mechanism to effectively remove redundant features in the feature extraction stage of the feature screening convolutional neural network, reduce the computational complexity, and improve the efficiency of model training and classification accuracy, especially when processing complex lymphoma pathological section images, thereby improving the robustness and stability of the feature screening convolutional neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of a method for automatic classification of lymphoma pathological sections based on a convolutional neural network described in the present invention. DETAILED DESCRIPTION

[0033] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0034] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0035] The following is a detailed description of a method for automatic classification of lymphoma pathological sections based on a convolutional neural network provided by the present invention in conjunction with the accompanying drawings.

[0036] Refer to the attached Figure 1 , which shows a flow chart of a method for automatic classification of lymphoma pathological sections based on a convolutional neural network provided by an embodiment of the present invention, the method comprising the following steps:

[0037] S1: Collect the original slice image and perform preprocessing to obtain the slice image; based on the slice image pixel value, introduce the adaptive enhancement algorithm to calculate the image enhancement factor;

[0038] The original slice images are collected and preprocessed. In the data preprocessing stage, the original slice images are denoised and scaled to obtain slice images. The specific implementation process of the preprocessing stage is as follows: first, the original slice images are denoised by using methods such as Gaussian filtering or median filtering to remove the noise in the original slice images to obtain denoised original slice images; all denoised original slice images are scaled to ensure that the denoised original slice images have consistent size and resolution, thereby improving the efficiency and accuracy of subsequent processing. The preprocessing is a well-known method for those skilled in the art and will not be described in detail here.

[0039] The slice image is divided into regions according to the expert experience method, and the pixel value of each region is adjusted by the adaptive enhancement algorithm to highlight the characteristics of the slice image. The adaptive enhancement algorithm is an enhancement technology based on the local characteristics of the image. It combines the gradient amplitude, local texture characteristics and local contrast of the slice image, and dynamically adjusts the pixel value of the image in an adaptive manner to highlight the details and contrast of the region.

[0040] In the adaptive enhancement algorithm, the image enhancement factor is calculated based on the gradient amplitude and local texture features of the slice image. The image enhancement factor indicates the degree to which a certain position in the slice image needs to be enhanced. For each divided area, the image enhancement factor is calculated using the following formula:

[0041] ,

[0042] in, is the image enhancement factor, indicating that the slice image is at position The degree to which the place needs to be strengthened; It is the pixel coordinate of the slice image, which indicates the horizontal and vertical coordinates of a specific position in the slice image; is the image area The total number of pixels in ; Represents the image area A pixel position in ; Represents the image area, which is divided according to the expert experience method; are the pixel coordinates of the image region; is the slice image at position The pixel value at ; is the slice image at position The gradient along the horizontal direction reflects the degree of change of the slice image in the horizontal direction; is the slice image at position The gradient along the vertical direction reflects the degree of change of the slice image in the vertical direction; is the slice image at position The gradient magnitude at , indicating the position The degree of change at is calculated as ; is the slice image at position The local texture features at the location reflect the position The texture features at the locations are calculated using existing techniques such as the Gray Level Co-occurrence Matrix (GLCM); is the image area The maximum value of all texture features in the slice image is used to normalize the texture features;

[0043] The image enhancement factor comprehensively considers the gradient amplitude and local texture features of the slice image, reflects the degree of change of the slice image in the horizontal and vertical directions through the gradient amplitude, combines the local texture features (such as the texture features calculated by the gray-level co-occurrence matrix), and performs maximum value normalization processing. It can more accurately enhance the detail contrast of the slice image, while avoiding excessive enhancement of flat areas, achieving a more balanced and effective image enhancement effect, and improving the stability and recognition of the slice image in subsequent processing.

[0044] S2: Calculate the local contrast based on the slice image pixel value; Calculate the adaptive enhancement factor based on the image enhancement factor and the local contrast; Enhance the slice image based on the adaptive enhancement factor to obtain an enhanced image;

[0045] In order to enhance the detail contrast of the slice image, the local standard deviation is used as an indicator to measure the local contrast. The local contrast effectively reflects the change of pixel values ​​in the local area of ​​the slice image by measuring the difference between each pixel and adjacent pixels and combining the pixel standard deviation of the image area, so as to capture the change of image details and provide support for subsequent enhancement adjustments.

[0046] The local contrast calculation formula is as follows:

[0047] ,

[0048] in, is the slice image at position The local contrast at is the offset in pixels; is the slice image at position The pixel value at ; is the slice image at position The pixel value at ; is the pixel standard deviation of the image area, The image area The standard deviation of the pixel values ​​within an image is used to measure the degree of variation of pixel values ​​within the image area.

[0049] The adaptive enhancement algorithm calculates the adaptive enhancement factor based on the image enhancement factor and the local contrast, and adaptively adjusts the image enhancement factor by introducing the local contrast for weighting to obtain the adaptive enhancement factor, thereby dynamically adjusting the enhancement effect of each pixel. The adaptive enhancement factor calculates a targeted enhancement value for each pixel by introducing the local contrast, thereby applying different degrees of image enhancement in different image areas, which can more accurately optimize the contrast and detail performance of the slice image and improve the image quality;

[0050] The calculation formula of the adaptive enhancement factor is:

[0051] ,

[0052] in, is the adaptive enhancement factor; is the image enhancement factor; It is the sum of the image enhancement factors of all pixels and is used for normalization; and is a constant used to control the effect of local contrast, and is set according to expert experience;

[0053] After calculating the adaptive enhancement factor, the adaptive enhancement factor is applied to the slice image pixel value after normalization to enhance the detail contrast of the slice image. Finally, the adaptive enhancement factor is combined with the local contrast to optimize the detail performance of the slice image, obtain an enhanced image, and improve the subsequent classification effect;

[0054] The enhancement adjustment formula is as follows:

[0055] ,

[0056] in, Indicates at location The pixel value of the enhanced image; Indicates that the slice image is at position The pixel value at ; is the adaptive enhancement factor; It is the sum of the adaptive enhancement factors of all pixels, which is used for normalization; It is an adjustment coefficient used to control the effect of local contrast on the final image enhancement effect and is set according to expert experience; is the slice image at position The local contrast at .

[0057] By combining the slice image pixel value, adaptive enhancement factor and local contrast, the slice image is enhanced to obtain an enhanced image. The adaptive enhancement factor determines the degree of enhancement of each image area according to the gradient amplitude and local texture characteristics of the slice image, while the local contrast helps to highlight the details. By adjusting the coefficient, the influence of the local contrast on the overall effect of the image can be accurately adjusted, thereby maintaining the visual balance of the image while enhancing the detail performance. Ultimately, the enhanced image is significantly improved in terms of detail, edge and texture presentation, which is helpful for subsequent classification tasks.

[0058] S3: Input the enhanced image into the feature screening convolutional neural network for classification to obtain the final classification result.

[0059] The enhanced image is input into a feature screening convolutional neural network for classification. The feature screening convolutional neural network is based on the traditional convolutional neural network and adds a feature screening operation. The traditional convolutional neural network usually extracts the preliminary feature matrix in the convolution layer and directly uses these features for subsequent classification tasks. However, when processing medical images, the feature matrix extracted by the convolutional neural network may contain a large number of redundant and irrelevant features, resulting in poor classification results and reduced computational efficiency. In order to solve this problem, a feature screening mechanism is introduced. Based on the matrix determinant, condition number and Frobenius norm, the preliminary feature matrix is ​​scored and screened to retain features with strong discriminative ability, thereby improving classification accuracy and computational efficiency.

[0060] Before image classification, it is necessary to first prepare a data set for training. Specifically, the original lymphoma pathology section images for training are selected according to the expert experience method, and corresponding classification labels are added to each original lymphoma pathology section image. The original lymphoma pathology section images can be obtained through the hospital pathology department or medical imaging platform, including different subtypes of lymphoma and normal tissue. The original lymphoma pathology section images used for training are subjected to the above-mentioned preprocessing operations and adaptive enhancement algorithms in sequence to obtain enhanced images for training as training sets, and the training sets are input into the feature screening convolutional neural network for training, and the weights and biases of the feature screening convolutional neural network are adjusted to identify and classify lymphoma images and normal tissue images. The training process of the feature screening convolutional neural network adopts existing technology and will not be repeated here.

[0061] The specific operation process of the feature screening convolutional neural network is as follows:

[0062] First, the enhanced image is convolved to obtain a preliminary feature matrix. Since the preliminary feature matrix extracted by the feature screening convolutional neural network contains a large amount of information and some features of the preliminary feature matrix contribute little to the classification task, it is necessary to perform feature screening on the preliminary feature matrix to remove redundant features and retain features with strong discriminative ability.

[0063] Specifically, the window size is set according to the expert experience method, and the preliminary feature matrix is ​​divided according to the set window size to generate multiple sub-matrices. The matrix determinant, condition number and Frobenius norm are introduced as screening criteria, and the scoring formula is used to calculate the score value of each sub-matrix;

[0064] The scoring formula is as follows:

[0065] ,

[0066] in, is the score value of the submatrix, which is used to measure the submatrix The amount and stability of information; It is The logarithm of the determinant of the submatrices; is the index variable of the submatrix; is a submatrix The determinant of is used to reflect the submatrix Global feature information of is a submatrix The condition number is used to measure the numerical stability of the matrix; is a submatrix The Frobenius norm of ;

[0067] After completing the scoring calculation of the submatrices, sort the scoring values ​​of all submatrices in descending order, and set the number of submatrices to be screened according to the expert experience method. Select several submatrices with the highest scores from the sorted submatrix list as optimization features. Optimization features can effectively retain key information and remove redundant features, thereby providing more stable input for subsequent classification tasks and improving the overall classification accuracy and robustness of the model;

[0068] The optimized features are used as new inputs, and further features are extracted through convolution operations to obtain high-level features, which are used to further mine high-level and complex features.

[0069] Finally, the extracted high-level features are processed through a fully connected layer, and the probability value of each category is output through a normalized exponential function (softmax function), and the category with the highest probability is selected as the final classification result.

[0070] After the feature screening convolutional neural network training is completed, the enhanced image to be classified is input into the trained feature screening convolutional neural network, and the category with the highest probability is selected as the final classification result. The lymphoma pathological sections are automatically classified, thereby realizing the automatic recognition and diagnosis of lymphoma pathological sections and providing auxiliary decision support for clinicians.

[0071] In summary, a convolutional neural network-based automatic classification method for lymphoma pathological sections was completed.

[0072] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0073] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for automatic classification of lymphoma pathological sections based on convolutional neural network, characterized in that: The following steps are involved: S1: collect original slice images and preprocess them to obtain slice images; Based on the pixel value of the slice image, an adaptive enhancement algorithm is introduced, and the image enhancement factor is calculated by combining the gradient amplitude and local texture features of the slice image. The specific formula of the image enhancement factor is: ， in, is in position Image enhancement factor at ; are the pixel coordinates of the slice image; is the image area The total number of pixels in ; Represents the image area The pixel position in ; Represents the image area; are the pixel coordinates of the image region; is the slice image at position The pixel value at ; is the slice image at position The gradient along the horizontal direction; is the slice image at position The gradient in the vertical direction; is the slice image at position The gradient magnitude at ; is the slice image at position Local texture features at is the image area The maximum value of all texture features in the mid-slice image; S2: Calculate the local contrast based on the slice image pixel value; Calculate the adaptive enhancement factor based on the image enhancement factor and the local contrast; Enhance the slice image based on the adaptive enhancement factor to obtain an enhanced image; S3: Input the enhanced image into the feature screening convolutional neural network for classification to obtain the final classification result.

2. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 1, characterized in that: The S1 specifically includes: The slice image is divided into regions, and the pixel value of the image is dynamically adjusted in an adaptive manner through an adaptive enhancement algorithm combined with the gradient amplitude, local texture characteristics and local contrast of the slice image.

3. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 2, characterized in that: The S2 specifically includes: The local contrast is calculated based on the difference between each pixel in the slice image and its adjacent pixels, combined with the pixel standard deviation of the image region.

4. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 3, characterized in that: The S2 specifically includes: The local contrast weighting method is introduced to adaptively adjust the image enhancement factor to obtain the adaptive enhancement factor.

5. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 4, characterized in that: The S2 specifically includes: The adaptive enhancement factor is applied to the slice image pixel value after normalization, and the slice image is enhanced in combination with the local contrast to obtain an enhanced image. The enhancement adjustment formula is as follows: ， in, Indicates at location The pixel value of the enhanced image; Indicates that the slice image is at position The pixel value at ; is the adaptive enhancement factor; It is the sum of the adaptive enhancement factors of all pixels; is the adjustment factor; is the slice image at position The local contrast at .

6. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 1, characterized in that: The S3 specifically includes: In the feature screening convolutional neural network, the enhanced image is convolved to obtain a preliminary feature matrix; the preliminary feature matrix is ​​divided to generate sub-matrices; a feature screening mechanism is introduced to calculate the score value of each sub-matrix based on the matrix determinant, condition number and Frobenius norm, and the score values ​​of all sub-matrices are sorted in descending order to screen out optimized features.

7. The method for automatic classification of lymphoma pathological sections based on convolutional neural network according to claim 6, characterized in that: The S3 specifically includes: The optimized features are used as new inputs, and features are extracted through convolution operations to obtain high-level features. The high-level features are processed through a fully connected layer, and the probability value of each category is output through a normalized exponential function. The category with the highest probability is selected as the final classification result.

Images