A deep learning-based MRI medical image segmentation method and system

Abstract

The present invention discloses a method and system for MRI medical image segmentation based on deep learning. The method comprises: acquiring a breast MRI image and performing normalization processing, removing non-interested regions to obtain a target MRI image; performing detail enhancement on the suspected lesion region to reconstruct it into an image to be preprocessed; performing noise reduction processing, contrast enhancement, and edge enhancement preprocessing on the image to be preprocessed, generating three versions of input images, and inputting them as three channels into a pre-trained deep learning model to complete the identification of breast tumors; performing image segmentation on the breast tumor region, and classifying the segmented breast tumors. The method and system for MRI medical image segmentation based on deep learning provided by the present invention achieve accurate and efficient processing of breast MRI images, efficient identification, accurate segmentation, and precise classification of breast tumors, can improve the accuracy and robustness of breast tumor detection, and provide efficient and reliable auxiliary diagnosis for clinical practice.

Description

Technical field

[0001] The present invention relates to the field of image recognition technology, and in particular to a deep learning-based MRI medical image segmentation method and system. [Background Technology]

[0002] With the advancement of medical imaging technology, early diagnosis and accurate identification of breast cancer have become critical components of clinical diagnosis. Breast cancer is one of the most common malignant tumors in women, and early detection and treatment significantly improve patient survival rates. The diagnosis of breast cancer typically relies on imaging studies, particularly breast MRI, which provides high-resolution images that help detect subtle lesions in breast tissue. MRI can provide more accurate information, especially in dense breast tissue, than traditional mammograms.

[0003] However, interpreting MRI images is not only time-consuming but also requires extensive professional experience on the part of radiologists. Furthermore, in image analysis for breast cancer diagnosis, the diversity and complexity of lesions, potentially related to interactions with dense tissue, can make it difficult for radiologists to accurately detect and analyze masses. Therefore, achieving breast MRI-assisted diagnosis and improving the accuracy and efficiency of breast tumor identification and classification have become pressing technical challenges in this field. [Summary of the invention]

[0004] In view of this, an embodiment of the present invention provides an MRI medical image segmentation method and system based on deep learning.

[0005] In a first aspect, an embodiment of the present invention provides a method for MRI medical image segmentation based on deep learning, the method comprising:

[0006] S1, obtain breast MRI images and perform normalization processing, remove non-interested regions to obtain target MRI images;

[0007] S2. Perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct it into an image to be preprocessed;

[0008] S3. Perform noise reduction, contrast enhancement, and edge enhancement on the image to be preprocessed, generating three versions of the input image. The three versions are input as three-channel inputs into a pre-trained deep learning model to complete breast tumor recognition.

[0009] S4. Perform image segmentation on the breast tumor region and classify the segmented breast tumor.

[0010] According to the above aspects and any possible implementation, an implementation is further provided, wherein S1 specifically includes:

[0011] Obtaining raw breast MRI image data from an MRI scanner, performing intensity normalization on the grayscale values ​​of the MRI image, mapping the grayscale values ​​to a preset range, and adjusting the spatial resolution to unify the pixel size and image size;

[0012] Calculate the global average threshold T based on the overall gray value distribution of the image g , preliminarily distinguish the breast tissue area from the background noise, the global average threshold T g The calculation formula is: Among them, I n (x, y) represents the grayscale value of the normalized image, and N represents the total number of pixels in the image;

[0013] For images with significant local grayscale changes, local contrast adjustment is further introduced to calculate the dynamic threshold T l , the calculation formula of the local contrast adjustment formula is: T l (x,y)=αT g +(1-α)·mean(N(x,y)), where α represents the weighting factor of the global and local thresholds, N(x,y) represents the local neighborhood of pixel (x,y), and mean(N(x,y)) represents the mean grayscale value of the local neighborhood around pixel (x,y);

[0014] According to the dynamic threshold T l Binarize the image and generate a segmentation mask M(x,y) to separate the breast tissue area from other non-interested areas, where M(x,y)=1 is the breast tissue area, and M(x,y)=0 is the non-breast tissue area;

[0015] Use Canny edge detection to perform contour detection on the segmentation mask M(x,y), detect edge points in the image, use contour tracking algorithm to generate the complete contour of the breast tissue area, calculate the minimum bounding rectangle containing the breast tissue area based on the complete contour, and generate the breast tissue area mask M based on the minimum bounding rectangle b (x,y), where B b It is represented by the minimum bounding rectangle of the breast tissue area;

[0016] Apply morphological opening operation to the segmentation mask M(x,y) to remove small area noise points and perform dilation operation on the mask boundary;

[0017] According to the breast tissue area mask M b (x,y) suppresses signals from subcutaneous fat and chest wall muscles, Among them, I f (x,y) represents the target MRI image containing only breast tissue;

[0018] Output target MRI image.

[0019] According to the above aspects and any possible implementation, a further implementation is provided, wherein the method for determining the significant local grayscale change is as follows:

[0020] Calculate the variance of the grayscale values ​​within the neighborhood N(x,y) of pixel (x,y):

[0021]

[0022] Calculate the ratio of the maximum to minimum grayscale values ​​in the neighborhood:

[0023]

[0024] Where μ(x,y) represents the neighborhood grayscale mean, |N(x,y)| represents the number of pixels in the neighborhood, I(i,j) represents the grayscale value of the pixel in the neighborhood, maxI(i,j) represents the maximum grayscale value in the neighborhood, minI(i,j) represents the minimum grayscale value in the neighborhood, and ε represents a positive number.

[0025] Compare the local variance with the preset variance threshold. If the local variance exceeds the preset variance threshold, there is a significant grayscale change in the area where the pixel (x, y) is located.

[0026] The local contrast is compared with the preset contrast threshold. If the local contrast exceeds the preset contrast threshold, there is a significant grayscale change in the area where the pixel (x, y) is located.

[0027] According to the above aspects and any possible implementation, an implementation is further provided, wherein S2 specifically includes:

[0028] From the target MRI image I f Extract the left and right unilateral areas from (x,y): I l (x,y)=I f (x,y), x∈[0,W / 2], I r (x,y)=I f (x,y), x∈[W / 2,W], W represents the image width;

[0029] For the left and right areas I l (x,y) and I r (x, y) generates a saliency map S(x, y), S(x, y) = |G x (x,y)|+|G y (x,y)|, where S(x,y) represents the significance value, G x (x,y) and G y(x, y) represent the gray-scale gradients of the pixel point (x, y) in the x and y directions respectively;

[0030] Calculate the average saliency value of the left and right saliency maps. S l (x, y) represents the saliency value in the left breast saliency map, |S l | and |S r | represent the number of non-zero pixels in the saliency map; Compare the differences between the left and right saliency maps. If |mean l -mean r | > T d , T d represents the saliency difference threshold, then select the side with higher saliency as the abnormal side. Otherwise, regard both the left and right unilateral regions as the abnormal side;

[0031] Perform dynamic threshold processing on the saliency map S(x, y) of the abnormal side: T S represents the dynamic threshold, T S =β·mean(s(x, y)), β represents the dynamic adjustment coefficient. Determine the minimum bounding rectangle of the suspected lesion area according to the mask M R (x, y), and mark the rectangle to generate the suspected lesion area image I R (x, y);

[0032] For the marked rectangle area of the suspected lesion area image I R (x, y), perform local enhancement. Decompose the marked rectangle area image I R (x, y) of the marked rectangle area to extract the base layer by Gaussian blur, I b (x, y)=Gaussian(I R (x, y), σ), σ represents the standard deviation of Gaussian blur, and Gaussian represents the Gaussian convolution function; Extract the detail layer, I d (x, y)=I R (x, y)-I b (x, y), I b (x, y) represents the low-frequency component extracted by Gaussian blur, and I d (x, y) represents the high-frequency component; Dynamically enhance the detail layer, I d,e (x, y)=η·I d (x, y), η represents the dynamic enhancement factor. k represents the enhancement ratio factor, std represents the standard deviation of the detail layer, and ε represents a positive number; Superimpose the enhanced detail layer and the base layer to generate the enhanced image, I e (x, y)=I b (x, y)+I d,e (x, y);

[0033] Image I of the suspected lesion area after local enhancement based on the marked rectangular frame area R (x,y) generates the image to be preprocessed I o (x,y).

[0034] According to the above aspects and any possible implementation, a further implementation is provided, wherein the mask M R (x, y) determines the minimum circumscribed rectangular frame of the suspected lesion area, marks the rectangular frame and generates the image I of the suspected lesion area R (x,y), specifically including:

[0035] According to the mask M R (x,y) Determine the minimum bounding rectangle B of the suspected lesion area min , B min =[x min ,x max ,y min ,y max ], x min and x max Indicates the minimum and maximum abscissas of the suspected lesion area, y min and y max Indicates the minimum and maximum ordinates of the suspected lesion area;

[0036] Calculate the extended rectangle B e , the minimum bounding rectangle B of the suspected lesion area min To expand, B e =[x min -d x ,x max +d x ,y min -d y ,y max +d y ], d x =φ x ·(x max -x min ), d y =φ y ·(y max -y min ),φ x Indicates the preset horizontal expansion ratio, φ y Indicates the preset vertical expansion ratio;

[0037] According to the expanded rectangular box B e Mark the original image to generate the extended suspected lesion area image I R (x,y).

[0038] According to the above aspects and any possible implementation, an implementation is further provided, wherein S3 specifically includes:

[0039] The image to be preprocessed is subjected to noise reduction, contrast enhancement and edge enhancement preprocessing respectively to generate three versions of input images;

[0040] Integrate the denoised image, contrast enhanced image and edge enhanced image into three-channel input;

[0041] Build a deep learning model for breast tumor recognition, use processed and annotated breast MRI images as an image dataset, and divide the dataset into training, validation, and test sets;

[0042] Use the cross entropy loss function to optimize model performance, train the deep learning model using the training set, and verify and evaluate it using the validation set and test set;

[0043] Use the trained deep learning model to identify the presence of breast tumors.

[0044] According to the above aspects and any possible implementation, an implementation is further provided, wherein the image to be preprocessed is subjected to noise reduction, contrast enhancement, and edge enhancement preprocessing, specifically comprising:

[0045] Define a pixel neighborhood N(x,y) for the preprocessed image, where N(x,y) is a local window around the center pixel (x,y);

[0046] Calculate the weighted value of each pixel in the neighborhood: w(x,p) represents the grayscale similarity weight between pixel p and the center pixel (x,y), I input (p) represents the grayscale value of pixel p, and h represents the smoothing parameter;

[0047] Perform weighted averaging of neighborhood pixels according to weights to generate a denoised image:

[0048]

[0049] Generate denoised image I d (x,y);

[0050] Divide the image to be preprocessed into 16×16 grid areas;

[0051] Calculate the histogram of grayscale values ​​in each grid area and clip the histogram: H c (i) = min(H(i), C), where H(i) represents the frequency of gray level i and C represents the clipping threshold;

[0052] Calculate the equalized grayscale value using the cropped histogram and reassemble the grid area to generate a contrast-enhanced image;

[0053] Generate contrast-enhanced image I c (x,y);

[0054] Calculate the Sobel gradient in the horizontal and vertical directions for the preprocessed image:

[0055] Calculate the gradient magnitude:

[0056] Superposition enhancement edge features: I e (x,y)=I o (x,y)+λ·G, λ represents the enhancement coefficient;

[0057] Generate edge-enhanced image I e (x,y).

[0058] According to the above aspects and any possible implementation, a further implementation is provided, wherein the building of a deep learning model for breast tumor recognition specifically includes:

[0059] Input layer: receives the preprocessed three-channel image I m =[I d ,I c ,I e ], each channel corresponds to the denoised image, contrast enhanced image and edge enhanced image, and the output size is 256×256×3;

[0060] Initial convolution and pooling, using a 7×7 convolution kernel to extract multi-scale features, followed by a 3×3 max pooling layer to reduce the feature size, including: Convolution layer 1: convolution kernel size 7×7, stride 2, number of channels 64, output size 128×128×64; Pooling layer: 3×3, stride 2, output size 64×64×64;

[0061] Construct three groups of residual modules, and add a channel attention mechanism after each group of residual modules, including: Residual block 1: two layers of 3×3 convolution, 64 channels, and an output size of 64×64×64, SE module 1: channel attention mechanism, output dimension 64, and an output size of 64×64×64, Residual block 2: two layers of 3×3 convolution, 128 channels, and an output size of 32×32×128, SE module 2: channel attention mechanism, output dimension 128, and an output size of 32×32×128, Residual block 3: two layers of 3×3 convolution, 256 channels, and an output size of 16×16×256, SE module 3: channel attention mechanism, output dimension 256, and an output size of 16×16×256;

[0062] Global pooling and classification layer: global average pooling is reduced to 1×1×256, and the output size is 1×1×256;

[0063] Fully connected layer: Sigmoid activation, outputting the probability value of whether the tumor exists;

[0064] The model parameters are optimized using the binary cross entropy loss function.

[0065] According to the above aspects and any possible implementation, an implementation is further provided, wherein S4 specifically includes:

[0066] Performing image segmentation on the breast tumor region to generate a tumor image, and extracting features of the tumor image, including but not limited to: morphological features, texture features, and color features. The morphological features include area, perimeter, and aspect ratio. The texture features are extracted using GLCM, including contrast, correlation, energy, and entropy. The color features are color histogram information in RGB and Lab color spaces.

[0067] Perform Z-score standardization on each eigenvector to generate a eigenvector;

[0068] A pre-trained first-level SVM classifier for each feature was used to perform first-level classification. Multiple pre-trained first-level SVM classifiers were used for first-level classification. Each classifier's decision boundary was based on an RBF kernel function. The input was a feature vector, and the output was a tumor classification result. Class 1 was benign tumors, and class 2 was malignant tumors. A voting mechanism was used to determine the final classification result.

[0069] If the classification result is category 2, then enter the secondary classification;

[0070] A secondary SVM classifier is pre-trained for each of the four malignant tumor subtypes. Multiple pre-trained secondary SVM classifiers are used for secondary classification. The output is the confidence score of the four categories: Luminal-A, Luminal-B, HER2+, and Basal-like. The confidence weighting strategy is used to determine the final classification result.

[0071] Generate the final classification results of breast tumors.

[0072] In a second aspect, an embodiment of the present invention provides an MRI medical image segmentation system based on deep learning, the system comprising:

[0073] a target image generation module, configured to acquire a breast MRI image and perform normalization processing to remove non-interested regions to acquire a target MRI image, wherein the non-interested regions include background noise, subcutaneous fat, and chest wall muscles;

[0074] The module for generating an image to be preprocessed is used to perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct the image to be preprocessed;

[0075] The recognition module performs noise reduction, contrast enhancement, and edge enhancement on the pre-processed image, generating three versions of the input image. These are then fed as three-channel inputs into a pre-trained deep learning model to identify breast tumors.

[0076] The classification module is used to perform image segmentation on the breast tumor area and classify the segmented breast tumor.

[0077] One of the above technical solutions has the following beneficial effects:

[0078] The method of the embodiment of the present invention proposes a deep learning-based MRI medical image segmentation method and system, which realizes accurate and efficient processing of breast MRI images, and realizes efficient identification, accurate segmentation and precise classification of breast tumors. It can improve the accuracy and robustness of breast tumor detection and provide an efficient and reliable auxiliary diagnostic tool for clinicians.

Brief Description of the Drawings

[0079] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0080] Figure 1 Schematic diagram of the process S1-S4 of the deep learning-based MRI medical image segmentation method provided in an embodiment of the present invention;

[0081] Figure 2 A schematic diagram of the structure of a deep learning model provided by an embodiment of the present invention;

[0082] Figure 3 This is a functional block diagram of the deep learning-based MRI medical image segmentation system provided in an embodiment of the present invention. [Specific implementation method]

[0083] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0084] Please refer to Figure 1 , which is a flow chart of a deep learning-based MRI medical image segmentation method S1-S4 provided by an embodiment of the present invention, such as Figure 1 As shown, the method includes the following steps:

[0085] S1, obtain breast MRI images and perform normalization processing, remove non-interested regions to obtain target MRI images;

[0086] S2. Perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct it into an image to be preprocessed;

[0087] S3. Perform noise reduction, contrast enhancement, and edge enhancement on the image to be preprocessed, generating three versions of the input image. The three versions are input as three-channel inputs into a pre-trained deep learning model to complete breast tumor recognition.

[0088] S4. Perform image segmentation on the breast tumor region and classify the segmented breast tumor.

[0089] The present invention effectively reduces background and irrelevant area interference by removing non-interest areas including non-breast tissue areas, focusing on the characteristics of the breast tissue area; the preliminary target determination step helps identify the abnormal side, then confirms the suspected lesion area in the breast tissue area, and performs detail enhancement processing on the initially identified suspected lesion area, reducing the amount of tasks and processing, further improving the visibility of the lesion area, and improving the accuracy of breast tumor recognition; the deep learning model is used for tumor recognition accurately, quickly and efficiently, noise reduction processing removes noise in the image, contrast enhancement makes the tumor area more prominent, and edge enhancement strengthens the edge information of the tumor area. The noise reduction image, contrast enhancement image and edge enhancement image are integrated into three-channel input. This multi-channel input method can provide richer image features for deep learning model analysis, helping the model to understand the characteristics of the tumor area from different angles and improving the accuracy and robustness of the model; segmenting the tumor area can provide a clear tumor outline, which is convenient for subsequent tumor classification and further clinical analysis. The model can perform high-precision classification of tumors, combined with the malignancy and subtype classification of the tumor, to provide more comprehensive diagnostic information. The deep learning-based MRI medical image segmentation method and system provided by the present invention realize the precise and efficient processing of breast MRI images, and achieve the efficient identification, accurate segmentation and precise classification of breast tumors. It can improve the accuracy and robustness of breast tumor detection and provide an efficient and reliable auxiliary diagnostic tool for clinicians.

[0090] In a preferred embodiment of the present invention, the S1 specifically includes:

[0091] Obtaining raw breast MRI image data from an MRI scanner, performing intensity normalization on the grayscale values ​​of the MRI image, mapping the grayscale values ​​to a preset range, and adjusting the spatial resolution to unify the pixel size and image size;

[0092] Calculate the global average threshold T based on the overall gray value distribution of the image g , preliminarily distinguish the breast tissue area from the background noise, the global average threshold T g The calculation formula is: Among them, I n (x, y) represents the grayscale value of the normalized image, and N represents the total number of pixels in the image;

[0093] For images with significant local grayscale changes, local contrast adjustment is further introduced to calculate the dynamic threshold T l , the calculation formula of the local contrast adjustment formula is: T l (x,y)=αT g +(1-α)·mean(N(x,y)), where α represents the weighting factor of the global and local thresholds, N(x,y) represents the local neighborhood of pixel (x,y), and mean(N(x,y)) represents the mean grayscale value of the local neighborhood around pixel (x,y);

[0094] According to the dynamic threshold T l Binarize the image and generate a segmentation mask M(x,y) to separate the breast tissue area from other non-interested areas, where M(x,y)=1 is the breast tissue area, and M(x,y)=0 is the non-breast tissue area;

[0095] Use Canny edge detection to perform contour detection on the segmentation mask M(x,y), detect edge points in the image, use contour tracking algorithm to generate the complete contour of the breast tissue area, calculate the minimum bounding rectangle containing the breast tissue area based on the complete contour, and generate the breast tissue area mask M based on the minimum bounding rectangle b (x,y), where B b It is represented by the minimum bounding rectangle of the breast tissue area;

[0096] Apply morphological opening operation to the segmentation mask M(x,y) to remove small area noise points and perform dilation operation on the mask boundary;

[0097] According to the breast tissue area mask M b (x,y) suppresses signals from subcutaneous fat and chest wall muscles, Among them, I f (x,y) represents the target MRI image containing only breast tissue;

[0098] Output target MRI image.

[0099] The present invention eliminates the deviation between different images caused by differences in scanning conditions, equipment or image quality by standardizing breast MRI images, thereby ensuring the consistency of subsequent processing steps; by removing background noise in the image, the interference of the model on irrelevant information is significantly reduced; subcutaneous fat and chest wall muscles and the anatomical structure of the tumor easily produce similar grayscale values, which will interfere with the accurate identification of the tumor. The present invention enhances the accuracy and robustness of removing non-interest regions through local contrast adjustment, dynamic threshold segmentation and morphological operations, and by separating the breast tissue area from other non-interest regions, it can effectively reduce interference and focus on the characteristics of the tumor area.

[0100] It should be noted that the method for determining the significant local grayscale change is as follows:

[0101] Calculate the variance of the grayscale values ​​within the neighborhood N(x,y) of pixel (x,y):

[0102]

[0103] Calculate the ratio of the maximum to minimum grayscale values ​​in the neighborhood:

[0104]

[0105] Where μ(x,y) represents the neighborhood grayscale mean, |N(x,y)| represents the number of pixels in the neighborhood, I(i,j) represents the grayscale value of the pixel in the neighborhood, maxI(i,j) represents the maximum grayscale value in the neighborhood, minI(i,j) represents the minimum grayscale value in the neighborhood, and ε represents a positive number.

[0106] Compare the local variance with the preset variance threshold. If the local variance exceeds the preset variance threshold, there is a significant grayscale change in the area where the pixel (x, y) is located.

[0107] The local contrast is compared with the preset contrast threshold. If the local contrast exceeds the preset contrast threshold, there is a significant grayscale change in the area where the pixel (x, y) is located.

[0108] The present invention measures the discreteness of the neighborhood grayscale values ​​based on the results of variance calculation, can determine whether there is a significant grayscale change in the region, and adjust the dynamic threshold accordingly. By calculating the contrast to measure the difference in brightness and darkness in the neighborhood, the edges and feature areas can be highlighted for dynamic threshold adjustment. The combination of the two provides a reliable grayscale characteristic measurement basis for dynamic threshold segmentation, which can effectively distinguish between breast areas and non-areas of interest.

[0109] In a preferred embodiment of the present invention, S2 specifically includes:

[0110] Extract the left and right unilateral regions from the target MRI image I f (x,y): I l (x,y) = I f (x,y), x ∈ [0, W / 2], I r (x,y) = I f (x,y), x ∈ [W / 2, W], where W represents the image width;

[0111] Generate saliency maps S(x,y) for the left and right regions I l (x,y) and I r (x,y) respectively, S(x,y) = |G x (x,y)| + |G y (x,y)|, where S(x,y) represents the saliency value, and G x (x,y) and G y (x,y) respectively represent the gray-scale gradients of the pixel point (x,y) in the x and y directions;

[0112] Calculate the average saliency values of the left and right saliency maps, S l (x,y) represents the saliency value in the left breast saliency map, |S l | and |S r | represent the number of non-zero pixels in the saliency map; compare the differences between the left and right saliency maps. If |mean l - mean r | > T d , where T d represents the saliency difference threshold, then select the side with higher saliency as the abnormal side; otherwise, consider both the left and right unilateral regions as abnormal sides;

[0113] Perform dynamic threshold processing on the saliency map S(x,y) of the abnormal side: T S represents the dynamic threshold, T S = β · mean(s(x,y)), where β represents the dynamic adjustment coefficient. Determine the minimum bounding rectangle of the suspected lesion area according to the mask M[[ID= 57]] R (x,y), and mark the rectangle to generate the suspected lesion area image I R (x,y);

[0114] Perform local enhancement on the marked rectangle area of the suspected lesion area image I R (x,y). Decompose the marked rectangle area image I R (x,y) of the marked rectangle area, perform Gaussian blur to extract the base layer, I b(x,y)=Gaussian(I R (x,y),σ), σ represents the standard deviation of Gaussian blur, Gaussian represents Gaussian convolution function; extract detail layer, I d (x,y)=I R (x,y)-I b (x,y),I b (x,y) represents the extraction of low-frequency components by Gaussian blur, I d (x,y) represents the high-frequency component; dynamic enhancement detail layer, I d,e (x,y)=η·I d (x,y), η represents the dynamic enhancement factor, k represents the enhancement scale factor, std represents the standard deviation of the detail layer, and ε represents a positive number; the enhanced detail layer is superimposed on the base layer to generate an enhanced image, I e (x,y)=I b (x,y)+I d,e (x,y);

[0115] Image I of the suspected lesion area after local enhancement based on the marked rectangular frame area R (x,y) generates the image to be preprocessed I o (x,y).

[0116] It should be noted that breast lesions are usually concentrated in unilateral areas. Gradient change detection and difference analysis of bilateral breasts can quickly locate the abnormal side. By quickly quantifying significant differences, only the abnormal side is processed to avoid repeated analysis of the normal side area, reduce the workload of bilateral breast analysis, and improve efficiency. If there is no significant difference, both the left and right unilateral areas are regarded as abnormal sides, avoiding the omission of tumor targets; at the same time, the suspected lesion area is accurately identified to provide a basis for subsequent detail enhancement. Local enhancement improves the contrast and resolution of the suspected lesion area, and the enhanced detail layer is superimposed on the base layer to ensure that the enhanced image is clearer without excessively amplifying noise, providing high-quality input for subsequent deep learning models and improving the accuracy of breast tumor recognition.

[0117] In a preferred embodiment of the present invention, the mask M R (x, y) determines the minimum circumscribed rectangular frame of the suspected lesion area, marks the rectangular frame and generates the image I of the suspected lesion area R (x,y), specifically including:

[0118] According to the mask M R (x,y) Determine the minimum bounding rectangle B of the suspected lesion area min , B min =[x min ,x max ,ymin ,y max ], x min and x max Indicates the minimum and maximum abscissas of the suspected lesion area, y min and y max Indicates the minimum and maximum ordinates of the suspected lesion area;

[0119] Calculate the extended rectangle B e , the minimum bounding rectangle B of the suspected lesion area min To expand, B e =[x min -d x ,x max +d x ,y min -d y ,y max +d y ], d x =φ x ·(x max -x min ), d y =φ y ·(y max -y min ),φ x Indicates the preset horizontal expansion ratio, φ y Indicates the preset vertical expansion ratio;

[0120] According to the expanded rectangular box B e Mark the original image to generate the extended suspected lesion area image I R (x,y).

[0121] The present invention expands the local enhancement range by dynamically expanding the rectangular frame to avoid missing tumors. In addition, it adds breast tissue information while retaining the tumor area, reasonably expands the recognition range, and can capture the boundary features of the tumor and surrounding normal tissues. This helps the model improve the accuracy of tumor recognition and learn the characteristic differences between lesions and normal tissues.

[0122] In a preferred embodiment of the present invention, S3 specifically includes:

[0123] The image to be preprocessed is subjected to noise reduction, contrast enhancement and edge enhancement preprocessing respectively to generate three versions of input images;

[0124] Integrate the denoised image, contrast enhanced image and edge enhanced image into three-channel input;

[0125] Build a deep learning model for breast tumor recognition, use processed and annotated breast MRI images as an image dataset, and divide the dataset into training, validation, and test sets;

[0126] Use the cross entropy loss function to optimize model performance, train the deep learning model using the training set, and verify and evaluate it using the validation set and test set;

[0127] Use the trained deep learning model to identify the presence of breast tumors.

[0128] The present invention removes noise from MRI images through denoising to highlight the main structures and features; contrast enhancement improves the difference between light and dark in the tumor area, making it easier for the model to extract significant features; edge enhancement highlights the tumor boundary and provides clear structured information; and multi-channel input features are learned through residual networks and attention mechanisms to optimize the recognition ability of breast tumors.

[0129] Specifically, the image to be preprocessed is subjected to noise reduction, contrast enhancement, and edge enhancement preprocessing, which specifically includes:

[0130] Define a pixel neighborhood N(x,y) for the preprocessed image, where N(x,y) is a local window around the center pixel (x,y);

[0131] Calculate the weighted value of each pixel in the neighborhood: w(x,p) represents the grayscale similarity weight between pixel p and the center pixel (x,y), I input (p) represents the grayscale value of pixel p, and h represents the smoothing parameter;

[0132] Perform weighted averaging of neighborhood pixels according to weights to generate a denoised image:

[0133]

[0134] Generate denoised image I d (x,y);

[0135] Divide the image to be preprocessed into 16×16 grid areas;

[0136] Calculate the histogram of grayscale values ​​in each grid area and clip the histogram: H c (i) = min(H(i), C), where H(i) represents the frequency of gray level i and C represents the clipping threshold;

[0137] Calculate the equalized grayscale value using the cropped histogram and reassemble the grid area to generate a contrast-enhanced image;

[0138] Generate contrast-enhanced image I c (x,y);

[0139] Calculate the Sobel gradient in the horizontal and vertical directions for the preprocessed image:

[0140] Calculate the gradient magnitude:

[0141] Superposition enhancement edge features: I e (x,y)=I o (x,y)+λ·G, λ represents the enhancement coefficient;

[0142] Generate edge-enhanced image I e (x,y).

[0143] The present invention adopts non-local mean filtering and calculates weights by neighborhood grayscale similarity to enhance the protection of tumor boundaries and textures, improve the ability to retain breast tumor boundary details, and simultaneously remove random noise from MRI images; dynamic adjustment of the cropping threshold, grid division and histogram equalization make contrast enhancement more adaptable, improve the light-dark contrast between the lesion area and normal tissue, and make tumor features more obvious; the present invention superimposes edge features with the original image to enhance the significance of the lesion area boundary, highlighting the tumor boundary while avoiding noise amplification caused by excessive enhancement.

[0144] In a preferred embodiment of the present invention, Figure 2 As shown, the construction of a deep learning model for breast tumor recognition specifically includes:

[0145] Input layer: receives the preprocessed three-channel image I m =[I d ,I c ,I e ], each channel corresponds to the denoised image, contrast enhanced image and edge enhanced image, and the output size is 256×256×3;

[0146] Initial convolution and pooling, using a 7×7 convolution kernel to extract multi-scale features, followed by a 3×3 max pooling layer to reduce the feature size, including: Convolution layer 1: convolution kernel size 7×7, stride 2, number of channels 64, output size 128×128×64; Pooling layer: 3×3, stride 2, output size 64×64×64;

[0147] Construct three groups of residual modules, and add a channel attention mechanism after each group of residual modules, including: Residual block 1: two layers of 3×3 convolution, 64 channels, and an output size of 64×64×64, SE module 1: channel attention mechanism, output dimension 64, and an output size of 64×64×64, Residual block 2: two layers of 3×3 convolution, 128 channels, and an output size of 32×32×128, SE module 2: channel attention mechanism, output dimension 128, and an output size of 32×32×128, Residual block 3: two layers of 3×3 convolution, 256 channels, and an output size of 16×16×256, SE module 3: channel attention mechanism, output dimension 256, and an output size of 16×16×256;

[0148] Global pooling and classification layer: global average pooling is reduced to 1×1×256, and the output size is 1×1×256;

[0149] Fully connected layer: Sigmoid activation, outputting the probability value of whether the tumor exists;

[0150] The model parameters are optimized using the binary cross entropy loss function.

[0151] The three preprocessing methods of the deep learning model of the present invention respectively extract noise reduction features, contrast features and edge features. Through the three-channel input network, the residual module extracts features of different scales layer by layer. Each group of residual modules contains two layers of convolution. The residual connection alleviates the gradient vanishing problem and realizes efficient feature extraction. Multiple groups of modules gradually extract features of different scales, capture the global and local information of the tumor area, and the attention mechanism dynamically adjusts the feature weights of each channel to strengthen the key features. Through the optimization of the binary cross entropy loss function, the prediction result of whether the tumor exists is output, which improves the detection accuracy of breast tumors, strengthens the fusion ability of noise reduction, contrast and edge features, enhances the comprehensive utilization ability of multi-channel features, adapts to the characteristics of breast MRI, and realizes efficient and accurate recognition.

[0152] In a preferred embodiment of the present invention, the S4 specifically includes:

[0153] Performing image segmentation on the breast tumor region to generate a tumor image, and extracting features of the tumor image, including but not limited to: morphological features, texture features, and color features. The morphological features include area, perimeter, and aspect ratio. The texture features are extracted using GLCM, including contrast, correlation, energy, and entropy. The color features are color histogram information in RGB and Lab color spaces.

[0154] Perform Z-score standardization on each eigenvector to generate a eigenvector;

[0155] A pre-trained first-level SVM classifier for each feature was used to perform first-level classification. Multiple pre-trained first-level SVM classifiers were used for first-level classification. Each classifier's decision boundary was based on an RBF kernel function. The input was a feature vector, and the output was a tumor classification result. Class 1 was benign tumors, and class 2 was malignant tumors. A voting mechanism was used to determine the final classification result.

[0156] If the classification result is category 2, then enter the secondary classification;

[0157] A secondary SVM classifier is pre-trained for each of the four malignant tumor subtypes. Multiple pre-trained secondary SVM classifiers are used for secondary classification. The output is the confidence score of the four categories: Luminal-A, Luminal-B, HER2+, and Basal-like. The confidence weighting strategy is used to determine the final classification result.

[0158] Generate the final classification results of breast tumors.

[0159] The features extracted from tumor images by this method include morphology, texture, and color, which cover the key information dimensions of tumors. A voting mechanism enhances the stability of the primary classification, preventing misclassifications caused by the failure of a single classifier. The confidence-weighted strategy for the secondary classification integrates the results of multiple classifiers, improving the accuracy of malignant tumor subtype classification. By combining the first- and second-stage SVM classifiers, the recognition rate of malignant tumors can be improved, and the classification of malignant tumor subtypes can be refined, providing diagnostic value to doctors and facilitating more accurate diagnosis and treatment decisions.

[0160] The embodiments of the present invention further provide device embodiments for implementing the steps and methods in the above method embodiments.

[0161] Please refer to Figure 3 , which is a functional block diagram of an MRI medical image segmentation system based on deep learning provided by an embodiment of the present invention, such as Figure 3 As shown, the deep learning-based MRI medical image segmentation system includes:

[0162] a target image generation module, configured to acquire a breast MRI image and perform normalization processing to remove non-interested regions to acquire a target MRI image, wherein the non-interested regions include background noise, subcutaneous fat, and chest wall muscles;

[0163] The module for generating an image to be preprocessed is used to perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct the image to be preprocessed;

[0164] The recognition module performs noise reduction, contrast enhancement, and edge enhancement on the pre-processed image, generating three versions of the input image. These are then fed as three-channel inputs into a pre-trained deep learning model to identify breast tumors.

[0165] The classification module is used to perform image segmentation on the breast tumor area and classify the segmented breast tumor.

[0166] Since each unit module in this embodiment can execute Figure 1 For the method shown in the embodiment, the part not described in detail in this embodiment can be referred to Figure 1 Related instructions.

[0167] At the hardware level, the device may include a processor and, optionally, an internal bus, a network interface, and memory. The memory may include internal memory, such as high-speed random-access memory (RAM), and may also include non-volatile memory, such as at least one disk drive. Of course, the device may also include other hardware required for the service.

[0168] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus. The bus can be divided into an address bus, a data bus, a control bus, and the like.

[0169] The memory is used to store programs. Specifically, the program may include program code, which includes computer operating instructions. The memory may include internal memory and non-volatile memory, and provides instructions and data to the processor.

[0170] The steps of the method disclosed in conjunction with the embodiments of the present invention can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules within the decoding processor. The software modules can be located in a storage medium well-known in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the method described above.

[0171] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0172] For the convenience of description, the above device is described as being divided into various units or modules according to their functions. Of course, when implementing the present invention, the functions of each unit or module can be implemented in the same or multiple software and / or hardware.

[0173] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0175] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0177] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0178] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0179] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0180] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0181] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0182] The present invention may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0183] The various embodiments of the present invention are described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so its description is relatively simple. For relevant portions, refer to the description of the method embodiment.

[0184] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A deep learning-based MRI medical image segmentation method, characterized in that: The method comprises: S1, obtain breast MRI images and perform normalization processing, remove non-interested regions to obtain target MRI images; S2. Perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct it into an image to be preprocessed; S3. Perform noise reduction, contrast enhancement, and edge enhancement on the image to be preprocessed, generating three versions of the input image. The three versions are input as three-channel inputs into a pre-trained deep learning model to complete breast tumor recognition. S4, performing image segmentation on the breast tumor region and classifying the segmented breast tumor; The S2 specifically includes: From the target MRI image I f Extract the left and right unilateral areas from (x,y): I l (x,y)=I f (x,y), x∈[0,W / 2], I r (x,y)=I f (x,y), x∈[W / 2,W], W represents the image width; For the left and right areas I l (x,y) and I r (x, y) generates a saliency map S(x, y), S(x, y) = |G x (x,y)|+|G y (x,y)|, where S(x,y) represents the significance value, G x (x,y) and G y (x,y) represents the grayscale gradient of the pixel point (x,y) in the x and y directions respectively; Calculate the average saliency value of the left and right saliency maps. S l (x, y) represents the saliency value in the left breast saliency map, |S l | and |S r | represent the number of non-zero pixels in the saliency map; compare the differences between the left and right saliency maps. If |mean l -mean r | > T d , T d represents the saliency difference threshold, then select the side with higher saliency as the abnormal side. Otherwise, consider both the left and right unilateral regions as abnormal sides. Perform dynamic threshold processing on the saliency map S(x,y) on the abnormal side: T S represents the dynamic threshold, T S =β·mean(s(x,y)), β represents the dynamic adjustment coefficient, according to the mask M R (x, y) determines the minimum circumscribed rectangular frame of the suspected lesion area, marks the rectangular frame and generates the image I of the suspected lesion area R (x,y); For the suspected lesion area image I R The marked rectangular area of ​​(x, y) is locally enhanced, and the marked rectangular area image I is decomposed R The (x, y) marked rectangular box area is Gaussian blurred to extract the base layer, I b (x,y)=Gaussian(I R (x,y),σ), σ represents the standard deviation of Gaussian blur, Gaussian represents Gaussian convolution function; extract detail layer, I d (x,y)=I R (x,y)-I b (x,y),I b (x,y) represents the extraction of low-frequency components by Gaussian blur, I d (x,y) represents the high-frequency component; dynamic enhancement detail layer, I d,e (x,y)=η·I d (x,y), η represents the dynamic enhancement factor, k represents the enhancement scale factor, std represents the standard deviation of the detail layer, and ε represents a positive number; the enhanced detail layer is superimposed on the base layer to generate an enhanced image, I e (x,y)=I b (x,y)+I d,e (x,y); Image I of the suspected lesion area after local enhancement based on the marked rectangular frame area R (x,y) generates the image to be preprocessed I o (x,y).

2. The method according to claim 1, characterized in that Said S1 specifically includes: Obtaining raw breast MRI image data from an MRI scanner, performing intensity normalization on the grayscale values ​​of the MRI image, mapping the grayscale values ​​to a preset range, and adjusting the spatial resolution to unify the pixel size and image size; Calculate the global average threshold T based on the overall gray value distribution of the image g , preliminarily distinguish the breast tissue area from the background noise, the global average threshold T g The calculation formula is: Among them, I n (x, y) represents the grayscale value of the normalized image, and N represents the total number of pixels in the image; For images with significant local grayscale changes, local contrast adjustment is further introduced to calculate the dynamic threshold T l , the calculation formula of the local contrast adjustment formula is: T l (x,y)=αT g +(1-α)·mean(N(x,y)), where α represents the weighting factor of the global and local thresholds, N(x,y) represents the local neighborhood of pixel (x,y), and mean(N(x,y)) represents the mean grayscale value of the local neighborhood around pixel (x,y); According to the dynamic threshold T l Binarize the image and generate a segmentation mask M(x,y) to separate the breast tissue area from other non-interested areas, where M(x,y)=1 is the breast tissue area, and M(x,y)=0 is the non-breast tissue area; Use Canny edge detection to perform contour detection on the segmentation mask M(x,y), detect edge points in the image, use contour tracking algorithm to generate the complete contour of the breast tissue area, calculate the minimum bounding rectangle containing the breast tissue area based on the complete contour, and generate the breast tissue area mask M based on the minimum bounding rectangle b (x,y), where B b It is represented by the minimum bounding rectangle of the breast tissue area; Apply morphological opening operation to the segmentation mask M(x,y) to remove small area noise points and perform dilation operation on the mask boundary; According to the breast tissue area mask M b (x,y) suppresses signals from subcutaneous fat and chest wall muscles, Among them, I f (x,y) represents the target MRI image containing only breast tissue; Output target MRI image.

3. The method according to claim 2, characterized in that The method for determining the significant local grayscale change is as follows: Calculate the variance of the grayscale values ​​within the neighborhood N(x,y) of pixel (x,y): Calculate the ratio of the maximum to minimum grayscale values ​​in the neighborhood: Where μ(x,y) represents the neighborhood grayscale mean, |N(x,y)| represents the number of pixels in the neighborhood, I(i,j) represents the grayscale value of the pixel in the neighborhood, maxI(i,j) represents the maximum grayscale value in the neighborhood, minI(i,j) represents the minimum grayscale value in the neighborhood, and ε represents a positive number. Compare the local variance with the preset variance threshold. If the local variance exceeds the preset variance threshold, there is a significant grayscale change in the area where the pixel (x, y) is located. The local contrast is compared with the preset contrast threshold. If the local contrast exceeds the preset contrast threshold, there is a significant grayscale change in the area where the pixel (x, y) is located.

4. The method according to claim 1, wherein The mask M R (x, y) determines the minimum circumscribed rectangular frame of the suspected lesion area, marks the rectangular frame and generates the image I of the suspected lesion area R (x,y), specifically including: According to the mask M R (x,y) Determine the minimum bounding rectangle B of the suspected lesion area min , B min =[x min ,x max ,y min ,y max ], x min and x max Indicates the minimum and maximum abscissas of the suspected lesion area, y min and y max Indicates the minimum and maximum ordinates of the suspected lesion area; Calculate the extended rectangle B e , the minimum bounding rectangle B of the suspected lesion area min To expand, B e =[x min -d x ,x max +d x ,y min -d y ,y max +d y ], d x =φ x ·(x max -x min ), d y =φ y ·(y max -y min ),φ x Indicates the preset horizontal expansion ratio, φ y Indicates the preset vertical expansion ratio; According to the expanded rectangular box B e Mark the original image to generate the extended suspected lesion area image I R (x,y).

5. The method according to claim 1, characterized in that The S3 specifically includes: The image to be preprocessed is subjected to noise reduction, contrast enhancement and edge enhancement preprocessing respectively to generate three versions of input images; Integrate the denoised image, contrast enhanced image and edge enhanced image into three-channel input; Build a deep learning model for breast tumor recognition, use processed and annotated breast MRI images as an image dataset, and divide the dataset into training, validation, and test sets; Use the cross entropy loss function to optimize model performance, train the deep learning model using the training set, and verify and evaluate it using the validation set and test set; Use the trained deep learning model to identify the presence of breast tumors.

6. The method according to claim 5, characterized in that The image to be preprocessed is subjected to noise reduction, contrast enhancement and edge enhancement preprocessing, specifically comprising: Define a pixel neighborhood N(x,y) for the preprocessed image, where N(x,y) is a local window around the center pixel (x,y); Calculate the weighted value of each pixel in the neighborhood: w(x,p) represents the grayscale similarity weight between pixel p and the center pixel (x,y), I input (p) represents the grayscale value of pixel p, and h represents the smoothing parameter; Perform weighted averaging of neighborhood pixels according to weights to generate a denoised image: Generate denoised image I dn (x,y); Divide the image to be preprocessed into 16×16 grid areas; Calculate the histogram of grayscale values ​​in each grid area and clip the histogram: H c (i) = min(H(i), C), where H(i) represents the frequency of gray level i and C represents the clipping threshold; Calculate the equalized grayscale value using the cropped histogram and reassemble the grid area to generate a contrast-enhanced image; Generate contrast-enhanced image I c (x,y); Calculate the Sobel gradient in the horizontal and vertical directions for the preprocessed image: Calculate the gradient magnitude: Superposition enhancement edge features: I e (x,y)=I o (x,y)+λ·G, λ represents the enhancement coefficient; Generate edge-enhanced image I e (x,y).

7. The method according to claim 5, characterized in that The construction of a deep learning model for breast tumor identification specifically includes: Input layer: receives the preprocessed three-channel image I m =[I dn ,I c ,I e ], each channel corresponds to the denoised image, contrast enhanced image and edge enhanced image, and the output size is 256×256×3; Initial convolution and pooling, using a 7×7 convolution kernel to extract multi-scale features, followed by a 3×3 max pooling layer to reduce the feature size, including: Convolution layer 1: convolution kernel size 7×7, stride 2, number of channels 64, output size 128×128×64; Pooling layer: 3×3, stride 2, output size 64×64×64; Construct three groups of residual modules, and add a channel attention mechanism after each group of residual modules, including: Residual block 1: two layers of 3×3 convolution, 64 channels, and an output size of 64×64×64, SE module 1: channel attention mechanism, output dimension 64, and an output size of 64×64×64, Residual block 2: two layers of 3×3 convolution, 128 channels, and an output size of 32×32×128, SE module 2: channel attention mechanism, output dimension 128, and an output size of 32×32×128, Residual block 3: two layers of 3×3 convolution, 256 channels, and an output size of 16×16×256, SE module 3: channel attention mechanism, output dimension 256, and an output size of 16×16×256; Global pooling and classification layer: global average pooling is reduced to 1×1×256, and the output size is 1×1×256; Fully connected layer: Sigmoid activation, outputting the probability value of whether the tumor exists; The model parameters are optimized using the binary cross entropy loss function.

8. The method according to claim 1, characterized in that The S4 specifically includes: Performing image segmentation on the breast tumor region to generate a tumor image, and extracting features of the tumor image, including but not limited to: morphological features, texture features, and color features. The morphological features include area, perimeter, and aspect ratio. The texture features are extracted using GLCM, including contrast, correlation, energy, and entropy. The color features are color histogram information in RGB and Lab color spaces. Perform Z-score standardization on each eigenvector to generate a eigenvector; A pre-trained first-level SVM classifier for each feature was used to perform first-level classification. Multiple pre-trained first-level SVM classifiers were used for first-level classification. Each classifier's decision boundary was based on an RBF kernel function. The input was a feature vector, and the output was a tumor classification result. Class 1 was benign tumors, and class 2 was malignant tumors. A voting mechanism was used to determine the final classification result. If the classification result is category 2, then enter the secondary classification; A secondary SVM classifier is pre-trained for each of the four malignant tumor subtypes. Multiple pre-trained secondary SVM classifiers are used for secondary classification. The output is the confidence score of the four categories: Luminal-A, Luminal-B, HER2+, and Basal-like. The confidence weighting strategy is used to determine the final classification result. Generate the final classification results of breast tumors.

9. A deep learning-based MRI medical image segmentation system using the method of claim 1, characterized in that: include: a target image generation module, configured to acquire a breast MRI image and perform normalization processing to remove non-interested regions to acquire a target MRI image, wherein the non-interested regions include background noise, subcutaneous fat, and chest wall muscles; The module for generating an image to be preprocessed is used to perform preliminary target determination on the target MRI image, determine the suspected lesion area based on the preliminary target, and perform detail enhancement on the suspected lesion area to reconstruct the image to be preprocessed; The recognition module performs noise reduction, contrast enhancement, and edge enhancement on the pre-processed image, generating three versions of the input image. These are then fed as three-channel inputs into a pre-trained deep learning model to identify breast tumors. The classification module is used to perform image segmentation on the breast tumor area and classify the segmented breast tumor.

Images