Whole breast information analysis for prediction of neoadjuvant therapy response in breast cancer

Abstract

The present application belongs to the technical field of medical image processing, and particularly relates to a whole breast information analysis method for predicting the curative effect of neoadjuvant therapy of breast cancer. The steps of the method are as follows: S1, collecting and arranging DCE-MRI breast imaging of breast cancer patients before and after receiving neoadjuvant therapy; S2, performing pixel matrix operation on the images collected and arranged in S1 and retaining breast pathological area images and axillary images by using a region growing algorithm; S3, arranging the images obtained in S2, inputting the arranged images into a neural network, performing model training, and obtaining a prediction model; S4, performing feature fusion on the model data of the breast cancer patients before and after receiving neoadjuvant therapy obtained in S3 and performing classification prediction to obtain effective prediction of pathological complete remission and prediction of whether lymph node metastasis occurs. Thus, the analysis and prediction efficiency is improved, the information extraction amount is improved, the accuracy of model prediction is improved, and better treatment reference opinions can be provided for patients.

Description

[Technical Field]

[0001] This invention belongs to the technical field of medical image processing, specifically relating to a whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer. [Background Technology]

[0002] Currently, neoadjuvant therapy has become an important part of clinical treatment for breast cancer, with neoadjuvant chemotherapy (NAC) being the primary treatment approach. Neoadjuvant therapy offers advantages such as downstaging surgery, breast-conserving surgery, understanding tumor sensitivity to treatment regimens, and determining the next treatment option. However, clinical trials have revealed that some patients are not sensitive to neoadjuvant therapy, and the efficacy of treatment varies from person to person. Therefore, it is necessary to predict the efficacy of neoadjuvant therapy to better identify suitable treatment methods for patients and ensure that they receive effective treatment as early as possible.

[0003] In the current technology for predicting the efficacy of neoadjuvant therapy for breast cancer, imaging examinations are one of the most commonly used methods, mainly including mammography, ultrasound, and magnetic resonance imaging (MRI). Information such as tumor size, margin, and density can be extracted from these imaging examinations. However, these intuitive information alone cannot be used to make a good assessment of the efficacy of neoadjuvant therapy. Therefore, image processing algorithms are needed to delve into the internal features of the lesion.

[0004] Currently, most studies evaluating the efficacy of neoadjuvant therapy (NAC) in breast cancer using MRI images only extract some features from breast MRI images, without extracting information from the entire breast for research. Furthermore, it is necessary to label the corresponding feature regions, which consumes a lot of doctors' time and has poor predictive results. Therefore, there is a need for a neoadjuvant therapy efficacy prediction method that extracts information from the entire breast to help patients obtain more suitable treatment plans.

[0005] The main routes of breast cancer metastasis include lymph node metastasis, hematogenous metastasis, and local spread, with lymph node metastasis being the most common. Whether or not lymph node metastasis is involved is extremely important for tumor staging, treatment planning, and prognostic assessment. Axillary lymph node metastasis is the most common type of lymph node metastasis. Therefore, surgical treatment for breast tumors mainly includes the management of the primary breast tumor and the axilla. Previous studies predicting neoadjuvant therapy did not include predictions of lymph node metastasis; therefore, treatment recommendations based on this lack of prediction have relatively limited significance.

[0006] Chinese invention patent application CN202110712181.0 discloses a method for constructing a lymph node metastasis prediction model for breast cancer patients based on radiomics. The steps include: collecting magnetic resonance imaging (MRI) image data and clinical feature data of patients from multiple centers, and dividing the patients into a training cohort, a neoadjuvant therapy validation cohort, and an external validation cohort according to the center and treatment plan; extracting MR image features of the primary breast tumor and axillary lymph nodes from multiple MR sequences based on the MRI image data, using a random forest algorithm to screen the MR image features to obtain several key image features, and establishing an image feature prediction model based on the key image features using a support vector machine (SVM) algorithm; performing univariate analysis on the clinical feature data based on the clinical feature data, selecting statistically significant clinical features as key clinical features, and establishing a clinical feature prediction model based on the key clinical features using an SVM algorithm; and establishing a comprehensive lymph node metastasis prediction model based on the key image features and key clinical features using an SVM algorithm, wherein the model category of the SVM algorithm is c-classification, the model penalty coefficient is 1, and the model kernel function is radial basis function. This method can effectively improve the stability of lymph node metastasis prediction models, making lymph node metastasis prediction more accurate. However, this method requires the collection of a large number of preliminary sample data to build the model, but the large individual differences among breast cancer patients make it difficult to establish the preliminary model. [Summary of the Invention]

[0007] This invention provides a whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer, in order to solve the technical problems of inaccurate and inefficient prediction of neoadjuvant therapy efficacy.

[0008] This invention is achieved through the following technical solution, providing a whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer:

[0009] The steps of a whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer are as follows: S1, collect and organize DCE-MRI breast images of breast cancer patients before and after neoadjuvant therapy; S2, perform pixel matrix operations on the images collected and organized in S1 and use a region growing algorithm to retain the images of the pathological area of ​​the breast and the axillary region; S3, organize the images obtained in S2, and then input the organized images into a neural network for model training to obtain a prediction model; S4, perform feature fusion and classification prediction on the model data of breast cancer patients before and after neoadjuvant therapy obtained in S3 to obtain predictions of effective pathological complete remission and whether lymph node metastasis has occurred.

[0010] The beneficial effects are: the adoption of fully automated feature extraction technology greatly reduces the workload of physicians and increases the robustness of the model. In addition, it skips the automatic segmentation of deep learning and instead analyzes the entire image, which can extract the tumor microenvironment and peritumoral features, and extract information of the entire breast, including lymph nodes. This enables the prediction of whether neoadjuvant therapy will achieve pathological complete remission and whether lymph node metastasis will occur, and greatly increases the amount of information extracted. It provides more sufficient information for predicting whether effective pathological complete remission will be achieved after neoadjuvant therapy, thus improving the accuracy of the prediction.

[0011] Furthermore, the prediction of whether S4 lymph node metastasis occurs includes the prediction of axillary lymph nodes and sentinel lymph nodes.

[0012] The beneficial effects are: it can simultaneously predict axillary lymph nodes, sentinel lymph nodes, and lymph nodes, and the effective information extraction is more comprehensive. While predicting whether neoadjuvant therapy can achieve complete pathological remission and whether lymph nodes will metastasize, it also improves the accuracy of prediction.

[0013] Furthermore, S1 includes: S11, selecting the first time window image and the image at the peak signal intensity on the TIC curve for DCE-MRI breast imaging of breast cancer patients before and after neoadjuvant therapy; S12, normalizing the image obtained in S11 using the Z-score method; S13, normalizing the gray value range of each pixel in the normalized image.

[0014] The beneficial effects are: using Z-score normalization facilitates data comparison and helps improve the accuracy of subsequent analysis. Furthermore, selecting the first time window image and the image at which the signal intensity reaches its peak allows for more effective and accurate delineation of lesion areas.

[0015] Furthermore, in S13, the grayscale value range is normalized by calculating the grayscale histogram and then equalizing it.

[0016] The beneficial effect is that by using a grayscale histogram to display the frequency of pixel grayscale occurrences and performing equalization, the visibility of the image is effectively improved.

[0017] Furthermore, S2 includes: S21, taking the images of breast cancer patients collected and organized in S1, calculating the difference between the pixel matrix of the first time window image on the TIC curve and the image where the signal intensity reaches its peak, the side with the larger change in pixel value is the lesion side, and deleting the breast image of the side without tumor; S22, processing the image obtained in S21 using a region growing algorithm, retaining the image matrix pixels of the pathological breast and axilla, or S22, processing the image obtained in S21 using a gradient vector flow algorithm, retaining the image matrix pixels of the pathological breast, or S22, processing the image obtained in S21 using a watershed algorithm, retaining the image matrix pixels of the pathological breast.

[0018] The beneficial effects are: by solving the matrix by finding the difference between the image pixel matrix, the lesion can be determined conveniently and accurately; the region growing algorithm has a good segmentation effect and can flexibly and accurately retain the image matrix pixels according to the image conditions to be processed; the gradient vector flow algorithm can interact directly with the model, and the model expression is compact and the implementation speed is fast; the watershed algorithm can also accurately and flexibly retain the image matrix pixels.

[0019] Furthermore, in S22, the region growing algorithm uses the similarity between image pixels for segmentation and calls the regiongrow function to process the image obtained in S21.

[0020] The beneficial effects are: it can effectively avoid the disadvantage of spatial discontinuity in image segmentation and has better regional features.

[0021] Furthermore, S3 includes: S31, normalizing the image matrix obtained from S2 using the Z-score method and normalizing the grayscale value range of each pixel; S32, inputting the image obtained from S31 into a Siamese network to construct a neural network model, passing two inputs through two neural networks to make them have the same architecture, hyperparameters and weights.

[0022] The beneficial effects are: using Z-score normalization facilitates data comparison and helps improve the accuracy of subsequent analysis. Furthermore, using a Siamese network model can also improve robustness.

[0023] Furthermore, when constructing convolutional neural networks in Siamese networks, the base network framework uses densely connected convolutional networks, or the base network framework uses VGG16.

[0024] The benefits are: DenseNet has fewer parameters through feature reuse and bypass settings, resulting in less computation and significant bandwidth savings. Using VGG16 reduces network parameters, enhances the network's fitting ability, extracts more information, and controls the scale of computation.

[0025] Furthermore, S4 includes S41, fusing features using bilinear pooling followed by two independent fully connected layers, and S42, using SoftMax classification to process the image data obtained in S41 to obtain predictions of effective pathological complete remission and lymph node metastasis in breast cancer patients.

[0026] The beneficial effects are: it can accurately identify and extract information data, ensure the accuracy of feature extraction, and select the most accurate prediction results. [Attached Image Description]

[0027] Figure 1 This is a flowchart of the whole breast information analysis method according to Embodiment 1 provided by the present invention;

[0028] Figure 2 This is a flowchart of the processing of two DCE-MRI images before neoadjuvant therapy in S2 of Embodiment 1 provided by the present invention;

[0029] Figure 3 This is a flowchart of the processing of two DCE-MRI images after neoadjuvant therapy in S2 of Embodiment 1 provided by the present invention;

[0030] Figure 4 This is a schematic diagram of the neural network architecture of Embodiment 1 provided by the present invention;

Detailed Implementation Methods

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings.

[0032] Specific Embodiment 1 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by this invention:

[0033] This embodiment provides a whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer, as shown in the attached figure. Figure 1 As shown, the whole breast information analysis method includes the following steps:

[0034] S1, collect and organize DCE-MRI (contrast-enhanced magnetic resonance imaging) breast images of breast cancer patients, both those who have undergone and those who have undergone neoadjuvant therapy.

[0035] Specifically, S1 includes:

[0036] S11. For DCE-MRI breast imaging of breast cancer patients before and after neoadjuvant therapy, images from the first time window at the start of the scan and images at the peak signal intensity were selected on the TIC curve (time-signal intensity curve). Selecting the first time window at the start of the scan minimizes the impact of intravenously injected contrast agent on the lesion area. Selecting the image at the peak signal intensity maximizes the impact of intravenously injected contrast agent on the lesion area. By selecting images from the first time window and images at the peak signal intensity, the lesion area can be effectively delineated.

[0037] S12, the image obtained in S11 is normalized using the Z-score method, that is:

[0038]

[0039] Where x is the individual observation, σ is the standard deviation of the data, and μ is the mean of the population data;

[0040] S13, normalize the grayscale value range of each pixel in the normalized image. Specifically, calculate the grayscale histogram and perform grayscale histogram equalization. The numerical formula for the grayscale histogram is:

[0041]

[0042] Where r represents k Dispersion gray level, n k r represents the number of pixels in the image that exhibit different gray levels. k Represents the total number of pixels in the image. That is, frequency;

[0043] When equalizing a gray-level histogram, the cumulative distribution function is used as the gray-level transformation function, and its formula is:

[0044] F X (x)=P(X≤x)

[0045] S2 takes the images of breast cancer patients collected and organized in S1, performs pixel matrix operations on the corresponding images, and uses a region growing algorithm to preserve the images of the breast pathological area and the axillary area.

[0046] Specifically, S2 includes:

[0047] S21. Take the images of breast cancer patients collected and organized in S1, and calculate the difference between the pixel matrix of the image in the first time window on the TIC curve and the image where the signal intensity reaches its peak. The formula is:

[0048]

[0049] Dxij That is, the absolute value of the difference between corresponding pixels in two images. The side with the larger change in pixel value (left / right) is the lesion side. This determines whether the patient's tumor is in the left or right breast. At the same time, the image of the breast without the tumor is cut off and deleted.

[0050] S22, as attached Figure 2 and 3 As shown, the image obtained from S21 is processed using the region seeding (RSG) algorithm, preserving the image matrix pixels of the pathological breast and axilla. A key issue in the practical application of the region seeding algorithm is the selection of the seed and the determination of the similarity region determination criteria. The specific steps of the region seeding algorithm are as follows: Select the first unassigned pixel (x, y), and consider whether the relationship between its 8 neighboring pixels satisfies the growth criteria. Pixels that satisfy the growth criteria are set to the same region as (x, y), and so on, until all pixels in the image have been assigned a region, at which point the growth process ends.

[0051] When using the region growing algorithm, segmentation is performed based on the pixel similarity of the obtained image. The regiongrow function is called to process the image obtained by S21, with S=255 and T=80, where S is the seed value and T is the threshold. The entire pathological breast and axillary regions are segmented, and the remaining pixel matrix is ​​set to zero.

[0052] S3: Organize the image obtained in S2, and then input the organized image into the neural network for model training to obtain the prediction model.

[0053] Specifically, S3 includes:

[0054] S31, the breast pathological area and axillary image matrix obtained in S2 before and after neoadjuvant treatment are normalized using the Z-score method, and the gray value range of each pixel in the normalized image is normalized. The Z-score gray value range normalization processing method is the same as the image processing methods in S12 and S13.

[0055] S32 inputs the image obtained in S31 into a Siamese network to construct a neural network model. Two inputs are passed through two neural networks, ensuring they have the same architecture, hyperparameters, and weights. Inputting the image without neoadjuvant treatment into the model yields a predicted output value 1, while inputting the image with neoadjuvant treatment yields a predicted output value 2. Both predicted output values ​​1 and 2 are image feature vectors obtained from the images processed by the neural networks. and The vector, used to calculate its loss function, is as follows:

[0056]

[0057] in Right now Representing two sample features and The Euclidean distance between the two samples is given, where P represents the feature dimension of the samples, Y is the label indicating whether the two samples match (Y=1 indicates a match, Y=0 indicates a mismatch), m is the threshold, and N is the number of samples. When Y=1 (i.e., when the samples are similar), the loss function is:

[0058]

[0059] That is, the loss value increases when the samples are dissimilar. When Y = 0 (i.e., the samples are dissimilar), the loss function is:

[0060]

[0061] When constructing convolutional neural networks in Siamese networks, the base network framework uses a densely connected convolutional network (DenseNet). In a densely connected convolutional network, each layer is connected to all preceding layers along the channel dimension and serves as the input to the next layer, i.e.:

[0062] x l =H l ([x0,x1,…,x l-1 ])

[0063] H l (·) represents a nonlinear transformation function, which includes batch normalization, ReLU activation, pooling, and convolution.

[0064] In this embodiment, the DenseNet-121 network is used, as shown in the attached diagram. Figure 4As shown, its network structure is as follows: first, after the FeatureBlock, it reaches the first DenseBlock, then passes through a TransitionBlock to enter the next DenseBlock (×12), and similarly passes through a TransitionBlock to enter the third DenseBlock (×24), then passes through another TransitionBlock to reach the fourth DenseBlock (×16), and finally reaches a 7×7 average pool to obtain a pooling layer with stride = 1 and a fully connected layer. The growthrate = 32 used indicates the number of feature maps output by each layer in each DenseBlock.

[0065] In this embodiment, the FeatureBlock includes convolution and pooling processes. The convolution is 7×7 with a stride of 2, and the pooling is 3×3 max pooling with a stride of 2. The DenseBlock contains six [1×1conv, 3×3conv] layers, which form the Bottleneck structure. The 1×1conv operation before the 3×3conv layer aims to reduce the number of input feature maps, thereby reducing dimensionality and computation while fusing features from various channels. The TransitionBlock sits between the two DenseBlocks, acting as a connector and controlling the model's complexity. It consists of a 1×1 convolutional layer with a stride of 2 and a 2×2 (average pooling) pooling layer with a stride of 2.

[0066] S4 uses the model data obtained in S3 before and after neoadjuvant therapy for breast cancer patients to perform feature fusion and classification prediction to obtain effective predictions of pathological complete remission and whether lymph node metastasis has occurred.

[0067] Specifically, S4 includes:

[0068] S41: Features are fused using bilinear pooling followed by two independent fully connected layers. Bilinear pooling achieves deep, high-level feature fusion, obtaining second-order statistical information through outer product. When first-order information is the same, the difference in second-order information is used for classification. The formula for classifying the bilinearly pooled vector is as follows:

[0069]

[0070] A bilinear model consists of a quadruple, b = (f A ,fB ,P,C), where f A ,f B The features come from different sources, P is the pooling operation, and C is the classifier. Specifically, the two different features are bilinearly mixed to obtain matrix b, where T is the number of channels. Then, sum pooling is performed on all the newly obtained feature matrices b to obtain a new matrix, which is treated as a vector, denoted as bilinearvector(x). After moment normalization and L2 normalization, the feature z is obtained. The obtained feature z is input into two independent fully connected layers, where the weight matrix is ​​multiplied by the input vector and the bias is added, as shown in the formula:

[0071] z j =w j ·z+b j

[0072] Among them, Z j For the input of the fully connected layer, w j This is the weight matrix, representing the importance of each feature dimension and its influence on the final score. The score for each category is obtained by weighted summation of the features. j This is a bias term.

[0073] S42 uses SoftMax classification on the two independent fully connected layers obtained from S41 to predict whether the patient has achieved pathological complete remission after neoadjuvant therapy and whether there is lymph node metastasis (including axillary lymph nodes and sentinel lymph nodes). The classification formula using SoftMax layers is as follows:

[0074]

[0075] Its cross-entropy function is:

[0076] Among them, y c p represents the true value. c softmax(z) j The cross-entropy function is used to output the loss value, which is then used in subsequent backpropagation to calculate the gradient and update the weights. Ultimately, this yields predictions of whether the patient can achieve effective pathological complete remission through neoadjuvant therapy and whether lymph node metastasis has occurred.

[0077] Specific Embodiment 2 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by this invention:

[0078] The only difference between this embodiment and Embodiment 1 is that in S4, the axillary lymph node images are primarily analyzed and preserved. In other embodiments, the sentinel lymph node images may also be primarily analyzed.

[0079] Specific embodiment 3 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by the present invention:

[0080] The only difference between this embodiment and Embodiment 1 is that grayscale value transformation normalization is used in S12 for image processing to improve image clarity. Other processing methods may also be used in other embodiments.

[0081] Specific embodiment 3 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by the present invention:

[0082] The only difference between this embodiment and Embodiment 1 is that S22 processes the image obtained in S21 using the Gradient Vector Flow (GVF) algorithm, preserving the image matrix pixels of the pathological breast tissue. This allows for direct interaction with the model, resulting in a compact model expression and fast implementation speed. The specific function of the Gradient Vector Flow (GVF) algorithm is as follows:

[0083]

[0084] in u represents the gradient of the edge image. x ,u y Let v represent the gradient value of the gradient field obtained by GVF at the position (x,y) along the x-direction. x ,v y This represents the gradient value of the gradient field obtained by GVF at the position (x,y) in the y direction.

[0085] Its iterative solution:

[0086]

[0087]

[0088] in Δt represents the time interval between each iteration, Δx and Δy represent the pixel intervals, μ is the regularization parameter, and i, j, n represent x, y, and t.

[0089] Specific embodiment 4 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by the present invention:

[0090] The difference between this embodiment and Embodiment 1 is that S22 processes the image obtained in S21 using a watershed algorithm, preserving the image matrix pixels of the pathological breast. First, all pixels in the gradient image are classified according to their grayscale values, and a geodesic distance threshold is set. Second, the pixel with the smallest grayscale value is found, and the threshold is increased starting from the minimum value. During the increase, the geodesic distance to the pixel is calculated. If the distance is less than the set threshold, these pixels are submerged; otherwise, a dam is set on these pixels. This classifies the neighboring pixels.

[0091] Specific embodiment 5 of the whole breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer provided by the present invention:

[0092] The only difference between this embodiment and Embodiment 1 is that when constructing the convolutional neural network in the Siamese network, the basic network framework uses VGG16. VGG16 has small convolutional kernels (3×3) and small pooling kernels (2×2), which can reduce network parameters. On the one hand, it can perform more nonlinear mappings, thereby enhancing the network's fitting ability. On the other hand, VGG16 also has the characteristics of having a large number of channels, deep layers, and wide feature maps, which can extract more information. In addition, because it has small convolutional kernels and small pooling kernels, the scale of computation can also be controlled.

Claims

1. A whole-breast information analysis method for predicting the efficacy of neoadjuvant therapy for breast cancer, characterized in that, The steps of the whole breast information analysis method are as follows: S1, collect and organize DCE-MRI breast imaging data before and after neoadjuvant therapy; S2, perform pixel matrix operations on the images collected and organized in S1 and use a region growing algorithm to preserve the images of the breast pathological area and the axillary region; S3, organize the images obtained in S2, and then input the organized images into a neural network for model training to obtain a prediction model; S3 includes: S31, normalize the image matrix obtained in S2 using the Z-score method and normalize the gray value range of each pixel; S32, input the images obtained in S31 into a twin network. S3 involves constructing a neural network model, passing two inputs through two neural networks, ensuring they have the same architecture, hyperparameters, and weights; S4 involves fusing features from the model data of breast cancer patients before and after neoadjuvant therapy obtained in S3 and performing classification predictions to obtain effective predictions of complete pathological remission and lymph node metastasis; S4 includes: S41, fusing features using bilinear pooling followed by two independent fully connected layers; S42, using SoftMax classification to process the image data obtained in S41 to obtain predictions of effective complete pathological remission and lymph node metastasis for breast cancer patients.

2. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 1, characterized in that, The prediction of whether lymph node metastasis occurs in S4 includes the prediction of axillary lymph nodes and sentinel lymph nodes.

3. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 2, characterized in that, S1 includes: S11, selecting the first time window image and the image at the peak signal intensity on the TIC curve for DCE-MRI breast imaging of breast cancer patients before and after neoadjuvant therapy; S12, normalizing the image obtained in S11 using the Z-score method; S13, normalizing the gray value range of each pixel in the normalized image.

4. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 3, characterized in that, The grayscale value range normalization method in S13 is to calculate the grayscale histogram and perform grayscale histogram equalization.

5. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 4, characterized in that, S2 includes: S21, taking the images of breast cancer patients collected and organized in S1, calculating the difference between the pixel matrix of the first time window image on the TIC curve and the image where the signal intensity reaches its peak, the side with the larger change in pixel value is the lesion side, and deleting the breast image of the side without tumor; S22, processing the image obtained in S21 using a region growing algorithm, retaining the image matrix pixels of the pathological breast and axilla, or S22, processing the image obtained in S21 using a gradient vector flow algorithm, retaining the image matrix pixels of the pathological breast, or S22, processing the image obtained in S21 using a watershed algorithm, retaining the image matrix pixels of the pathological breast.

6. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 5, characterized in that, In S22, the region growing algorithm uses the similarity between image pixels for segmentation and calls the regiongrow function to process the image obtained in S21.

7. The method for whole-breast information analysis for predicting the efficacy of neoadjuvant therapy for breast cancer according to claim 1, characterized in that, When constructing the convolutional neural network in the Siamese network, the base network framework uses a densely connected convolutional network, or the base network framework uses VGG16.

Images