A method and system for classifying tumor ultrasound images

Abstract

This invention relates to the field of digital image processing technology, specifically to a method and system for classifying tumor ultrasound images. The tumor ultrasound image classification method, applied to a server, mainly includes: acquiring and preprocessing a dataset composed of tumor ultrasound images; inputting the preprocessed dataset based on a pre-constructed multi-region image recognition model; and fusing information from the tumor's internal and surrounding regions in the preprocessed dataset for automatic classification. This invention can preprocess ultrasound images and then, through a multi-region image recognition model, fully fuse image information from the tumor's interior and surrounding areas to classify the ultrasound images, significantly improving classification accuracy.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, specifically to a method and system for classifying tumor ultrasound images. Background Technology

[0002] Digital image processing technology has developed rapidly and is used in many fields. For example, it is widely used in facial recognition. Furthermore, digitizing medical images to achieve automatic classification also has broad application prospects.

[0003] In ultrasound images of tumors, information about the area surrounding the tumor is crucial. Current tumor image classification methods primarily focus on acquiring information from individual regions, neglecting the advantage of effectively fusing information from both the tumor's internal (intratumoral) and surrounding (peritumoral) areas. This leads to low accuracy in tumor ultrasound image classification. Summary of the Invention

[0004] The purpose of this invention is to provide a tumor ultrasound image classification method and system to overcome the problem that existing tumor image classification methods rely solely on acquiring information from a single region, resulting in poor accuracy in tumor ultrasound image classification.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A tumor ultrasound image classification method, applied to a server, the method includes:

[0007] Acquire a dataset consisting of tumor ultrasound images and preprocess it;

[0008] The pre-processed dataset is input based on a pre-built multi-region image recognition model.

[0009] The multi-region image recognition model integrates preprocessed information from the tumor's internal region and the region surrounding the tumor in the dataset for automatic classification.

[0010] The above scheme preprocesses ultrasound images and then uses a multi-region image recognition model to fully integrate image information from inside and around the tumor to classify the ultrasound images, which can greatly improve the classification accuracy.

[0011] As a further aspect of the present invention, the preprocessing steps are as follows:

[0012] Fill along the shorter side of the tumor ultrasound image to make the height and width equal;

[0013] Adjust the size of the tumor ultrasound image and the size of the segmentation labels;

[0014] The tumor outline is delineated, and combined tumor region labels are obtained by expanding the delineated tumor outline.

[0015] Subtract the intratumoral region label from the combined tumor region label to obtain the peritumoral region label.

[0016] As a further aspect of the present invention:

[0017] In the multi-region image recognition model, the processed tumor image is first processed by the first residual module and the second residual module in sequence, and then output to the first information enhancement module. After processing by the first information enhancement module, it is output to the third residual module. After processing by the third residual module, it is output to the second information enhancement module. After processing by the second information enhancement module, it is transmitted to the fourth residual module. After processing by the fourth residual module, it is transmitted to the multi-region image information fusion module to classify the tumor image.

[0018] The information enhancement module aims to progressively enhance the features of combined tumor regions. Simultaneously, the multi-region information fusion module extracts and fuses features from intratumoral, peritumoral, and combined tumor regions. A channel attention mechanism adaptively fuses features from these three regions, thereby improving classification accuracy.

[0019] As a further aspect of the present invention: the first information enhancement module and the second information enhancement module include a supervision module, and the execution steps of the supervision module are as follows:

[0020] Obtain the original feature map of the tumor ultrasound image;

[0021] Based on the acquired original image feature map, it is added to the original input feature map to obtain the first output feature map.

[0022] As a further aspect of the present invention: the acquisition of the original image feature map includes:

[0023] The original image feature map is obtained using formula (1);

[0024]

[0025] Among them, F comb G represents the original image feature map obtained by the supervision module. C*H*W G′ represents the input feature map of the supervision module. 1*H*W This represents the label feature map obtained through supervision;

[0026] The process of obtaining the first output feature map includes:

[0027] Based on the obtained original image feature map and the supervised feature map, the first output feature map is obtained through formula (2);

[0028]

[0029] Where F represents the original first input feature map, and F′ represents the first output feature map. This indicates the element-wise addition of feature maps.

[0030] As a further aspect of the present invention: in the supervision module, the supervision module uses 1×1 convolution to compress the number of channels to 1, and uses the sigmoid activation function to obtain the feature map.

[0031] As a further aspect of the present invention: the multi-region information fusion module includes:

[0032] Obtain at least one branch fusion coefficient;

[0033] Based on the fusion coefficients of the aforementioned branches, a second output feature map is obtained.

[0034] As a further aspect of the present invention: obtaining the second output feature map includes:

[0035] At least one branch fusion system is multiplied with the second input feature map, and then concatenated through the feature maps.

[0036] As a further aspect of the present invention: before transmitting the preprocessed dataset to the multi-region image recognition model fusion step, the method further includes: feeding the original tumor ultrasound image, classification label, and tumor multi-region image segmentation label into the multi-region image recognition model, using the sum of the losses of the supervision module, enhancement module, and multi-region fusion module as the final total loss, and backpropagating to update the parameters to optimize the network.

[0037] This application also provides a tumor ultrasound image classification system, including:

[0038] The acquisition module is used to acquire and preprocess a dataset composed of tumor ultrasound images;

[0039] The input module is used to transmit the preprocessed dataset to the multi-region image recognition model;

[0040] The classification module is used to automatically classify the preprocessed dataset using a multi-region image recognition model.

[0041] In summary, the embodiments of the present invention have the following beneficial effects compared with the prior art:

[0042] 1. The tumor ultrasound image classification method provided by the present invention first preprocesses the ultrasound image, and then uses a multi-region image recognition model to fully integrate the image information inside and around the tumor to classify the ultrasound image, which can greatly improve the classification accuracy.

[0043] 2. The tumor ultrasound image classification method provided by this invention is a novel medical image classification method capable of fusing and classifying multi-region tumor images. The multi-region image recognition model fusion module adopts a residual classification network framework. To fully explore the potential relationships between multi-region images, an information enhancement module for combined tumor regions (intratumoral and peritumoral regions) is introduced, aiming to progressively enhance the features of combined tumor regions. Simultaneously, a multi-region information fusion module is used to extract and fuse features from intratumoral, peritumoral, and combined tumor regions, and a channel attention mechanism adaptively fuses features from the three regions.

[0044] 3. Compared to existing image-based methods that require precise segmentation labels for the tumor periphery during the testing phase, which is cumbersome and time-consuming in clinical applications, the tumor ultrasound image classification method provided by this invention preprocesses ultrasound images and extracts and fuses features from intratumoral, peritumoral, and combined tumor regions, enabling adaptive fusion of information from multiple tumor regions without the use of segmentation labels. Attached Figure Description

[0045] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a schematic diagram of the tumor ultrasound image classification method provided by the present invention.

[0047] Figure 2 This is a schematic diagram of the structure of the tumor ultrasound image classification system provided by the present invention.

[0048] Figure 3 This is a schematic diagram of the structure of the multi-region image recognition model provided by the present invention.

[0049] Figure 4 The segmentation result of the supervision module provided by this invention.

[0050] Figure 5 The classification results of the multi-region image recognition model provided by this invention after training and testing on multiple datasets. Detailed Implementation

[0051] 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 and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0053] It should be understood that, although the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion in embodiments of the invention, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0054] The applicant has discovered that current tumor ultrasound image classification methods only aim to obtain information from a single region, resulting in inaccurate classification. Therefore, this invention provides a tumor ultrasound image classification method that involves acquiring and preprocessing a dataset composed of tumor ultrasound images; inputting the preprocessed dataset based on a pre-constructed multi-region image recognition model; and fusing information from the tumor's internal and surrounding regions within the preprocessed dataset for automatic classification. This invention preprocesses ultrasound images and then, through a multi-region image recognition model, fully integrates image information from both the tumor's interior and surrounding areas to classify the ultrasound images, significantly improving classification accuracy.

[0055] The tumor ultrasound image classification method will be described in detail below with reference to the specific accompanying drawings.

[0056] Example 1

[0057] See Figure 1 As shown, a tumor ultrasound image classification method, mainly applied to a server, includes the following steps:

[0058] S10. Obtain the dataset and preprocess it;

[0059] In step S1, the dataset mainly consists of a series of tumor ultrasound images. To ensure dataset diversity, images can be collected beforehand from patients of different ages and durations of illness using existing ultrasound equipment.

[0060] The preprocessing steps are as follows:

[0061] S11. In order to unify the input size of the tumor ultrasound image, first fill along the shorter side of the tumor ultrasound image to make the height and width equal.

[0062] S12. Next, the original tumor ultrasound image is adjusted to 224×224 using bilinear interpolation, and the segmentation labels are adjusted to 224×224 using nearest neighbor interpolation.

[0063] Bilinear interpolation and nearest neighbor interpolation are existing technologies and will not be described in detail here.

[0064] S13. Then, outline the tumor and obtain combined tumor region labels by performing a standard morphological expansion operation of 1.5 times the edge on the outlined tumor outline.

[0065] S14. Finally, subtract the intratumoral region label from the combined tumor region label to obtain the peritumoral region label. The intratumoral region label can be obtained using existing computer equipment, so it will not be described in detail here.

[0066] Furthermore, in this application, the dot product of the original image of the tumor region and the label is the tumor image of the corresponding region.

[0067] It is understandable that the outline of a tumor can be drawn manually or using computer equipment. This part is existing technology and will not be explained in detail.

[0068] It should be noted that, considering the size of the dataset, methods such as horizontal flipping, vertical flipping, and random rotation between -30 and 30 degrees are used to increase the size of the training set.

[0069] S20. Based on a pre-built multi-region image recognition model, input the pre-processed dataset;

[0070] See Figure 3 As shown, in the multi-region image recognition model, the processed tumor image is first processed by the first residual module and the second residual module in sequence, and then output to the first information enhancement module. After processing by the first information enhancement module, it is output to the third residual module. After processing by the third residual module, it is output to the second information enhancement module. After processing by the second information enhancement module, it is transmitted to the fourth residual module. After processing by the fourth residual module, it is transmitted to the multi-region image information fusion module to classify the tumor image.

[0071] To facilitate understanding of step S2 of this application, the construction of the tumor multi-region image recognition model will be described in detail below.

[0072] The main steps in constructing a multi-region image recognition model are as follows:

[0073] S21. Construct a residual baseline classification network model; In this application, the residual baseline classification network model may include a first residual module, a second residual module, a third residual module, and a fourth residual module.

[0074] In this application, the structure of each residual module is based on a residual network as the baseline network for classification.

[0075] Residual network models refer to encoding networks similar to ResNet. The encoding stage of the original ResNet network consists of a series of convolutional layers, batch normalization layers, ReLU nonlinear activation functions, and max pooling layers.

[0076] In this invention, for compatibility and to maintain more high-resolution detail, the first 7×7 convolutional layer of ResNet34 is replaced with two 3×3 convolutional layers that maintain their size. This improves the feature extraction capability of the ResNet network and avoids gradient vanishing and gradient exploding phenomena.

[0077] Furthermore, the ResNet network mentioned in this application is an existing network, so a detailed description of the ResNet network will not be provided here.

[0078] S22. Construct an information enhancement module to enhance information on the combined intratumoral and peritumoral tumor regions.

[0079] This invention constructs an information enhancement module to fully extract intratumoral and peritumoral information in the tumor region and improve the accuracy of classifying tumors as benign or malignant.

[0080] The information enhancement module mainly consists of two parts: the supervision module and the enhancement module.

[0081] The execution steps of the supervision module are as follows:

[0082] S221. Obtain the original tumor ultrasound image feature map using formula (1);

[0083]

[0084] Among them, F comb G represents the original tumor ultrasound image feature map obtained by the supervision module. C*H*W G′ represents the input feature map of the supervision module. 1*H*W This represents the label feature map obtained through supervision.

[0085] S222. Based on the acquired original image feature map, add it to the original input feature map to obtain the first output feature map. Specifically, based on the obtained tumor ultrasound image feature map and the original first input feature map, obtain the output feature map through formula (2).

[0086]

[0087] Where F represents the original first input feature map, and F′ represents the first output feature map. This indicates the element-wise addition of feature maps.

[0088] The supervision module plays an important role in the information enhancement module of the combined area.

[0089] See Figure 4 The images processed by the supervision module allow for better observation of the areas surrounding and inside the tumor.

[0090] The supervised module proposed in this application uses a U-shaped module to segment labels in a supervised manner during the training phase, which solves the problem that current classification models based on the peritumoral region still need to segment labels during the testing phase.

[0091] The U-shaped module is existing technology and will not be discussed further here.

[0092] Secondly, this application progressively samples the feature maps to the size of the input image. Considering the number of computational parameters, the number of channels in the model can be gradually reduced through decoder blocks.

[0093] Finally, the supervision module uses 1×1 convolutions to compress the number of channels to 1 and uses the sigmoid activation function to obtain feature maps.

[0094] After the original feature map is processed by multiple multi-scale convolutional layers in the supervised module, the inventors use a 1×1 convolutional layer and a sigmoid activation layer to generate a segmentation label map. Finally, the original feature map is multiplied by the attention map to obtain the output feature map.

[0095] The structure of the enhancement module in this application mainly serves to enhance the characteristics of the combined tumor region.

[0096] S23. Construct an information fusion module to fuse image information from multiple tumor regions;

[0097] Step S23 mainly uses the information fusion module to fully extract multi-region image information of the tumor and explores an effective method for fusing multi-region image information of the tumor.

[0098] The multi-regional information fusion module can be summarized as follows:

[0099] S231. Obtain the fusion coefficients of several branches using formula (3);

[0100] α i =CA(C(F′) i )), i = [1, 2…n] (3)

[0101] i represents several distinct branches, n is a positive integer, C represents the convolution operation, and CA represents the channel attention module.

[0102] S232. Multiply at least one branch fusion system with the second input feature map, and then concatenate the feature maps to obtain the output feature map of the fusion module through formula (4), as follows:

[0103]

[0104] Where F′ i The second input feature map is represented by (which can be obtained from formula (2)), and Con represents the concatenation of feature maps. F″ represents element-wise multiplication, and F″ represents the second output feature map.

[0105] In this step, the multi-region information fusion module can be placed in a general network to update parameters through self-learning and explore more effective multi-region information fusion methods.

[0106] In the network structure design of the information fusion module, a supervision module can be introduced to supervise the segmentation labels of multiple tumor regions, and three branches can be used to extract features from the three tumor regions respectively. That is, i = [1, 2, 3] in this application.

[0107] In addition, the applicant found that the features of different tumor regions are quite similar, making it impossible to effectively fuse feature maps. To address this, the applicant introduces a one-dimensional channel attention mechanism using formulas (3) and (4) to model the channel dependencies between feature maps of different tumor regions and uses the generated channel weights to effectively fuse feature maps of different regions.

[0108] Furthermore, the one-dimensional channel attention mechanism can not only effectively reduce the number of parameters, but also ensure sufficient local cross-channel interaction, significantly improving the performance of the model.

[0109] It is worth mentioning that by multiplying the channel weights generated by the one-dimensional attention mechanism model with the feature maps of different regions, and then adding them to the original image using residual connections, a feature map that effectively integrates contextual information at different scales can be obtained.

[0110] S24. Establish a multi-region image recognition model.

[0111] See Figure 1 Based on the baseline network established in steps S21, S22, and S23, this invention inserts an enhancement module and an information fusion module to establish a multi-region image recognition model for the entire tumor.

[0112] To enhance combined tumor region information across multiple stages of the multi-region image recognition model, this invention employs two information enhancement modules during the feature extraction stage. These two modules model long-term dependencies, guiding the entire multi-region image recognition model to gradually focus on combined tumor region information, utilizing intratumoral and peritumoral texture and edge information to obtain more accurate classification results.

[0113] The multi-region image fusion module is used in the final stage of the multi-region image recognition model. It can capture and fully fuse the feature information of multiple regions of the tumor.

[0114] This application compresses the number of channels to 1 using 1×1 convolution and introduces an advanced channel attention mechanism through formulas (3) and (4), which can adaptively measure the contribution of different regional features to the final classification result, exploring and finding the most effective way to fuse multi-regional tumor information. The channel attention mechanism updates parameters through self-learning, guiding the optimization of the entire classification network.

[0115] Furthermore, before step S2 is executed, participants Figure 5 It also includes inputting multiple training sets to train a multi-region image recognition model.

[0116] Training a multi-region image recognition model requires feeding the original tumor ultrasound images and classification labels into the entire network for supervised learning, as well as the segmentation labels for the multi-region tumor images. The sum of the losses from the combined region information enhancement module and the multi-region fusion module is used as the final total loss, and backpropagation is used to update the parameters, guiding the optimization of the entire network.

[0117] To further optimize the model, the applicant designed a joint loss function that includes a segmentation loss function for supervising the three tumor region labels and a cross-entropy loss function for the classification task.

[0118] In particular, this application may include five segmentation loss functions: two for information enhancement modules of two intratumoral and peritumoral combined regions, used to supervise the combined region labels; and three for multi-region information fusion modules of the tumor, used to supervise the three tumor region labels.

[0119] The loss function is as follows:

[0120]

[0121] Where L C L represents the cross-entropy loss function. D L represents the segmentation loss function. total Let λ represent the joint loss of the entire network, and λ1 and λ2 represent the coefficients of each loss function.

[0122] For S30, the multi-region image recognition model fuses the preprocessed information of the tumor's internal region and the information of the region surrounding the tumor in the dataset to achieve automatic classification.

[0123] When testing the classification network, you only need to input the image to be tested; you do not need to input the tumor multi-region segmentation map labels. The classification network will automatically integrate the tumor multi-region information based on the test image to classify the tumor as benign or malignant.

[0124] Working principle:

[0125] The tumor ultrasound image classification method provided by this invention is a novel medical image classification method that can achieve fusion classification of tumor images from multiple regions.

[0126] The multi-region image recognition model fusion module employs a residual classification network framework. To fully explore the potential relationships between multi-region images, an information enhancement module for the combined tumor region (intratumoral and peritumoral region) is introduced, aiming to progressively enhance the features of the combined tumor region. Simultaneously, a multi-region information fusion module is used to extract and fuse features from the intratumoral, peritumoral, and combined tumor regions, and a channel attention mechanism adaptively fuses features from the three regions.

[0127] Simultaneously, it enables preprocessing of ultrasound images and can extract and fuse features from intratumoral, peritumoral, and combined tumor regions, achieving adaptive fusion of information from multiple tumor regions without using segmentation labels.

[0128] Example 2

[0129] This application also provides a tumor ultrasound image classification system, including:

[0130] Module 10 is used to acquire and preprocess a dataset composed of tumor ultrasound images; it is also used for:

[0131] The preprocessing steps are as follows:

[0132] Fill along the shorter side of the tumor ultrasound image to make the height and width equal;

[0133] Adjust the size of the tumor ultrasound image and the size of the segmentation labels;

[0134] The tumor outline is delineated, and combined tumor region labels are obtained by expanding the delineated tumor outline.

[0135] Subtract the intratumoral region label from the combined tumor region label to obtain the peritumoral region label.

[0136] Input module 20 is used to input the preprocessed dataset based on a pre-built multi-region image recognition model; and is also used to:

[0137] In the multi-region image recognition model, ultrasound images including combined tumor region labels (intratumoral and peritumoral) are first processed sequentially by a first residual module and a second residual module, and then output to a first information enhancement module. After processing by the first information enhancement module, the images are output to a third residual module. After processing by the third residual module, the images are output to a second information enhancement module. After processing by the second information enhancement module, the images are transmitted to a fourth residual module. After processing by the fourth residual module, the images are transmitted to a multi-region image information fusion module to classify the tumor images.

[0138] The first information enhancement module and the second information enhancement module include a supervision module. The execution steps of the supervision module are: acquiring the original tumor ultrasound image feature map; and calculating the first output feature map by adding the acquired original image feature map to the original input feature map.

[0139] The process of obtaining the original image feature map includes:

[0140] The original image feature map is obtained using formula (1);

[0141]

[0142] Among them, F comb G represents the original image feature map obtained by the supervision module. C*H*W G′ represents the input feature map of the supervision module. 1*H*W This represents the label feature map obtained through supervision;

[0143] The process of obtaining the first output feature map includes:

[0144] Based on the obtained original image feature map and the supervised feature map, the first output feature map is obtained through formula (2).

[0145]

[0146] Where F represents the original first input feature map, and F′ represents the first output feature map. This indicates the element-wise addition of feature maps.

[0147] In the supervision module, the number of channels is compressed to 1 using 1×1 convolution, and sigmoid activation is used to obtain the feature map.

[0148] The multi-regional information fusion module includes:

[0149] Obtain at least one branch fusion coefficient;

[0150] Based on the fusion coefficients of the aforementioned branches, a second output feature map is obtained.

[0151] Obtaining the second output feature map includes: multiplying at least one branch fusion system with the second input feature map, and then concatenating the feature maps.

[0152] Before transmitting the preprocessed dataset to the multi-region image recognition model fusion step, the method further includes: feeding the original tumor ultrasound image, classification label, and tumor multi-region image segmentation label to the multi-region image recognition model; using the sum of the losses of the supervision module, enhancement module, and multi-region fusion module as the final total loss; and backpropagating to update the parameters to optimize the network.

[0153] The classification module 30 is used to automatically classify the preprocessed dataset using a multi-region image recognition model.

[0154] Those skilled in the art will understand that information and signals can be represented using any number of different techniques and methods. For example, the data, instructions, commands, information, signals, bits, symbols, and chips described above can be represented by voltage, current, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0155] Those skilled in the art should also understand that the various illustrative logical functional blocks, modules, circuits, and algorithmic steps associated with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination thereof. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, functional blocks, modules, circuits, and steps are typically described according to their functionality. Whether this functionality is implemented in hardware or software depends on the specific application and design constraints of the entire system. Those skilled in the art can implement the described functionality in varying ways for each specific application, but such implementation should not be construed as a departure from the scope of the invention.

[0156] The technical principles of the present invention have been described above with reference to specific embodiments, which are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection scope. Those skilled in the art can conceive of other specific embodiments of the present invention without creative effort, and these embodiments will all fall within the protection scope of the present invention.

Claims

1. A method for classifying tumor ultrasound images, characterized in that, Applied to a server, the method includes: Acquire and preprocess a dataset consisting of tumor ultrasound images; The preprocessing steps are as follows: Fill along the shorter side of the tumor ultrasound image to make the height and width equal; Adjust the size of the tumor ultrasound image and the size of the segmentation labels; The tumor outline is delineated, and combined tumor region labels are obtained by expanding the delineated tumor outline. Subtract the intratumoral region label from the combined tumor region label to obtain the peritumoral region label; The preprocessed dataset is then transferred to the multi-region image recognition model fusion point. In the multi-region image recognition model, ultrasound images including combined tumor region labels (intratumoral and peritumoral) are first processed sequentially by a first residual module and a second residual module, and then output to a first information enhancement module. After processing by the first information enhancement module, the images are output to a third residual module. After processing by the third residual module, the images are output to a second information enhancement module. After processing by the second information enhancement module, the images are transmitted to a fourth residual module. After processing by the fourth residual module, the images are transmitted to a multi-region image information fusion module to classify the tumor images. The multi-region image recognition model integrates preprocessed information from the tumor's internal region and the region surrounding the tumor in the dataset to achieve automatic classification.

2. The tumor ultrasound image classification method according to claim 1, characterized in that, Both the first information enhancement module and the second information enhancement module include a supervision module, and the execution steps of the supervision module are as follows: Obtain the original feature map of the tumor ultrasound image; Based on the acquired original image feature map, it is added to the original input feature map to obtain the first output feature map.

3. The tumor ultrasound image classification method according to claim 2, characterized in that, The process of obtaining the original image feature map includes: The original image feature map is obtained using formula (1); Among them, F comb G represents the original image feature map obtained by the supervision module. C*H*W G′ represents the input feature map of the supervision module. 1*H*W This represents the label feature map obtained through supervision; The process of obtaining the first output feature map includes: Based on the obtained original image feature map and the supervised feature map, the first output feature map is obtained through formula (2); Where F represents the original first input feature map, and F′ represents the first output feature map. This indicates the element-wise addition of feature maps.

4. The method of using the tumor ultrasound image classification method according to claim 2, characterized in that, In the supervision module, the number of channels is compressed to 1 using 1×1 convolution, and the sigmoid activation function is used to obtain the feature map.

5. The tumor ultrasound image classification method according to claim 1, characterized in that, The multi-regional information fusion module includes: Obtain at least one branch fusion coefficient; Based on the fusion coefficients of the aforementioned branches, a second output feature map is obtained.

6. The method of using the tumor ultrasound image classification method according to claim 5, characterized in that, Obtaining the second output feature map includes: At least one branch fusion coefficient is multiplied with the second input feature map, and the second output feature map is obtained by the result of the concatenated multiplication of the feature maps.

7. The method of using the tumor ultrasound image classification method according to claim 1, characterized in that, Before transmitting the preprocessed dataset to the multi-region image recognition model fusion step, the method further includes: feeding the original tumor ultrasound image, classification label, and tumor multi-region image segmentation label into the multi-region image recognition model; using the sum of the loss functions of the supervision module, the first information enhancement module, the second information enhancement module, and the multi-region fusion module as the final total loss function; and backpropagating to update the parameters to optimize the network.

8. A tumor ultrasound image classification system, characterized in that, include: The acquisition module is used to acquire and preprocess a dataset composed of tumor ultrasound images. The preprocessing steps are as follows: Fill along the shorter side of the tumor ultrasound image to make the height and width equal; Adjust the size of the tumor ultrasound image and the size of the segmentation labels; The tumor outline is delineated, and combined tumor region labels are obtained by expanding the delineated tumor outline. Subtract the intratumoral region label from the combined tumor region label to obtain the peritumoral region label; The input module is used to transmit the preprocessed dataset to the multi-region image recognition model. The multi-region image recognition model includes ultrasound images with combined tumor region labels inside and around the tumor. The images are first processed by the first residual module and the second residual module, and then output to the first information enhancement module. After processing by the first information enhancement module, the images are output to the third residual module. After processing by the third residual module, the images are output to the second information enhancement module. After processing by the second information enhancement module, the images are transmitted to the fourth residual module. After processing by the fourth residual module, the images are transmitted to the multi-region image information fusion module to classify the tumor images. The classification module is used to automatically classify the preprocessed dataset using a multi-region image recognition model.

Images