Image processing method and device for breast cancer recognition

Abstract

The invention discloses an image processing method and device for breast cancer recognition. The method comprises the following steps: acquiring to-be-recognized image information; preprocessing the to-be-recognized image information to obtain initial recognition image information; performing identification processing on the initial identification image information based on a target image identification model to obtain a target image identification result; the target image recognition result is used for indicating analysis and recognition of the breast cancer. Therefore, the image recognition accuracy and the image processing efficiency can be improved, and then the breast cancer image analysis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image processing method and apparatus for breast cancer identification. Background Technology

[0002] Currently, mammography is the most common method for detecting breast cancer. However, early diagnosis is very difficult due to the large amount of data and poor imaging characteristics of early breast cancer. With the development of image processing technology and early diagnostic technology, image processing of breast pathology has become an important method for the early diagnosis of breast cancer. Because the boundaries of internal organs, muscles, blood vessels, and other organs are often very complex, the diseased areas of organs are often mixed with normal tissue, and the differences in grayscale and boundaries between the diseased areas and normal tissue are relatively small, making it difficult to identify these organs and diseased areas in medical images. Existing image recognition algorithms cannot independently complete the image recognition task. Therefore, this paper provides an image processing method and apparatus for breast cancer identification to improve image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an image processing method and apparatus for breast cancer identification, which is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0004] To address the aforementioned technical problems, a first aspect of the present invention discloses an image processing method for breast cancer identification, the method comprising: Obtain the image information to be recognized; The image information to be identified is preprocessed to obtain initial recognition image information; The initial image information is processed based on the target image recognition model to obtain the target image recognition result; the target image recognition result is used to indicate the analysis and recognition of breast cancer.

[0005] As an optional implementation, in a first aspect of the present invention, the preprocessing of the image information to be identified to obtain initial identification image information includes: The image information to be identified is sliced ​​to obtain sliced ​​image information; the sliced ​​image information includes M sliced ​​images; M is a positive even number; The sliced ​​image information is subjected to dimensionality reduction processing to obtain dimensionality-reduced image information; The sliced ​​image information is subjected to position encoding processing to obtain image position encoding information; The sliced ​​image information, the reduced-dimensional image information, and the image location encoding information are integrated to obtain the initial recognition image information.

[0006] As an optional implementation, in a first aspect of the present invention, the target image recognition model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and a target image perception and parsing sub-model; wherein, The input of the first target encoding sub-model is connected to the model input of the target image recognition model, and the output of the first target encoding sub-model is connected to the input of the second target encoding sub-model; the output of the second target encoding sub-model is connected to the input of the third target encoding sub-model; the output of the third target encoding sub-model is connected to the input of the fourth target encoding sub-model and the input of the fourth sampling module; the output of the fourth target encoding sub-model is connected to the input of the fifth target encoding sub-model; the output of the target encoding sub-model is connected to the input of the sixth target encoding sub-model and the input of the third sampling module; the output of the sixth target encoding sub-model is connected to the input of the seventh target encoding sub-model; the output of the seventh target encoding sub-model is connected to the input of the eighth target encoding sub-model and the input of the second sampling module; the output of the eighth target encoding sub-model is connected to the... The input terminal of the ninth target encoding sub-model is connected to the input terminal of the tenth target encoding sub-model; the output terminal of the tenth target encoding sub-model is connected to the input terminal of the eleventh target encoding sub-model; the output terminal of the eleventh target encoding sub-model is connected to the input terminal of the first sampling module; the output terminals of the first, second, third, and fourth sampling modules are respectively connected to the input terminals of the first, second, third, and fourth convolution modules; the output terminals of the first, second, third, and fourth convolution modules are all connected to the input terminal of the target image perception and analysis sub-model; the output terminal of the target image perception and analysis sub-model is connected to the model output of the target image recognition model.

[0007] As an optional implementation, in the first aspect of the present invention, the model architecture of the first target coding sub-model to the eleventh target coding sub-model is consistent; The first target encoding sub-model includes a first fusion normalization module, a multi-head attention module, a second fusion normalization module, a first neural network, and a third fusion normalization module; wherein, The input of the first fusion normalization module is connected to the model input of the target image recognition model; the output of the first fusion normalization module is connected to the input of the multi-head attention module and the input of the second fusion normalization module; the output of the multi-head attention module is connected to the input of the second fusion normalization module; the output of the second fusion normalization module is connected to the input of the first neural network and the input of the third fusion normalization module; the output of the third fusion normalization module is connected to the input of the second target encoding sub-model. The target image perception and parsing sub-model includes a first pooling module, a fourth fusion normalization module, a fifth fusion normalization module, a sixth fusion normalization module, a feature fusion module, a fifth convolution module, and a classification module; wherein, The input of the first pooling module is connected to the output of the fourth convolution module. The output of the first pooling module is connected to the input of the fourth fusion normalization module and the input of the feature fusion module. The input of the fourth fusion normalization module is also connected to the output of the third convolution module. The output of the fourth fusion normalization module is also connected to the input of the fifth fusion normalization module and the input of the feature fusion module. The input of the fifth fusion normalization module is also connected to the output of the second convolution module. The output of the fifth fusion normalization module is also connected to the input of the sixth fusion normalization module and the input of the feature fusion module. The input of the sixth fusion normalization module is also connected to the output of the first convolution module. The output of the sixth fusion normalization module is connected to the input of the feature fusion module. The output of the feature fusion module is connected to the input of the fifth convolution module. The output of the fifth convolution module is connected to the input of the classification module. The output of the classification module is connected to the model output of the target image recognition model.

[0008] As an optional implementation, in the first aspect of the present invention, the target image recognition model is obtained based on the following steps: Obtain the initial sample set and initialize the target image recognition model; The initial sample set is augmented to obtain the target sample set; The parameters of the initialized target image recognition model are updated based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model.

[0009] As an optional implementation, in a first aspect of the present invention, the first training model includes a first training encoding sub-model, a second training encoding sub-model, a third training encoding sub-model, a fourth training encoding sub-model, a fifth training encoding sub-model, a sixth training encoding sub-model, a seventh training encoding sub-model, an eighth training encoding sub-model, a ninth training encoding sub-model, a tenth training encoding sub-model, an eleventh training encoding sub-model, a twelfth training encoding sub-model, a first training decoding sub-model, a second training decoding sub-model, a sixth convolution module, a linear normalization module, and a first training activation module; The second training model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a training image perception parsing sub-model, a seventh training fusion normalization module, a fully convolutional network, and a second training activation module; The step of updating the parameters of the initialized target image recognition model based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model includes: The first training model is trained using the target sample set to obtain the target first training model; The model parameters corresponding to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh training coding sub-models in the first training model of the target image recognition model are respectively transferred to the first, second, third, fourth, fifth, sixth, eighth, ninth, tenth, and eleventh target coding sub-models of the initialized target image recognition model and the second training model, to obtain the first updated target image recognition model and the updated second training model; The model parameters of the first target encoding sub-model, the second target encoding sub-model, the third target encoding sub-model, the fourth target encoding sub-model, the fifth target encoding sub-model, the sixth target encoding sub-model, the seventh target encoding sub-model, the eighth target encoding sub-model, the ninth target encoding sub-model, the tenth target encoding sub-model, the eleventh target encoding sub-model, the first sampling module, the second sampling module, the third sampling module, the fourth sampling module, the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module in the updated second training model are locked, and the updated second training model is trained using the target sample set to obtain the target second training model; The model parameters corresponding to the training image perception parsing sub-model in the second training model of the target are transferred to the target image perception parsing sub-model of the first updated target image recognition model to obtain the updated target image recognition model.

[0010] As an optional implementation, in the first aspect of the present invention, the first training decoding sub-model includes a second training multi-head attention module, a fourth training fusion normalization module, a third training multi-head attention module, a fifth training fusion normalization module, a second training neural network, and a sixth training fusion normalization module. The input of the first training encoding sub-model is connected to the input of the first training model; the output of the first training encoding sub-model is connected to the input of the second training encoding sub-model; the output of the second training encoding sub-model is connected to the input of the third training encoding sub-model; the output of the third training encoding sub-model is connected to the input of the fourth training encoding sub-model; the output of the fourth training encoding sub-model is connected to the input of the fifth training encoding sub-model; the output of the training encoding sub-model is connected to the input of the sixth training encoding sub-model; the output of the sixth training encoding sub-model is connected to the input of the seventh training encoding sub-model; the output of the seventh training encoding sub-model is connected to the input of the eighth training encoding sub-model; the output of the eighth training encoding sub-model is connected to the input of the ninth training encoding sub-model; the output of the ninth training encoding sub-model is connected to the input of the tenth training encoding sub-model; the output of the tenth training encoding sub-model is connected to the input of the eleventh training encoding sub-model; the output of the eleventh training encoding sub-model is connected to the input of the twelfth training encoding sub-model; the output of the twelfth training encoding sub-model is connected to... The second training multi-head attention module has its input terminals connected to the fourth training fusion normalization module and the third training multi-head attention module. The output terminal of the second training multi-head attention module is connected to the input terminal of the fourth training fusion normalization module. The output terminal of the fourth training fusion normalization module is connected to the input terminals of the third training multi-head attention module and the fifth training fusion normalization module. The output terminal of the third training multi-head attention module is connected to the input terminal of the fifth training fusion normalization module. The output terminal of the fifth training fusion normalization module is connected to the input terminals of the second training neural network and the sixth training fusion normalization module. The output terminal of the second training neural network is connected to the input terminal of the sixth training fusion normalization module. The output terminal of the sixth training fusion normalization module is connected to the input terminal of the second training decoding sub-model. The output terminal of the second training decoding sub-model is connected to the input terminal of the fifth convolution module. The output terminal of the fifth convolution module is connected to the input terminal of the linear normalization module. The output terminal of the linear normalization module is connected to the input terminal of the first training activation module. The output terminal of the first training activation module is connected to the output of the first training model.

[0011] A second aspect of this invention discloses an image processing apparatus for breast cancer identification, the apparatus comprising: The acquisition module is used to acquire information about the image to be recognized. The first processing module is used to preprocess the image information to be identified to obtain initial recognition image information; The second processing module is used to perform recognition processing on the initial recognition image information based on the target image recognition model to obtain the target image recognition result; the target image recognition result is used to indicate the analysis and recognition of breast cancer.

[0012] A third aspect of the present invention discloses another image processing apparatus for breast cancer identification, the apparatus comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the image processing method for breast cancer identification disclosed in the first aspect of the present invention.

[0013] The fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked, are used to perform some or all of the steps in the image processing method for breast cancer identification disclosed in the first aspect of the present invention.

[0014] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, image information to be identified is acquired; the image information to be identified is preprocessed to obtain initial image information for identification; the initial image information for identification is processed based on a target image recognition model to obtain a target image recognition result; the target image recognition result is used to indicate the analysis and identification of breast cancer. Therefore, this application is beneficial for improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic flowchart of an image processing method for breast cancer identification disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an image processing device for breast cancer identification disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of another image processing device for breast cancer identification disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a target image recognition model disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a first training model disclosed in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a second training model disclosed in an embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] This invention discloses an image processing method and apparatus for breast cancer identification, which improves image recognition accuracy and image processing efficiency, thereby enhancing the efficiency of breast cancer image analysis. These will be described in detail below.

[0021] Example 1 Please see Figure 1 , Figure 1 This is a schematic flowchart of an image processing method for breast cancer identification disclosed in an embodiment of the present invention. Figure 1 The described image processing method for breast cancer identification is applied in an image management system, such as a local server or cloud server for image processing management of breast cancer identification, etc., and the embodiments of the present invention are not limited thereto. Figure 1As shown, the image processing method for breast cancer identification may include the following operations: 101. Obtain the image information to be recognized.

[0022] 102. Preprocess the image information to be recognized to obtain the initial recognition image information.

[0023] 103. Based on the target image recognition model, the initial recognition image information is processed to obtain the target image recognition result.

[0024] It should be noted that the above-mentioned image information representing the breast region can be obtained by breast ultrasound, X-ray, breast MRI, etc., and the embodiments of the present invention are not limited thereto.

[0025] In this embodiment of the invention, the target image recognition result is used to indicate the analysis and identification of breast cancer. Furthermore, the above-mentioned target image recognition result is only used to assist medical personnel in their judgment and does not directly represent the detection result. This embodiment of the invention does not impose any limitations.

[0026] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0027] In an optional embodiment, the preprocessing of the image information to be recognized to obtain initial recognition image information includes: The image information to be recognized is sliced ​​to obtain sliced ​​image information; the sliced ​​image information includes M sliced ​​images; M is a positive even number. Dimensionality reduction processing is performed on the sliced ​​image information to obtain dimensionality-reduced image information; Position encoding is performed on the sliced ​​image information to obtain image position encoding information; The sliced ​​image information, the reduced image information, and the image location encoding information are integrated to obtain the initial recognition image information.

[0028] Preferably, M is an even number between 20 and 28. Furthermore, choosing a relatively large and even value for M is primarily to balance the computational load of the subsequent target image recognition model with the preservation of image information after slicing. If the number of slices is too small, each block will be very large, which, while containing more complete object context information, will significantly increase the computational burden on the model. If the number of slices is too large (i.e., cut too finely), each block will be very small, potentially causing a complete object (especially a small target) to be divided into multiple blocks. No single small block can provide enough features for the model to recognize, resulting in missed detections. An even number between 20 and 28 achieves a better balance between computational efficiency and the completeness of target recognition, and each slice contains richer spatial context information, helping the target image recognition model better understand the relationships between local features. Furthermore, choosing an even number ensures that the sliced ​​image blocks can fit the symmetrical structure of the target image recognition model, avoiding feature map misalignment or information loss due to size parity issues, and ensuring the accuracy of feature alignment. This embodiment of the invention does not impose limitations on this.

[0029] It should be noted that the aforementioned positional encoding provides positional information for the input image sequence, enabling the target image recognition model to capture the sequential information of the images in the input sequence. Furthermore, positional encoding assigns a feature vector representing the original position of each image patch, allowing the target image recognition model to remember the relative or absolute position of the image patch within the complete image. Without positional encoding, the model would treat the input as an unordered set and would be unable to correctly understand the structural relationships of objects in the image. This embodiment of the invention does not impose limitations on this.

[0030] It should be noted that the above-mentioned dimensionality reduction image information includes M dimensionality reduction image block information, which is not limited in the embodiments of the present invention.

[0031] In this optional embodiment, as an optional implementation, the above-described dimensionality reduction processing of the sliced ​​image information to obtain dimensionality-reduced image information includes: The pixel values ​​of the sliced ​​image information are normalized to obtain normalized image information; the normalized image information includes M normalized image block information. For any normalized image block information in the normalized image information, perform two convolution processes on the normalized image block information in sequence to obtain the first convolution image block information corresponding to the normalized image block information; Pooling is performed on the first convolutional image block information corresponding to the normalized image block information to obtain the first pooled image block information corresponding to the normalized image block information. The first pooling image block information corresponding to the normalized image block information is convolved to obtain the second convolutional image block information corresponding to the normalized image block information. Pooling is performed on the second convolutional image block information corresponding to the normalized image block information to obtain the second pooled image block information corresponding to the normalized image block information. Flatten the second pooled image block information corresponding to the normalized image block information to obtain the flattened image block information corresponding to the normalized image block information. Fully connected processing is performed on the flattened image block information corresponding to the normalized image block information to obtain the dimensionality-reduced image block information corresponding to the normalized image block information.

[0032] It should be noted that the number of convolution channels in the three convolution processes in the above dimensionality reduction process increases sequentially, such as 16, 32, and 64. Furthermore, the first convolution is to extract low-level features, the second convolution is to extract more complex features, and the third convolution is to extract deeper features. This embodiment of the invention does not limit the scope of the invention.

[0033] It should be noted that the two pooling operations in the above dimensionality reduction process are both performed using max pooling layers, with the functions of dimensionality reduction to retain the most salient features and encoding the output feature map, respectively. This embodiment of the invention does not impose any limitations on these processes.

[0034] It should be noted that the flattening process in the above-mentioned dimensionality reduction process flattens the feature map into a vector, and then uses the fully connected layer to obtain the core feature vector after dimensionality reduction, that is, the dimensionality reduction image patch information. This embodiment of the invention does not limit this.

[0035] It should be noted that the aforementioned dimensionality reduction process maps high-dimensional image data to a low-dimensional space. By preserving the linear relationships between data, the data is mapped from the original high-dimensional space to a low-dimensional space, enabling better understanding and processing of the data. Furthermore, directly processing high-resolution original images or large sets of slices using a target image recognition model would be computationally intensive. However, by performing dimensionality reduction processing on the sliced ​​image data to obtain dimensionality-reduced image information, not only are the most essential and highly discriminative features extracted, but the dimensionality of the data is also significantly reduced. This accelerates the model training and inference process, reduces computational resource consumption, and effectively filters out some noise and redundant information in the image, thereby improving the model's generalization ability and preventing overfitting. This embodiment of the invention is not limited to this.

[0036] It should be noted that the integration of sliced ​​image information (providing local details and content of the image), dimensionality-reduced image information (providing condensed and more representative features), and image location encoding information (providing the spatial structural relationships of the image) mentioned above is to integrate the information into a single data set in the order described above. Furthermore, the fusion and complementary information of sliced ​​information, dimensionality-reduced features, and location encoding can enhance feature representation capabilities. Specifically, simple slicing can disrupt the inherent spatial layout relationships of the image, while the integration of location encoding and sliced ​​information ensures that the model, while focusing on local features, does not lose overall spatial structural information. Furthermore, the concatenation and integration of the above three types of information is equivalent to constructing a multi-view image representation, which can generate a more comprehensive and robust feature representation than any single information source. This effectively balances multiple requirements such as information preservation, computational efficiency, feature enhancement, and model adaptation, and can improve the accuracy of breast feature recognition in noisy, occluded, or complex scenes. This embodiment of the invention does not impose limitations on this.

[0037] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0038] In another alternative embodiment, such as Figure 4 As shown, the target image recognition model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and a target image perception and parsing sub-model; wherein, The input of the first target encoding sub-model is connected to the input of the target image recognition model; the output of the first target encoding sub-model is connected to the input of the second target encoding sub-model; the output of the second target encoding sub-model is connected to the input of the third target encoding sub-model; the output of the third target encoding sub-model is connected to the input of the fourth target encoding sub-model and the input of the fourth sampling module; the output of the fourth target encoding sub-model is connected to the input of the fifth target encoding sub-model; the output of the target encoding sub-model is connected to the input of the sixth target encoding sub-model and the input of the third sampling module; the output of the sixth target encoding sub-model is connected to the input of the seventh target encoding sub-model; the output of the seventh target encoding sub-model is connected to the input of the eighth target encoding sub-model and the input of the second sampling module; the output of the eighth target encoding sub-model is connected to the input of the ninth target encoding sub-model; the output of the ninth target encoding sub-model is connected to the input of the tenth target encoding sub-model; the output of the tenth target encoding sub-model is connected to the input of the eleventh target encoding sub-model; and the output of the eleventh target encoding sub-model is connected to the first sampling module. The input terminals of the first, second, third, and fourth sampling modules are respectively connected to the input terminals of the first, second, third, and fourth convolution modules; the output terminals of the first, second, third, and fourth convolution modules are all connected to the input terminals of the target image perception and analysis sub-model; the output terminal of the target image perception and analysis sub-model is connected to the model output of the target image recognition model. The target image recognition model includes the first target encoding sub-model, the second target encoding sub-model, the third target encoding sub-model, the fourth target encoding sub-model, the fifth target encoding sub-model, the sixth target encoding sub-model, the seventh target encoding sub-model, the eighth target encoding sub-model, the ninth target encoding sub-model, the tenth target encoding sub-model, the eleventh target encoding sub-model, the first sampling module, the second sampling module, the third sampling module, the fourth sampling module, the first convolution module, the second convolution module, the third convolution module, the fourth convolution module, and the target image perception and analysis sub-model; wherein, The input of the first target encoding sub-model is connected to the model input of the target image recognition model; the output of the first target encoding sub-model is connected to the input of the second target encoding sub-model; the output of the second target encoding sub-model is connected to the input of the third target encoding sub-model; the output of the third target encoding sub-model is connected to the input of the fourth target encoding sub-model and the input of the fourth sampling module; the output of the fourth target encoding sub-model is connected to the input of the fifth target encoding sub-model; the output of the target encoding sub-model is connected to the input of the sixth target encoding sub-model and the input of the third sampling module; the output of the sixth target encoding sub-model is connected to the input of the seventh target encoding sub-model; the output of the seventh target encoding sub-model is connected to the input of the eighth target encoding sub-model and the input of the second sampling module; the output of the eighth target encoding sub-model is connected to... The input terminal of the ninth target encoding sub-model is connected to the input terminal of the tenth target encoding sub-model; the output terminal of the tenth target encoding sub-model is connected to the input terminal of the eleventh target encoding sub-model; the output terminal of the eleventh target encoding sub-model is connected to the input terminal of the first sampling module; the output terminals of the first, second, third, and fourth sampling modules are respectively connected to the input terminals of the first, second, third, and fourth convolution modules; the output terminals of the first, second, third, and fourth convolution modules are all connected to the input terminal of the target image perception and analysis sub-model; the output terminal of the target image perception and analysis sub-model is connected to the model output of the target image recognition model.

[0039] It should be noted that the first sampling module, the second sampling module, the third sampling module, and the fourth sampling module mentioned above are all upsampling modules.

[0040] It should be noted that the first, second, third, and fourth convolutional modules mentioned above are convolutional layers. Furthermore, through the first, second, third, and fourth sampling modules and the first, second, third, and fourth convolutional modules, feature size scaling of 4×, 2×, 1×, and 0.5× can be achieved. For example, the output size of the fifth target encoding sub-model is 400×768, where 400 is considered the feature size and 768 is considered the number of channels. After reshaping the feature to 20*20 and scaling the feature side length by 4×, the feature dimension becomes 80×80×768.

[0041] It should be noted that the model architecture of the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh target coding sub-models is consistent, and their output dimension is 400×768.

[0042] It should be noted that the core innovation of the above target image recognition model lies in the use of multi-scale feature extraction and progressive reconstruction for image recognition. Its core is to first perform multi-level and multi-scale feature abstraction on the input image, then fuse the features at different levels of abstraction, and finally make accurate predictions based on the rich information after fusion.

[0043] Furthermore, the first to eleventh target encoding sub-models can be regarded as deep encoders of the target image recognition model. They are the backbone of the target image recognition model's feature extraction. The input image first passes through a series of encoding sub-models to progressively extract image features. The shallow encoding sub-models (such as the first and second layers) focus more on local details. As the layers deepen and the receptive field increases, the deep encoding sub-models (such as the seventh and eleventh layers) begin to capture global and abstract information, thereby transforming the initial input recognition image information into a series of feature maps with rich semantics. This embodiment of the invention is not limited to these specific features.

[0044] Furthermore, the sampling module and convolution module can be viewed as multi-scale feature fusion units in the target image recognition model, integrating feature maps from different stages of the encoding process through skip connections. Although the feature maps obtained from the end of the deep encoder (after the eleventh layer) are rich in semantic information, they have low resolution and lack the details required for precise localization. Therefore, the target image recognition model of this application gradually increases the resolution of the deep feature maps through the first to fourth sampling modules. At the same time, through skip connections, the feature maps from the same period (corresponding resolution) in the encoder are passed over and concatenated or added with the upsampled feature maps. The following first to fourth convolution modules are used to efficiently fuse these features from different depths, combining abstract semantic information with precise spatial details.

[0045] Furthermore, the target image perception and parsing sub-model makes a final decision based on the previously fused multi-scale features that combine semantics and details; this embodiment of the invention is not limited in this respect.

[0046] In summary, the target image recognition model described above is based on multi-scale feature fusion. The model can utilize both local details and global contextual information simultaneously, making it more robust and accurate than a simple single-path model. The jump connection mechanism directly transmits high-resolution detail information from the encoder to the decoder path, effectively compensating for the spatial information loss caused by operations such as pooling, and ensuring that the boundaries of the identified objects are clear and their positions are accurate.

[0047] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0048] In yet another optional embodiment, the model architecture of the first target coding sub-model to the eleventh target coding sub-model is consistent; The first target encoding sub-model includes a first fusion normalization module, a multi-head attention module, a second fusion normalization module, a first neural network, and a third fusion normalization module; wherein, The input of the first fusion normalization module is connected to the model input of the target image recognition model. The output of the first fusion normalization module is connected to the input of the multi-head attention module and the input of the second fusion normalization module. The output of the multi-head attention module is connected to the input of the second fusion normalization module. The output of the second fusion normalization module is connected to the input of the first neural network and the input of the third fusion normalization module. The output of the third fusion normalization module is connected to the input of the second target encoding sub-model. The target image perception and parsing sub-model includes a first pooling module, a fourth fusion and normalization module, a fifth fusion and normalization module, a sixth fusion and normalization module, a feature fusion module, a fifth convolution module, and a classification module; among them, The input of the first pooling module is connected to the output of the fourth convolution module. The output of the first pooling module is connected to the input of the fourth fusion normalization module and the input of the feature fusion module. The input of the fourth fusion normalization module is also connected to the output of the third convolution module. The output of the fourth fusion normalization module is connected to the input of the fifth fusion normalization module and the input of the feature fusion module. The input of the fifth fusion normalization module is also connected to the output of the second convolution module. The output of the fifth fusion normalization module is connected to the input of the sixth fusion normalization module and the input of the feature fusion module. The input of the sixth fusion normalization module is also connected to the output of the first convolution module. The output of the sixth fusion normalization module is connected to the input of the feature fusion module. The output of the feature fusion module is connected to the input of the fifth convolution module. The output of the fifth convolution module is connected to the input of the classification module. The output of the classification module is connected to the model output of the target image recognition model.

[0049] It should be noted that the first fusion normalization module, the second fusion normalization module, the third fusion normalization module, the fourth fusion normalization module, the fifth fusion normalization module, the sixth fusion normalization module, and the feature fusion module are all Add & Norm layers.

[0050] It should be noted that the aforementioned multi-head attention module refers to the multi-head attention mechanism module.

[0051] It should be noted that the first neural network is a feedforward neural network.

[0052] It should be noted that the first pooling module is the global pooling layer.

[0053] It should be noted that the fifth convolutional module is a 3×3 convolutional layer.

[0054] It should be noted that the classification module is a classifier.

[0055] It should be noted that the first target encoding sub-model first fuses and normalizes the input raw signal with the output of the multi-head attention module, providing a stable starting point for subsequent complex attention calculations. Then, the stabilized features are fed into the multi-head attention module, allowing each image patch to be updated based on its relationship with all other patches, thereby capturing long-range dependencies between image patches. This is then fused and normalized again with the initial input through a second fusion and normalization module, and a first neural network performs a non-linear transformation to enhance feature representation. Finally, the result is stabilized again through a third fusion and normalization module before being output to the next encoding sub-model. Thus, the first target encoding sub-model can transform a sequence containing M independent image patches into a feature sequence where each unit contains global contextual information. This is more conducive to the progressive understanding of image content by the entire target image recognition model, ultimately achieving accurate recognition. This embodiment of the invention does not limit the scope of the invention.

[0056] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0057] In yet another alternative embodiment, the target image recognition model is obtained based on the following steps: Obtain the initial sample set and initialize the target image recognition model; The initial sample set is augmented to obtain the target sample set; The parameters of the initialized target image recognition model are updated based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model.

[0058] It should be noted that the loss function for training the above model can be the cross-entropy function, the number of model iteration training rounds can be no less than 50 rounds, and the training hardware is a GPU of 4050 or higher. This embodiment of the invention does not impose any limitations.

[0059] It should be noted that the initial sample set includes a training set, a test set, and a validation set. The training set contains 200 images, each with edges annotated by 5 people, using five slightly different edge labels. The test set also contains 200 images, each with edges annotated by 5 people, again using five slightly different edge labels. The validation set contains 100 images, with labels also provided by 5 people.

[0060] It should be noted that the sample augmentation of the initial sample set is performed using methods such as image enhancement and image segmentation.

[0061] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0062] In an optional embodiment, such as Figure 5 and Figure 6 As shown, the first training model includes a first training encoding sub-model, a second training encoding sub-model, a third training encoding sub-model, a fourth training encoding sub-model, a fifth training encoding sub-model, a sixth training encoding sub-model, a seventh training encoding sub-model, an eighth training encoding sub-model, a ninth training encoding sub-model, a tenth training encoding sub-model, an eleventh training encoding sub-model, a twelfth training encoding sub-model, a first training decoding sub-model, a second training decoding sub-model, a sixth convolution module, a linear normalization module, and a first training activation module. The second training model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a training image perception parsing sub-model, a seventh training fusion normalization module, a fully convolutional network, and a second training activation module. The parameters of the initialized target image recognition model are updated based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model, including: The first training model is trained using the target sample set to obtain the target first training model; The model parameters corresponding to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh training coding sub-models in the first training model of the target image recognition model are transferred to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh target coding sub-models of the initialized target image recognition model and the second training model, respectively, to obtain the first updated target image recognition model and the updated second training model; The model parameters of the first target encoding sub-model, the second target encoding sub-model, the third target encoding sub-model, the fourth target encoding sub-model, the fifth target encoding sub-model, the sixth target encoding sub-model, the seventh target encoding sub-model, the eighth target encoding sub-model, the ninth target encoding sub-model, the tenth target encoding sub-model, the eleventh target encoding sub-model, the first sampling module, the second sampling module, the third sampling module, the fourth sampling module, the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module in the updated second training model are locked, and the updated second training model is trained using the target sample set to obtain the target second training model; The model parameters corresponding to the training image perception parsing sub-model in the second training model of the target are transferred to the target image perception parsing sub-model of the first updated target image recognition model to obtain the updated target image recognition model.

[0063] It should be noted that the above training process is a phased knowledge transfer and modular training strategy, which combines the technical implementation logic of model parameter transfer, module parameter locking and phased training, so as to gradually transfer and optimize the knowledge and capabilities of a large and complex model (the first training model) to a more practical target model, thereby obtaining a target image recognition model with better performance. This embodiment of the invention does not limit the scope of the invention.

[0064] Furthermore, in the first stage, a pre-trained, high-performance complex model (the first training model) is used to inject powerful feature extraction capabilities into the target model and the second training model. First, the first training model is trained using the target sample set to obtain a high-performance target first training model. Then, its eleventh training encoding sub-model and all previous encoder parameters are completely copied into the corresponding encoding sub-models of the target image recognition model and the second training model. This greatly accelerates the subsequent training process and improves the initial performance of the model, avoiding the huge cost and risk of training the encoder from scratch. This embodiment of the invention is not limited to this approach.

[0065] Furthermore, in the second stage, without compromising existing feature extraction capabilities, a prediction head adept at specific downstream tasks is trained. The parameters of all encoding sub-models, sampling, and convolutional modules in the second training model are locked. Then, using the target sample set, only the remaining parts are trained: the image-aware parsing sub-model, the seventh training fusion normalization module, the fully convolutional network, and the second training activation module. This prevents overfitting and catastrophic forgetting, and by locking the encoder, forces the model to train only the trainable parts, such as the image-aware parsing sub-model, improving its focus on the specific task of breast region recognition.

[0066] Furthermore, in the third stage, the results of the first two stages—a powerful encoder and a refined decoder—are combined into the final target model. The parameters of the task-optimized training image perception parsing sub-model trained in the second stage are transferred to the target image perception parsing sub-model of the target image recognition model. Thus, the third stage ensures both the powerful general feature encoding capabilities inherited from the first training model and the specialized decoding capabilities optimized for the target sample set and breast region recognition task inherited from the second training model.

[0067] As can be seen, the model training logic of this application not only avoids reinventing the wheel and maximizes the value of the pre-trained model, but also effectively prevents overfitting and accelerates the convergence speed by using the parameter locking strategy; it can also focus on solving a specific sub-problem at each stage, reducing the overall training complexity. The embodiments of this invention are not limited.

[0068] It should be noted that the model architectures of the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, eleventh, and twelfth training coding sub-models are consistent.

[0069] It should be noted that the model architecture of the first and second trained decoding sub-models mentioned above is consistent.

[0070] It should be noted that the sixth convolutional module mentioned above is a 3×3 convolutional layer.

[0071] It should be noted that the above linear normalization module is a normalization layer built based on normalize.

[0072] It should be noted that both the first and second training activation modules mentioned above are constructed based on the sigmoid function.

[0073] It should be noted that the seventh training fusion normalization module is the Add & Norm layer.

[0074] It should be noted that fully convolutional networks are models based on convolutional neural networks.

[0075] It should be noted that the aforementioned training image perception parsing sub-model includes a first training pooling module, a fourth training fusion normalization module, a fifth training fusion normalization module, a sixth training fusion normalization module, a training feature fusion module, a fifth training convolution module, and a training classification module. Furthermore, the model architecture of each sub-module of the training image perception parsing sub-model corresponds to that of the target image perception parsing sub-model.

[0076] It should be noted that the input of the above-mentioned training image perception parsing sub-model is connected to the output of the first convolutional module, the second convolutional module, the third convolutional module, and the fourth convolutional module, respectively.

[0077] It should be noted that the input of the seventh training fusion normalization module is connected to the outputs of the first, second, third, and fourth convolutional modules, respectively. Furthermore, the output of the seventh training fusion normalization module is connected to the input of the fully convolutional network.

[0078] It should be noted that the output of the fully convolutional network and the output of the trained image perception parsing sub-model are both connected to the input of the second training activation module.

[0079] It should be noted that in the second training model, the input of the first target encoding sub-model is connected to the input of the second training model.

[0080] It should be noted that the output of the second training activation module is connected to the output of the second training model.

[0081] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0082] In another optional embodiment, the first training decoding sub-model includes a second training multi-head attention module, a fourth training fusion normalization module, a third training multi-head attention module, a fifth training fusion normalization module, a second training neural network, and a sixth training fusion normalization module. The input of the first training encoding sub-model is connected to the input of the first training model; the output of the first training encoding sub-model is connected to the input of the second training encoding sub-model; the output of the second training encoding sub-model is connected to the input of the third training encoding sub-model; the output of the third training encoding sub-model is connected to the input of the fourth training encoding sub-model; the output of the fourth training encoding sub-model is connected to the input of the fifth training encoding sub-model; the output of the training encoding sub-model is connected to the input of the sixth training encoding sub-model; the output of the sixth training encoding sub-model is connected to the input of the seventh training encoding sub-model; the output of the seventh training encoding sub-model is connected to the input of the eighth training encoding sub-model; the output of the eighth training encoding sub-model is connected to the input of the ninth training encoding sub-model; the output of the ninth training encoding sub-model is connected to the input of the tenth training encoding sub-model; the output of the tenth training encoding sub-model is connected to the input of the eleventh training encoding sub-model; the output of the eleventh training encoding sub-model is connected to the input of the twelfth training encoding sub-model; and the output of the twelfth training encoding sub-model is connected to the second training encoding sub-model. The input terminals of the training multi-head attention module, the fourth training fusion normalization module, and the third training multi-head attention module are connected. The output terminal of the second training multi-head attention module is connected to the input terminal of the fourth training fusion normalization module. The output terminal of the fourth training fusion normalization module is connected to the input terminals of the third and fifth training fusion normalization modules. The output terminal of the third training multi-head attention module is connected to the input terminal of the fifth training fusion normalization module. The output terminal of the fifth training fusion normalization module is connected to the input terminals of the second and sixth training neural networks. The output terminal of the second training neural network is connected to the input terminal of the sixth training fusion normalization module. The output terminal of the sixth training fusion normalization module is connected to the input terminal of the second training decoding sub-model. The output terminal of the second training decoding sub-model is connected to the input terminal of the fifth convolution module. The output terminal of the fifth convolution module is connected to the input terminal of the linear normalization module. The output terminal of the linear normalization module is connected to the input terminal of the first training activation module. The output terminal of the first training activation module is connected to the output terminal of the first training model.

[0083] It should be noted that the first training decoding sub-model mentioned above ensures that the generation of the output sequence is strictly from left to right and step by step through the mask self-attention mechanism, thus guaranteeing the execution order; through the encoder-decoder attention mechanism, it can fully and selectively utilize all the information of the input sequence and deeply fuse contextual feature information; the extensive use of residual connections and layer normalization enables this complex model, which may contain up to 12 encoder layers, to be effectively trained, avoiding common problems of deep networks such as gradient vanishing. This embodiment of the invention does not limit this.

[0084] It is evident that implementing the image processing method for breast cancer identification described in the embodiments of the present invention is beneficial to improving the accuracy of image identification and the efficiency of image processing, thereby improving the efficiency of breast cancer image analysis.

[0085] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an image processing device for breast cancer identification disclosed in an embodiment of the present invention. Figure 2 The described apparatus can be applied in image management systems, such as local servers or cloud servers for image processing management in breast cancer identification, etc., and the embodiments of the present invention are not limited thereto. Figure 2 As shown, the device may include: The acquisition module 201 is used to acquire information about the image to be recognized; The first processing module 202 is used to preprocess the image information to be recognized to obtain initial recognition image information; The second processing module 203 is used to process the initial recognition image information based on the target image recognition model to obtain the target image recognition result; the target image recognition result is used to indicate the analysis and recognition of breast cancer.

[0086] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0087] In another alternative embodiment, such as Figure 2 As shown, the first processing module 202 preprocesses the image information to be recognized to obtain initial recognition image information, including: The image information to be recognized is sliced ​​to obtain sliced ​​image information; the sliced ​​image information includes M sliced ​​images; M is a positive even number. Dimensionality reduction processing is performed on the sliced ​​image information to obtain dimensionality-reduced image information; Position encoding is performed on the sliced ​​image information to obtain image position encoding information; The sliced ​​image information, the reduced image information, and the image location encoding information are integrated to obtain the initial recognition image information.

[0088] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0089] In yet another alternative embodiment, such as Figure 2 As shown, the target image recognition model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and a target image perception and parsing sub-model; wherein, The input of the first target encoding sub-model is connected to the model input of the target image recognition model; the output of the first target encoding sub-model is connected to the input of the second target encoding sub-model; the output of the second target encoding sub-model is connected to the input of the third target encoding sub-model; the output of the third target encoding sub-model is connected to the input of the fourth target encoding sub-model and the input of the fourth sampling module; the output of the fourth target encoding sub-model is connected to the input of the fifth target encoding sub-model; the output of the target encoding sub-model is connected to the input of the sixth target encoding sub-model and the input of the third sampling module; the output of the sixth target encoding sub-model is connected to the input of the seventh target encoding sub-model; the output of the seventh target encoding sub-model is connected to the input of the eighth target encoding sub-model and the input of the second sampling module; the output of the eighth target encoding sub-model is connected to... The input terminal of the ninth target encoding sub-model is connected to the input terminal of the tenth target encoding sub-model; the output terminal of the tenth target encoding sub-model is connected to the input terminal of the eleventh target encoding sub-model; the output terminal of the eleventh target encoding sub-model is connected to the input terminal of the first sampling module; the output terminals of the first, second, third, and fourth sampling modules are respectively connected to the input terminals of the first, second, third, and fourth convolution modules; the output terminals of the first, second, third, and fourth convolution modules are all connected to the input terminal of the target image perception and analysis sub-model; the output terminal of the target image perception and analysis sub-model is connected to the model output of the target image recognition model.

[0090] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0091] In yet another alternative embodiment, such as Figure 2 As shown, the model architectures of the first target encoding sub-model to the eleventh target encoding sub-model are consistent; The first target encoding sub-model includes a first fusion normalization module, a multi-head attention module, a second fusion normalization module, a first neural network, and a third fusion normalization module; wherein, The input of the first fusion normalization module is connected to the model input of the target image recognition model. The output of the first fusion normalization module is connected to the input of the multi-head attention module and the input of the second fusion normalization module. The output of the multi-head attention module is connected to the input of the second fusion normalization module. The output of the second fusion normalization module is connected to the input of the first neural network and the input of the third fusion normalization module. The output of the third fusion normalization module is connected to the input of the second target encoding sub-model. The target image perception and parsing sub-model includes a first pooling module, a fourth fusion and normalization module, a fifth fusion and normalization module, a sixth fusion and normalization module, a feature fusion module, a fifth convolution module, and a classification module; among them, The input of the first pooling module is connected to the output of the fourth convolution module. The output of the first pooling module is connected to the input of the fourth fusion normalization module and the input of the feature fusion module. The input of the fourth fusion normalization module is also connected to the output of the third convolution module. The output of the fourth fusion normalization module is connected to the input of the fifth fusion normalization module and the input of the feature fusion module. The input of the fifth fusion normalization module is also connected to the output of the second convolution module. The output of the fifth fusion normalization module is connected to the input of the sixth fusion normalization module and the input of the feature fusion module. The input of the sixth fusion normalization module is also connected to the output of the first convolution module. The output of the sixth fusion normalization module is connected to the input of the feature fusion module. The output of the feature fusion module is connected to the input of the fifth convolution module. The output of the fifth convolution module is connected to the input of the classification module. The output of the classification module is connected to the model output of the target image recognition model.

[0092] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0093] In yet another alternative embodiment, such as Figure 2As shown, the target image recognition model is obtained by performing the following steps based on the second processing module 203: Obtain the initial sample set and initialize the target image recognition model; The initial sample set is augmented to obtain the target sample set; The parameters of the initialized target image recognition model are updated based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model.

[0094] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0095] In yet another alternative embodiment, such as Figure 2 As shown, the first training model includes a first training encoding sub-model, a second training encoding sub-model, a third training encoding sub-model, a fourth training encoding sub-model, a fifth training encoding sub-model, a sixth training encoding sub-model, a seventh training encoding sub-model, an eighth training encoding sub-model, a ninth training encoding sub-model, a tenth training encoding sub-model, an eleventh training encoding sub-model, a twelfth training encoding sub-model, a first training decoding sub-model, a second training decoding sub-model, a sixth convolutional module, a linear normalization module, and a first training activation module; The second training model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a training image perception parsing sub-model, a seventh training fusion normalization module, a fully convolutional network, and a second training activation module. The second processing module 203 updates the parameters of the initialized target image recognition model based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model, including: The first training model is trained using the target sample set to obtain the target first training model; The model parameters corresponding to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh training coding sub-models in the first training model of the target image recognition model are transferred to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh target coding sub-models of the initialized target image recognition model and the second training model, respectively, to obtain the first updated target image recognition model and the updated second training model; The model parameters of the first target encoding sub-model, the second target encoding sub-model, the third target encoding sub-model, the fourth target encoding sub-model, the fifth target encoding sub-model, the sixth target encoding sub-model, the seventh target encoding sub-model, the eighth target encoding sub-model, the ninth target encoding sub-model, the tenth target encoding sub-model, the eleventh target encoding sub-model, the first sampling module, the second sampling module, the third sampling module, the fourth sampling module, the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module in the updated second training model are locked, and the updated second training model is trained using the target sample set to obtain the target second training model; The model parameters corresponding to the training image perception parsing sub-model in the second training model of the target are transferred to the target image perception parsing sub-model of the first updated target image recognition model to obtain the updated target image recognition model.

[0096] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0097] In yet another alternative embodiment, such as Figure 2 As shown, the first training decoding sub-model includes a second training multi-head attention module, a fourth training fusion normalization module, a third training multi-head attention module, a fifth training fusion normalization module, a second training neural network, and a sixth training fusion normalization module. The input of the first training encoding sub-model is connected to the input of the first training model; the output of the first training encoding sub-model is connected to the input of the second training encoding sub-model; the output of the second training encoding sub-model is connected to the input of the third training encoding sub-model; the output of the third training encoding sub-model is connected to the input of the fourth training encoding sub-model; the output of the fourth training encoding sub-model is connected to the input of the fifth training encoding sub-model; the output of the training encoding sub-model is connected to the input of the sixth training encoding sub-model; the output of the sixth training encoding sub-model is connected to the input of the seventh training encoding sub-model; the output of the seventh training encoding sub-model is connected to the input of the eighth training encoding sub-model; the output of the eighth training encoding sub-model is connected to the input of the ninth training encoding sub-model; the output of the ninth training encoding sub-model is connected to the input of the tenth training encoding sub-model; the output of the tenth training encoding sub-model is connected to the input of the eleventh training encoding sub-model; the output of the eleventh training encoding sub-model is connected to the input of the twelfth training encoding sub-model; and the output of the twelfth training encoding sub-model is connected to the second training encoding sub-model. The input terminals of the training multi-head attention module, the fourth training fusion normalization module, and the third training multi-head attention module are connected. The output terminal of the second training multi-head attention module is connected to the input terminal of the fourth training fusion normalization module. The output terminal of the fourth training fusion normalization module is connected to the input terminals of the third and fifth training fusion normalization modules. The output terminal of the third training multi-head attention module is connected to the input terminal of the fifth training fusion normalization module. The output terminal of the fifth training fusion normalization module is connected to the input terminals of the second and sixth training neural networks. The output terminal of the second training neural network is connected to the input terminal of the sixth training fusion normalization module. The output terminal of the sixth training fusion normalization module is connected to the input terminal of the second training decoding sub-model. The output terminal of the second training decoding sub-model is connected to the input terminal of the fifth convolution module. The output terminal of the fifth convolution module is connected to the input terminal of the linear normalization module. The output terminal of the linear normalization module is connected to the input terminal of the first training activation module. The output terminal of the first training activation module is connected to the output terminal of the first training model.

[0098] It is evident that implementation Figure 2 The described image processing device for breast cancer identification is beneficial to improving image recognition accuracy and image processing efficiency, thereby improving the efficiency of breast cancer image analysis.

[0099] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of another image processing device for breast cancer identification disclosed in an embodiment of the present invention. Figure 3The described apparatus can be applied in image management systems, such as local servers or cloud servers for image processing management in breast cancer identification, etc., and the embodiments of the present invention are not limited thereto. Figure 3 As shown, the device may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps in the image processing method for breast cancer identification described in Embodiment 1.

[0100] Example 4 This invention discloses a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform the steps in the image processing method for breast cancer identification described in Embodiment 1.

[0101] Example 5 This invention discloses a computer program product comprising a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform the steps in the image processing method for breast cancer identification described in Embodiment 1.

[0102] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0103] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0104] Finally, it should be noted that the image processing method and apparatus for breast cancer identification disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image processing method for breast cancer identification, characterized in that, The method includes: Obtain the image information to be recognized; The image information to be identified is preprocessed to obtain initial recognition image information; The initial image information is processed based on the target image recognition model to obtain the target image recognition result; the target image recognition result is used to indicate the analysis and recognition of breast cancer.

2. The image processing method for breast cancer identification according to claim 1, characterized in that, The preprocessing of the image information to be identified to obtain initial recognition image information includes: The image information to be identified is sliced ​​to obtain sliced ​​image information; the sliced ​​image information includes M sliced ​​images; M is a positive even number; The sliced ​​image information is subjected to dimensionality reduction processing to obtain dimensionality-reduced image information; The sliced ​​image information is subjected to position encoding processing to obtain image position encoding information; The sliced ​​image information, the reduced-dimensional image information, and the image location encoding information are integrated to obtain the initial recognition image information.

3. The image processing method for breast cancer identification according to claim 1, characterized in that, The target image recognition model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and a target image perception and parsing sub-model; wherein, The input of the first target encoding sub-model is connected to the model input of the target image recognition model, and the output of the first target encoding sub-model is connected to the input of the second target encoding sub-model; the output of the second target encoding sub-model is connected to the input of the third target encoding sub-model; the output of the third target encoding sub-model is connected to the input of the fourth target encoding sub-model and the input of the fourth sampling module; the output of the fourth target encoding sub-model is connected to the input of the fifth target encoding sub-model; the output of the target encoding sub-model is connected to the input of the sixth target encoding sub-model and the input of the third sampling module; the output of the sixth target encoding sub-model is connected to the input of the seventh target encoding sub-model; the output of the seventh target encoding sub-model is connected to the input of the eighth target encoding sub-model and the input of the second sampling module; the output of the eighth target encoding sub-model is connected to the... The input terminal of the ninth target encoding sub-model is connected to the input terminal of the tenth target encoding sub-model; the output terminal of the tenth target encoding sub-model is connected to the input terminal of the eleventh target encoding sub-model; the output terminal of the eleventh target encoding sub-model is connected to the input terminal of the first sampling module; the output terminals of the first, second, third, and fourth sampling modules are respectively connected to the input terminals of the first, second, third, and fourth convolution modules; the output terminals of the first, second, third, and fourth convolution modules are all connected to the input terminal of the target image perception and analysis sub-model; the output terminal of the target image perception and analysis sub-model is connected to the model output of the target image recognition model.

4. The image processing method for breast cancer identification according to claim 3, characterized in that, The model architectures of the first target encoding sub-model to the eleventh target encoding sub-model are consistent; The first target encoding sub-model includes a first fusion normalization module, a multi-head attention module, a second fusion normalization module, a first neural network, and a third fusion normalization module; wherein, The input of the first fusion normalization module is connected to the model input of the target image recognition model; the output of the first fusion normalization module is connected to the input of the multi-head attention module and the input of the second fusion normalization module; the output of the multi-head attention module is connected to the input of the second fusion normalization module; the output of the second fusion normalization module is connected to the input of the first neural network and the input of the third fusion normalization module; the output of the third fusion normalization module is connected to the input of the second target encoding sub-model. The target image perception and parsing sub-model includes a first pooling module, a fourth fusion normalization module, a fifth fusion normalization module, a sixth fusion normalization module, a feature fusion module, a fifth convolution module, and a classification module; wherein, The input of the first pooling module is connected to the output of the fourth convolution module. The output of the first pooling module is connected to the input of the fourth fusion normalization module and the input of the feature fusion module. The input of the fourth fusion normalization module is also connected to the output of the third convolution module. The output of the fourth fusion normalization module is also connected to the input of the fifth fusion normalization module and the input of the feature fusion module. The input of the fifth fusion normalization module is also connected to the output of the second convolution module. The output of the fifth fusion normalization module is also connected to the input of the sixth fusion normalization module and the input of the feature fusion module. The input of the sixth fusion normalization module is also connected to the output of the first convolution module. The output of the sixth fusion normalization module is connected to the input of the feature fusion module. The output of the feature fusion module is connected to the input of the fifth convolution module. The output of the fifth convolution module is connected to the input of the classification module. The output of the classification module is connected to the model output of the target image recognition model.

5. The image processing method for breast cancer identification according to claim 3, characterized in that, The target image recognition model is obtained based on the following steps: Obtain the initial sample set and initialize the target image recognition model; The initial sample set is augmented to obtain the target sample set; The parameters of the initialized target image recognition model are updated based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model.

6. The image processing method for breast cancer identification according to claim 5, characterized in that, The first training model includes a first training encoding sub-model, a second training encoding sub-model, a third training encoding sub-model, a fourth training encoding sub-model, a fifth training encoding sub-model, a sixth training encoding sub-model, a seventh training encoding sub-model, an eighth training encoding sub-model, a ninth training encoding sub-model, a tenth training encoding sub-model, an eleventh training encoding sub-model, a twelfth training encoding sub-model, a first training decoding sub-model, a second training decoding sub-model, a sixth convolution module, a linear normalization module, and a first training activation module; The second training model includes a first target encoding sub-model, a second target encoding sub-model, a third target encoding sub-model, a fourth target encoding sub-model, a fifth target encoding sub-model, a sixth target encoding sub-model, a seventh target encoding sub-model, an eighth target encoding sub-model, a ninth target encoding sub-model, a tenth target encoding sub-model, an eleventh target encoding sub-model, a first sampling module, a second sampling module, a third sampling module, a fourth sampling module, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a training image perception parsing sub-model, a seventh training fusion normalization module, a fully convolutional network, and a second training activation module; The step of updating the parameters of the initialized target image recognition model based on the target sample set, the first training model, and the second training model to obtain the updated target image recognition model includes: The first training model is trained using the target sample set to obtain the target first training model; The model parameters corresponding to the first, second, third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, and eleventh training coding sub-models in the first training model of the target image recognition model are respectively transferred to the first, second, third, fourth, fifth, sixth, eighth, ninth, tenth, and eleventh target coding sub-models of the initialized target image recognition model and the second training model, to obtain the first updated target image recognition model and the updated second training model; The model parameters of the first target encoding sub-model, the second target encoding sub-model, the third target encoding sub-model, the fourth target encoding sub-model, the fifth target encoding sub-model, the sixth target encoding sub-model, the seventh target encoding sub-model, the eighth target encoding sub-model, the ninth target encoding sub-model, the tenth target encoding sub-model, the eleventh target encoding sub-model, the first sampling module, the second sampling module, the third sampling module, the fourth sampling module, the first convolution module, the second convolution module, the third convolution module, and the fourth convolution module in the updated second training model are locked, and the updated second training model is trained using the target sample set to obtain the target second training model; The model parameters corresponding to the training image perception parsing sub-model in the second training model of the target are transferred to the target image perception parsing sub-model of the first updated target image recognition model to obtain the updated target image recognition model.

7. The image processing method for breast cancer identification according to claim 6, characterized in that, The first training decoding sub-model includes a second training multi-head attention module, a fourth training fusion normalization module, a third training multi-head attention module, a fifth training fusion normalization module, a second training neural network, and a sixth training fusion normalization module; The input of the first training encoding sub-model is connected to the input of the first training model; the output of the first training encoding sub-model is connected to the input of the second training encoding sub-model; the output of the second training encoding sub-model is connected to the input of the third training encoding sub-model; the output of the third training encoding sub-model is connected to the input of the fourth training encoding sub-model; the output of the fourth training encoding sub-model is connected to the input of the fifth training encoding sub-model; the output of the training encoding sub-model is connected to the input of the sixth training encoding sub-model; the output of the sixth training encoding sub-model is connected to the input of the seventh training encoding sub-model; the output of the seventh training encoding sub-model is connected to the input of the eighth training encoding sub-model; the output of the eighth training encoding sub-model is connected to the input of the ninth training encoding sub-model; the output of the ninth training encoding sub-model is connected to the input of the tenth training encoding sub-model; the output of the tenth training encoding sub-model is connected to the input of the eleventh training encoding sub-model; the output of the eleventh training encoding sub-model is connected to the input of the twelfth training encoding sub-model; the output of the twelfth training encoding sub-model is connected to... The second training multi-head attention module has its input terminals connected to the fourth training fusion normalization module and the third training multi-head attention module. The output terminal of the second training multi-head attention module is connected to the input terminal of the fourth training fusion normalization module. The output terminal of the fourth training fusion normalization module is connected to the input terminals of the third training multi-head attention module and the fifth training fusion normalization module. The output terminal of the third training multi-head attention module is connected to the input terminal of the fifth training fusion normalization module. The output terminal of the fifth training fusion normalization module is connected to the input terminals of the second training neural network and the sixth training fusion normalization module. The output terminal of the second training neural network is connected to the input terminal of the sixth training fusion normalization module. The output terminal of the sixth training fusion normalization module is connected to the input terminal of the second training decoding sub-model. The output terminal of the second training decoding sub-model is connected to the input terminal of the fifth convolution module. The output terminal of the fifth convolution module is connected to the input terminal of the linear normalization module. The output terminal of the linear normalization module is connected to the input terminal of the first training activation module. The output terminal of the first training activation module is connected to the output of the first training model.

8. An image processing device for breast cancer identification, characterized in that, The device includes: The acquisition module is used to acquire information about the image to be recognized. The first processing module is used to preprocess the image information to be identified to obtain initial recognition image information; The second processing module is used to perform recognition processing on the initial recognition image information based on the target image recognition model to obtain the target image recognition result; the target image recognition result is used to indicate the analysis and recognition of breast cancer.

9. An image processing device for breast cancer identification, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the image processing method for breast cancer identification as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when invoked, are used to perform the image processing method for breast cancer identification as described in any one of claims 1-7.

Images