Auxiliary identification method and device for peritoneal false mucus tumor image and computer equipment

Abstract

The invention relates to an auxiliary identification method and device for a peritoneal false mucus tumor image and computer equipment. The method comprises the following steps: acquiring a scanning image of a focus area, and extracting a focus feature map through a focus feature extraction network based on the scanning image; on the basis of the focus feature map, generating a feature mark sequence map, and on the basis of the feature mark sequence map, extracting a semantic feature map through a hierarchical down-sampling encoder; and based on the semantic feature map and the feature marking sequence map, generating a focus mask identification map of the focus area through a hierarchical up-sampling decoder. By adopting the method, the focus guiding accuracy of the peritoneal false mucus tumor image can be improved.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of medical artificial intelligence and computer vision, and in particular to an auxiliary recognition method, device and computer equipment for peritoneal pseudomyxoma images. Background Technology

[0002] Pseudomyxoma peritonei (PMP) is a rare malignant tumor of the peritoneum characterized by massive accumulation of mucinous ascites and widespread dissemination of intraperitoneal tumor implants. Currently, contrast-enhanced computed tomography (CECT) is the primary imaging modality for preoperative diagnosis and evaluation of PMP. However, manual interpretation of traditional CT images has limitations such as significant subjectivity and time-consuming reading, and in areas with complex anatomical structures such as the mesentery, perihepatic spaces, and subphrenic fossa, lesions are prone to being missed or underestimated. Therefore, improving the accuracy of auxiliary identification of PMP is a current research focus.

[0003] Existing technical solutions use convolutional neural networks (CNNs) based on deep learning technology to automatically and objectively delineate lesions. However, CNNs can only receive local information, making it difficult for traditional CNNs to effectively associate long-range spatial dependencies and capture global contextual information. The computational complexity is too large in high-resolution medical image processing, affecting the fine texture details required for precise boundary localization, thus resulting in low accuracy in lesion guidance for peritoneal pseudomyxoma images. Summary of the Invention

[0004] Therefore, it is necessary to provide an auxiliary identification method, device, and computer equipment for peritoneal pseudomyxoma images to address the aforementioned technical problems.

[0005] In a first aspect, this application provides an auxiliary method for identifying peritoneal pseudomyxoma images, comprising:

[0006] Acquire scanned images of the lesion area, and based on the scanned images, extract lesion feature maps using a lesion feature extraction network;

[0007] Based on the lesion feature map, a feature marker sequence map is generated, and based on the feature marker sequence map, a semantic feature map is extracted through a hierarchical downsampling encoder;

[0008] Based on the semantic feature map and the feature label sequence map, a lesion mask identification map of the lesion region is generated through a hierarchical upsampling decoder.

[0009] Optionally, the step of extracting lesion feature maps based on the scanned image using a lesion feature extraction network includes:

[0010] Based on the scanned images, lesion features are extracted using a lesion feature extraction network;

[0011] Identify the image range of the lesion features and the image content corresponding to the lesion features, and use the image range of the lesion features and the image content corresponding to the lesion features as each lesion feature map.

[0012] Optionally, generating a feature marker sequence map based on the lesion feature map includes:

[0013] The lesion feature map is projected to generate a query matrix and a key matrix, and the attention weight distribution information of the query matrix is ​​calculated based on a preset learnable parameter vector.

[0014] Based on the attention weight distribution information of the query matrix, a global attention query vector is calculated, and based on the global attention query vector and the key matrix, a feature label sequence map of the global context containing the lesion feature map is generated.

[0015] Optionally, the step of extracting semantic feature maps based on the feature-labeled sequence map using a hierarchical downsampling encoder includes:

[0016] The feature marker sequence map is downsampled using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. The local representation features of the downsampled feature map are then extracted using the hybrid processing module of the hierarchical downsampling encoder.

[0017] If the preset first iteration condition is not met, the downsampled feature map is used as the feature label sequence map, and the process is returned to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map, until the first preset iteration condition is met.

[0018] The downsampled feature map obtained in the last iteration, and the local representation features obtained in each iteration, are used as the semantic feature map.

[0019] Optionally, the step of generating a lesion mask map of the lesion region based on the semantic feature map and the feature label sequence map through a hierarchical upsampling decoder includes:

[0020] The semantic feature map is upsampled by the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and the feature layer level of the upsampled feature map is identified.

[0021] Based on the feature layer level of the upsampled feature map, in each downsampled feature map and the feature label sequence map, the target feature map corresponding to the upsampled feature map is selected, and based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder;

[0022] If the preset second iteration condition is not met, the new upsampled image is used as the semantic feature image, and the process is returned to execute the upsampling module through the hierarchical upsampling decoder to perform upsampling operation on the semantic feature image to obtain the upsampled feature image corresponding to the semantic feature image, until the second preset iteration condition is met.

[0023] The newly upsampled image obtained from the last iteration is used as the lesion mask identifier for the lesion region.

[0024] Optionally, the step of generating a new upsampled map based on the upsampled feature map and the target feature map through the alignment module of the hierarchical upsampled decoder includes:

[0025] The upsampled feature map is projected into a first self-attention matrix, and the target feature map is projected into a second self-attention matrix;

[0026] Based on the first self-attention matrix and the second self-attention matrix, a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix are generated through a dual cross-attention strategy.

[0027] Based on the first feature concatenation matrix and the second feature concatenation matrix, the upsampled feature map and the target feature map are subjected to image fusion processing to obtain a new upsampled map.

[0028] Secondly, this application also provides an auxiliary recognition device for peritoneal pseudomyxoma images, comprising:

[0029] The acquisition module is used to acquire scanned images of the lesion area and, based on the scanned images, extract lesion feature maps through a lesion feature extraction network.

[0030] The hierarchical upsampling module is used to generate a feature marker sequence map based on the lesion feature map, and extract a semantic feature map based on the feature marker sequence map through a hierarchical downsampling encoder;

[0031] The hierarchical downsampling module is used to generate a lesion mask map of the lesion region based on the semantic feature map and the feature label sequence map through a hierarchical upsampling decoder.

[0032] Optionally, the acquisition module is specifically used for:

[0033] Based on the scanned images, lesion features are extracted using a lesion feature extraction network;

[0034] Identify the image range of the lesion features and the image content corresponding to the lesion features, and use the image range of the lesion features and the image content corresponding to the lesion features as each lesion feature map.

[0035] Optionally, the hierarchical downsampling module is specifically used for:

[0036] The lesion feature map is projected to generate a query matrix and a key matrix, and the attention weight distribution information of the query matrix is ​​calculated based on a preset learnable parameter vector.

[0037] Based on the attention weight distribution information of the query matrix, a global attention query vector is calculated, and based on the global attention query vector and the key matrix, a feature label sequence map of the global context containing the lesion feature map is generated.

[0038] Optionally, the hierarchical downsampling module is specifically used for:

[0039] The feature marker sequence map is downsampled using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. The local representation features of the downsampled feature map are then extracted using the hybrid processing module of the hierarchical downsampling encoder.

[0040] If the preset first iteration condition is not met, the downsampled feature map is used as the feature label sequence map, and the process is returned to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map, until the first preset iteration condition is met.

[0041] The downsampled feature map obtained in the last iteration, and the local representation features obtained in each iteration, are used as the semantic feature map.

[0042] Optionally, the hierarchical upsampling module is specifically used for:

[0043] The semantic feature map is upsampled by the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and the feature layer level of the upsampled feature map is identified.

[0044] Based on the feature layer level of the upsampled feature map, in each downsampled feature map and the feature label sequence map, the target feature map corresponding to the upsampled feature map is selected, and based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder;

[0045] If the preset second iteration condition is not met, the new upsampled image is used as the semantic feature image, and the process is returned to execute the upsampling module through the hierarchical upsampling decoder to perform upsampling operation on the semantic feature image to obtain the upsampled feature image corresponding to the semantic feature image, until the second preset iteration condition is met.

[0046] The newly upsampled image obtained from the last iteration is used as the lesion mask identifier for the lesion region.

[0047] Optionally, the hierarchical upsampling module is specifically used for:

[0048] The upsampled feature map is projected into a first self-attention matrix, and the target feature map is projected into a second self-attention matrix;

[0049] Based on the first self-attention matrix and the second self-attention matrix, a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix are generated through a dual cross-attention strategy.

[0050] Based on the first feature concatenation matrix and the second feature concatenation matrix, the upsampled feature map and the target feature map are subjected to image fusion processing to obtain a new upsampled map.

[0051] Thirdly, this application provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described in any one of the first aspects.

[0052] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0053] Fifthly, this application provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0054] The aforementioned auxiliary identification method, apparatus, and computer equipment for peritoneal pseudomyxoma images acquire scanned images of the lesion region and, based on the scanned images, extract lesion feature maps using a lesion feature extraction network; generate feature marker sequence maps based on the lesion feature maps, and extract semantic feature maps based on the feature marker sequence maps using a hierarchical downsampling encoder; and generate lesion mask identification maps of the lesion region using a hierarchical upsampling decoder based on the semantic feature maps and the feature marker sequence maps. This solution, through a hierarchical downsampling encoder designed by the inventors, can extract local features from the lesion feature map in a context-rich scan image in a hierarchical manner, thereby obtaining a semantic feature map. Then, through the hierarchical downsampling encoder designed by the inventors, a high-resolution segmentation map is reconstructed from the semantic feature map of abstract multi-scale features step by step. The high-resolution images in the downsampling extraction process are aligned and fused with the reconstructed high-resolution segmentation map, thereby ensuring that both deep semantic context and fine spatial details can be represented in the lesion mask identification map. Thus, while ensuring the extraction of local features of the image, it effectively improves the representation of global context information of the scan and fine texture details, thereby comprehensively improving the accuracy of lesion guidance for peritoneal pseudomyxoma images. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating an auxiliary identification method for peritoneal pseudomyxoma images in one embodiment;

[0057] Figure 2 This is a module architecture diagram of the downsampling module of a hierarchical downsampling encoder in one embodiment;

[0058] Figure 3 This is a schematic diagram of the upsampling and alignment process in one embodiment;

[0059] Figure 4 This is a flowchart illustrating an example of auxiliary identification of peritoneal pseudomyxoma images in one embodiment;

[0060] Figure 5 This is a structural block diagram of an auxiliary recognition device for peritoneal pseudomyxoma images in one embodiment;

[0061] Figure 6This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0064] The auxiliary identification method for peritoneal pseudomyxoma images provided in this application embodiment can be applied to a system for auxiliary identification of peritoneal pseudomyxoma images. This system can be applied to a terminal, which can be, but is not limited to, various personal computers, laptops, mid-range computers, etc. Specifically, the terminal, through a hierarchical downsampling encoder designed by the inventors, can extract local features from the lesion feature map in a context-rich scan image in a hierarchical manner, thereby obtaining a semantic feature map. Then, through the hierarchical downsampling encoder designed by the inventors, a high-resolution segmentation map is reconstructed from the abstract multi-scale feature semantic feature map step by step. The high-resolution images obtained during the downsampling extraction process are aligned and fused with the reconstructed high-resolution segmentation map, thereby ensuring that both deep semantic context and fine spatial details can be represented in the lesion mask identification map. This effectively improves the representation of global contextual information and fine texture details of the scan while ensuring the extraction of local features of the image, thus comprehensively improving the accuracy of lesion guidance in peritoneal pseudomyxoma images.

[0065] In one exemplary embodiment, such as Figure 1 As shown, an auxiliary recognition method for peritoneal pseudomyxoma images is provided. Taking the application of this method to a terminal as an example, the method includes the following steps S101 to S103. Wherein:

[0066] Step S101: Obtain scanned images of the lesion area, and extract lesion feature maps based on the scanned images using a lesion feature extraction network.

[0067] In this embodiment, the terminal receives scan data of the lesion region acquired by a Contrast-Enhanced Computed Tomography (CECT) device, and obtains a scanned image of the lesion region. This lesion region can be the trunk region where a pseudomyxoma of the peritoneum is located, or any other type of lesion located in the trunk region. Then, the terminal uses a convolutional neural network (i.e., a lesion feature extraction network) based on deep learning technology to mark the feature map of the lesion region in the scanned image, obtaining a lesion feature map. This marking method can be any image marking method, and this solution is not limited to any particular method. The specific extraction process will be described in detail later.

[0068] Step S102: Based on the lesion feature map, generate a feature label sequence map, and based on the feature label sequence map, extract a semantic feature map through a hierarchical downsampling encoder.

[0069] In this embodiment, the terminal generates a feature-labeled sequence map based on the lesion feature map, and extracts a semantic feature map based on the feature-labeled sequence map through a hierarchical downsampling encoder. The generation of the feature-labeled sequence map is achieved by using a hybrid encoder based on additive visual attention, designed in this scheme, to transform the lesion feature map into a feature-labeled sequence map. This transformation method extracts global contextual information from the lesion feature map through a self-attention mechanism, thereby outputting a feature sequence map that ensures stable and efficient propagation of global information. The hierarchical structure of this hierarchical downsampling encoder constructs a rich multi-scale representation system, which can effectively handle the heterogeneous features of different PMP manifestations, ranging from small peritoneal implantations to large-area mucus accumulation. Furthermore, the hybrid characteristics of each stage ensure that the convolutional encoder can capture the fine texture features of mucus deposition, while the efficient attention module integrates this information into a global spatial context. The linear complexity of the attention mechanism allows this powerful dual-representation learning method to be applied at each stage, ensuring both the consistency and robustness of feature learning while significantly reducing computational costs. The specific process of generating the semantic feature map will be described in detail later.

[0070] Step S103: Based on the semantic feature map and the feature label sequence map, a lesion mask label map of the lesion area is generated through a hierarchical upsampling decoder.

[0071] In this embodiment, the terminal generates a lesion mask map of the lesion region based on the semantic feature map and the feature label sequence map, through a hierarchical upsampling decoder. The hierarchical upsampling encoder is a multi-level decoder with cross-attention. This decoder architecture adopts a symmetrical design, mirroring the hierarchical structure of the encoder. Its main goal is to progressively reconstruct a high-resolution segmentation map from abstract multi-scale features. The core efficiency of this hierarchical upsampling encoder lies in its Multiscale Dual-Representation Alignment Filter (MDAF) module, which implements fine-grained fusion between the encoder and decoder feature maps at each level. The specific fusion process will be described in detail later.

[0072] Based on the above scheme, the hierarchical downsampling encoder designed by the inventors can extract local features from the lesion feature map in the context-rich scanned image in a hierarchical manner to obtain a semantic feature map. Then, the hierarchical downsampling encoder designed by the inventors reconstructs a high-resolution segmentation map from the semantic feature map of abstract multi-scale features step by step. The high-resolution images in the downsampling extraction process are aligned and fused with the reconstructed high-resolution segmentation map, thereby ensuring that both deep semantic context and fine spatial details can be represented in the lesion mask map. This effectively improves the representation of global context information and fine texture details of the scan while ensuring the extraction of local features of the image, thus comprehensively improving the accuracy of lesion guidance for peritoneal pseudomyxoma images.

[0073] Optionally, based on the scanned image, a lesion feature map is extracted using a lesion feature extraction network, including: extracting lesion features based on the scanned image using a lesion feature extraction network; identifying the image range of the lesion features and the image content corresponding to the lesion features, and using the image range of the lesion features and the image content corresponding to the lesion features as each lesion feature map.

[0074] In this embodiment, the terminal extracts lesion features based on the scanned image using a lesion feature extraction network. Then, the terminal adapts the image content corresponding to the lesion feature in the scanned image and uses the range of this image content within the scanned image as the image range of the lesion feature. Finally, the terminal performs image labeling processing on the scanned image, combining the image range of the lesion feature and the corresponding image content, to obtain a lesion feature map. This image labeling processing involves marking the range and content of the lesion feature within the scanned image; this labeling can be achieved using various methods such as color and symbols.

[0075] Based on the above scheme, a lesion feature extraction network is first used to mark lesion feature maps in the scanned image, which facilitates the subsequent lesion image recognition by combining lesion features for screening, thereby improving the recognition efficiency of the lesion feature range of the scanned image of the lesion area.

[0076] Optionally, based on the lesion feature map, a feature label sequence map is generated, including: projecting the lesion feature map to generate a query matrix and a key matrix, and calculating the attention weight distribution information of the query matrix based on a preset learnable parameter vector; calculating a global attention query vector based on the attention weight distribution information of the query matrix, and generating a feature label sequence map containing the global context of the lesion feature map based on the global attention query vector and the key matrix.

[0077] In this embodiment, the terminal projects the lesion feature map to generate a query matrix and a key matrix, and calculates the attention weight distribution information of the query matrix based on a preset learnable parameter vector. The lesion feature map can be characterized as having a C-channel... . Then, the terminal projects the lesion feature map onto... , and In the matrix, generate Matrix (i.e., query matrix) and Matrix (i.e., key matrix).

[0078] The calculation formula corresponding to the above projection method is the same as the projection formula for attention technology, and the specific algorithm for this formula is as follows:

[0079]

[0080] In the above formula, Q is the query matrix, K is the key matrix, and V is the value matrix. These are preset parameters.

[0081] Then, the terminal calculates the global attention query vector based on the attention weight distribution information of the query matrix. This attention weight distribution information includes each query in the query matrix. The attention weights. The formula for calculating the global attention query vector is:

[0082]

[0083] In the above formula, Let Q be the global attention query vector, and let Q be the query matrix.

[0084] Let C be the learnable parameter vector, and C be the number of channels in the lesion feature map.

[0085] Then, based on the global attention query vector and the key matrix, the terminal generates a feature-labeled sequence map of the global context, containing lesion feature maps. Specifically, the terminal now transforms the global attention query vector into a global query representation, and the calculation formula corresponding to this transformation is as follows:

[0086]

[0087] In the above formula, q represents the global query. To query each vector point in the global attention query vector, For each query in the query matrix , where i is a virtual number.

[0088] Then, the terminal calculates each query through element-wise multiplication (⊙) between the global query representation and the key matrix. Each key in In relation to the relevance of the global query representation, the terminal ultimately assigns each of the above keywords... The relevance of the global query representation is returned to the normalized query matrix to obtain a feature-labeled sequence graph, thus ensuring the stable and efficient propagation of global information. The calculation formula for the above execution process is as follows:

[0089]

[0090] In the above formula, X out For feature-labeled sequence maps, For querying the matrix, For the prior transformation formula, For broadcast element-wise multiplication, The key matrix, This represents a global query.

[0091] Based on the above scheme, by converting the lesion feature map into a feature marker sequence map, the stable and efficient propagation of global information in the lesion feature map is increased.

[0092] Optionally, based on the feature label sequence map, a semantic feature map is extracted using a hierarchical downsampling encoder, including: downsampling the feature label sequence map using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map, and extracting local representation features of the downsampled feature map using the hybrid processing module of the hierarchical downsampling encoder; if the preset first iteration condition is not met, the downsampled feature map is used as the feature label sequence map, and the step of downsampling the feature label sequence map using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map is returned, until the first preset iteration condition is met; the downsampled feature map obtained in the last iteration and the local representation features obtained in each iteration are used as the semantic feature map.

[0093] In this embodiment, the terminal downsamples the feature marker sequence map using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. Then, the terminal extracts local representation features from the downsampled feature map using the hybrid processing module of the hierarchical downsampling encoder. This hybrid processing module employs a convolutional encoder structure—containing a 3×3 depthwise convolutional layer, followed by two layers with… Pointwise convolutional layers with activation functions. Used to learn local representation features of the downsampled feature map. The hierarchical architecture diagram of this downsampling module is shown below. Figure 2 As shown, a hierarchical backbone network architecture is adopted, consisting of four consecutive stages. The design goal of each stage is to capture features with progressively decreasing spatial resolution. After the initial block embedding step, the features are processed sequentially through each stage.

[0094] If the preset first iteration condition is not met, the terminal returns the downsampled feature map as a feature label sequence map to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map. This process continues until the first preset iteration condition is met. The first iteration condition is the number of iterations, specifically four consecutive iterations, corresponding to the four consecutive stages of the hierarchical downsampling encoder. Each downsampling process yields a downsampled feature map at a different level. After each iteration, the terminal transmits the resulting downsampled feature map to the efficient overlay attention module, thereby associating it with the global context information of the feature label sequence map.

[0095] The calculation formula for this hierarchical downsampling encoder is:

[0096]

[0097] In the above formula, For downsampled feature maps, This is a sequence map of feature labels.

[0098] The calculation formula for this overlay attention module is:

[0099]

[0100] In the above formula, To overlay downsampled feature maps containing global context information, This is a downsampled feature map.

[0101] Finally, the terminal uses the downsampled feature map obtained from the last iteration and the local representation features obtained from each iteration as the semantic feature map.

[0102] Based on the above scheme, by combining the efficient convolution module with the advanced additive attention mechanism, it is possible to achieve high-fidelity local feature extraction and effectively capture the long-range spatial dependence characteristics unique to diffuse PMP lesions while maintaining computational efficiency, thus laying a solid foundation for multi-scale feature representation.

[0103] Optionally, based on the semantic feature map and the feature marker sequence map, a lesion mask identification map of the lesion region is generated through a hierarchical upsampling decoder, including: performing an upsampling operation on the semantic feature map through the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and identifying the feature layer level of the upsampled feature map; based on the feature layer level of the upsampled feature map, selecting the target feature map corresponding to the upsampled feature map from each downsampled feature map and the feature marker sequence map, and generating a new upsampled map through the alignment module of the hierarchical upsampling decoder based on the upsampled feature map and the target feature map; if the preset second iteration condition is not met, using the new upsampled map as the semantic feature map, returning to execute the step of performing an upsampling operation on the semantic feature map through the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, until the second preset iteration condition is met; using the new upsampled map obtained in the last iteration as the lesion mask identification map of the lesion region.

[0104] In this embodiment, the terminal performs an upsampling operation on the semantic feature map through the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and identifies the feature layer level of the upsampled feature map. This feature layer level is the level corresponding to the upsampled feature map, and this level corresponds one-to-one with the level of the downsampled feature map. The level of the upsampled feature map in the last iteration is the same as the level corresponding to the feature label sequence map.

[0105] Then, based on the feature layer level of the upsampled feature map, the terminal filters the target feature map corresponding to the upsampled feature map from each downsampled feature map and the feature label sequence map. Based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder. This alignment module is a multi-scale dual representation alignment filter (MDAF) module. The specific alignment process will be explained in detail later. Figure 3 The diagram shows the processing flow for upsampling and alignment.

[0106] If the preset second iteration condition is not met, the terminal uses the newly upsampled image as the semantic feature map and returns it to the upsampling module of the hierarchical upsampling decoder to perform an upsampling operation on the semantic feature map, obtaining the upsampled feature map corresponding to the semantic feature map, until the second preset iteration condition is met. This second iteration condition is the number of iterations, which is the same as the number of iterations corresponding to the hierarchical upsampling encoder, both being four iterations. Finally, since the resolution of the newly upsampled image generated after four iterations is only one-quarter of the original input, the terminal performs another upsampling process on the newly upsampled image obtained in the last iteration to obtain a full-resolution binary segmentation mask. Finally, the terminal uses this full-resolution binary segmentation mask as the lesion mask identifier map for the lesion region.

[0107] Based on the above scheme, the core advantage of the hierarchical, attention-driven decoding strategy designed in this scheme effectively resolves the inherent contradiction between semantic abstraction and spatial accuracy. The MDAF module, through the rich contextual information of the deep learning network, effectively guides the localization and optimization of shallow features, achieving accurate presentation of local details and thus accurately and reliably delineating PMP lesions.

[0108] Optionally, based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder, including: projecting the upsampled feature map into a first self-attention matrix and projecting the target feature map into a second self-attention matrix; generating a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix through a dual cross-attention strategy based on the first and second self-attention matrices; and performing image fusion processing on the upsampled feature map and the target feature map based on the first and second feature concatenation matrices to obtain the new upsampled map.

[0109] In this embodiment, the terminal projects the upsampled feature map into a first self-attention matrix and the target feature map into a second self-attention matrix. The projection method is as follows: The multi-scale mapping module projects the upsampled feature map and the target feature map into a query matrix, a key matrix, and a value matrix, respectively. This multi-scale mapping module employs parallel strip convolutional structures with different kernel sizes, enabling it to capture multi-scale features and significantly improve representation capabilities. The first self-attention matrix includes a first query matrix, a first key matrix, and a first value matrix; the second self-attention matrix includes a second query matrix, a second key matrix, and a second value matrix.

[0110] The calculation formulas for the first self-attention matrix and the second self-attention matrix are as follows:

[0111]

[0112] In the above formula, , , These are the first query matrix, the first key matrix, and the first value matrix of the upsampled feature map, respectively. , , These are the second query matrix, the second key matrix, and the second value matrix of the target feature map, respectively. For upsampled feature maps, and For target feature map.

[0113] Then, based on the first self-attention matrix and the second self-attention matrix, the terminal generates a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix using a dual cross-attention strategy. This dual-attention strategy is a feature map alignment strategy for bidirectional semantic queries using a dual cross-attention mechanism. The terminal generates the first feature concatenation matrix and the second feature concatenation matrix by exchanging the query matrix in the first self-attention matrix with the key matrix and value matrix in the second self-attention matrix. Then, based on the first and second feature concatenation matrices, the terminal performs image fusion processing on the upsampled feature map and the target feature map to obtain a new upsampled image. The calculation formulas for this new upsampled image are as follows:

[0114]

[0115] Among them, two feature maps This is the first fused feature map obtained by fusing the target feature map into the upsampled map, and This is the second fused feature map after fusing the upsampled image into the target feature map. Finally, the terminal inputs the first and second fused feature maps into the convolutional projection layer and outputs a new upsampled image. This new upsampled image can be characterized as... .

[0116] Based on the above scheme, a precise alignment module is used to ensure that deep semantic context and fine spatial details can be comprehensively represented, thus solving the contradiction between global context association, local detail fidelity and computational efficiency.

[0117] This application also provides an example of auxiliary identification of peritoneal pseudomyxoma images, such as... Figure 4 As shown, the specific processing procedure includes the following steps:

[0118] Step S401: Obtain a scan image of the lesion area.

[0119] Step S402: Based on the scanned image, extract lesion features using a lesion feature extraction network.

[0120] Step S403: Identify the image range of the lesion features and the image content corresponding to the lesion features, and use the image range of the lesion features and the image content corresponding to the lesion features as the feature maps of each lesion.

[0121] Step S404: Project the lesion feature map to generate a query matrix and a key matrix, and calculate the attention weight distribution information of the query matrix based on the preset learnable parameter vector.

[0122] Step S405: Based on the attention weight distribution information of the query matrix, calculate the global attention query vector, and based on the global attention query vector and the key matrix, generate a feature label sequence map of the global context containing lesion feature maps.

[0123] Step S406: The feature marker sequence map is downsampled using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. The local representation features of the downsampled feature map are then extracted using the hybrid processing module of the hierarchical downsampling encoder.

[0124] Step S407: If the preset first iteration condition is not met, the downsampled feature map is used as a feature label sequence map and returned to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map, until the first preset iteration condition is met.

[0125] Step S408: The downsampled feature map obtained in the last iteration and the local representation features obtained in each iteration are used as semantic feature maps.

[0126] Step S409: The semantic feature map is upsampled by the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and the feature layer level of the upsampled feature map is identified.

[0127] Step S410: Based on the feature layer level of the upsampled feature map, select the target feature map corresponding to the upsampled feature map in each downsampled feature map and the feature label sequence map, and generate a new upsampled map based on the upsampled feature map and the target feature map through the alignment module of the hierarchical upsampled decoder.

[0128] Step S411: If the preset second iteration condition is not met, the new upsampled image is used as the semantic feature map, and the upsampling module of the hierarchical upsampling decoder is returned to perform upsampling operation on the semantic feature map to obtain the upsampled feature map corresponding to the semantic feature map, until the second preset iteration condition is met.

[0129] Step S412: Use the new upsampled image obtained in the last iteration as the lesion mask identifier image of the lesion area.

[0130] Step S413: Project the upsampled feature map into a first self-attention matrix, and project the target feature map into a second self-attention matrix.

[0131] Step S414: Based on the first self-attention matrix and the second self-attention matrix, a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix are generated through a dual cross-attention strategy.

[0132] Step S415: Based on the first feature concatenation matrix and the second feature concatenation matrix, perform image fusion processing on the upsampled feature map and the target feature map to obtain a new upsampled map.

[0133] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0134] Based on the same inventive concept, this application also provides an auxiliary identification device for peritoneal pseudomyxoma images to implement the auxiliary identification method for peritoneal pseudomyxoma images described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more auxiliary identification device embodiments for peritoneal pseudomyxoma images provided below can be found in the limitations of the auxiliary identification method for peritoneal pseudomyxoma images described above, and will not be repeated here.

[0135] In one exemplary embodiment, such as Figure 5 As shown, an auxiliary recognition device for peritoneal pseudomyxoma images is provided, comprising: an acquisition module 510, a layered upsampling module 520, and a layered downsampling module 530, wherein:

[0136] The acquisition module 510 is used to acquire scanned images of the lesion area and, based on the scanned images, extract lesion feature maps through a lesion feature extraction network.

[0137] The hierarchical upsampling module 520 is used to generate a feature marker sequence map based on the lesion feature map, and extract a semantic feature map based on the feature marker sequence map through a hierarchical downsampling encoder.

[0138] The hierarchical downsampling module 530 is used to generate a lesion mask identification map of the lesion region based on the semantic feature map and the feature label sequence map through the hierarchical upsampling decoder.

[0139] Optionally, the acquisition module 510 is specifically used for:

[0140] Based on the scanned images, lesion features are extracted using a lesion feature extraction network;

[0141] Identify the image range of the lesion features and the image content corresponding to the lesion features, and use the image range of the lesion features and the image content corresponding to the lesion features as each lesion feature map.

[0142] Optionally, the hierarchical downsampling module 520 is specifically used for:

[0143] The lesion feature map is projected to generate a query matrix and a key matrix, and the attention weight distribution information of the query matrix is ​​calculated based on a preset learnable parameter vector.

[0144] Based on the attention weight distribution information of the query matrix, a global attention query vector is calculated, and based on the global attention query vector and the key matrix, a feature label sequence map of the global context containing the lesion feature map is generated.

[0145] Optionally, the hierarchical downsampling module 520 is specifically used for:

[0146] The feature marker sequence map is downsampled using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. The local representation features of the downsampled feature map are then extracted using the hybrid processing module of the hierarchical downsampling encoder.

[0147] If the preset first iteration condition is not met, the downsampled feature map is used as the feature label sequence map, and the process is returned to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map, until the first preset iteration condition is met.

[0148] The downsampled feature map obtained in the last iteration, and the local representation features obtained in each iteration, are used as the semantic feature map.

[0149] Optionally, the hierarchical upsampling module 530 is specifically used for:

[0150] The semantic feature map is upsampled by the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and the feature layer level of the upsampled feature map is identified.

[0151] Based on the feature layer level of the upsampled feature map, in each downsampled feature map and the feature label sequence map, the target feature map corresponding to the upsampled feature map is selected, and based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder;

[0152] If the preset second iteration condition is not met, the new upsampled image is used as the semantic feature image, and the process is returned to execute the upsampling module through the hierarchical upsampling decoder to perform upsampling operation on the semantic feature image to obtain the upsampled feature image corresponding to the semantic feature image, until the second preset iteration condition is met.

[0153] The newly upsampled image obtained from the last iteration is used as the lesion mask identifier for the lesion region.

[0154] Optionally, the hierarchical upsampling module 530 is specifically used for:

[0155] The upsampled feature map is projected into a first self-attention matrix, and the target feature map is projected into a second self-attention matrix;

[0156] Based on the first self-attention matrix and the second self-attention matrix, a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix are generated through a dual cross-attention strategy.

[0157] Based on the first feature concatenation matrix and the second feature concatenation matrix, the upsampled feature map and the target feature map are subjected to image fusion processing to obtain a new upsampled map.

[0158] The modules in the aforementioned auxiliary recognition device for peritoneal pseudomyxoma images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0159] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements an auxiliary identification method for peritoneal pseudomyxoma images. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0160] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0161] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a beer warehouse inventory optimization method.

[0162] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a beer warehouse inventory optimization method.

[0163] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of a beer warehouse inventory optimization method.

[0164] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0165] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0166] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0167] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for auxiliary identification of peritoneal pseudomyxoma images, characterized in that, The method includes: Acquire scanned images of the lesion area, and based on the scanned images, extract lesion feature maps using a lesion feature extraction network; Based on the lesion feature map, a feature marker sequence map is generated, and based on the feature marker sequence map, a semantic feature map is extracted through a hierarchical downsampling encoder; Based on the semantic feature map and the feature label sequence map, a lesion mask identification map of the lesion region is generated through a hierarchical upsampling decoder.

2. The method according to claim 1, characterized in that, The step of extracting lesion feature maps based on the scanned image using a lesion feature extraction network includes: Based on the scanned images, lesion features are extracted using a lesion feature extraction network; Identify the image range of the lesion features and the image content corresponding to the lesion features, and use the image range of the lesion features and the image content corresponding to the lesion features as each lesion feature map.

3. The method according to claim 2, characterized in that, The step of generating a feature marker sequence map based on the lesion feature map includes: The lesion feature map is projected to generate a query matrix and a key matrix, and the attention weight distribution information of the query matrix is ​​calculated based on a preset learnable parameter vector. Based on the attention weight distribution information of the query matrix, a global attention query vector is calculated, and based on the global attention query vector and the key matrix, a feature label sequence map of the global context containing the lesion feature map is generated.

4. The method according to claim 1, characterized in that, The step of extracting semantic feature maps based on the feature-labeled sequence map using a hierarchical downsampling encoder includes: The feature marker sequence map is downsampled using the downsampling module of the hierarchical downsampling encoder to obtain a downsampled feature map. The local representation features of the downsampled feature map are then extracted using the hybrid processing module of the hierarchical downsampling encoder. If the preset first iteration condition is not met, the downsampled feature map is used as the feature label sequence map, and the process is returned to the downsampling module of the hierarchical downsampling encoder to perform downsampling processing on the feature label sequence map to obtain the downsampled feature map, until the first preset iteration condition is met. The downsampled feature map obtained in the last iteration, and the local representation features obtained in each iteration, are used as the semantic feature map.

5. The method according to claim 4, characterized in that, The step of generating a lesion mask map of the lesion region based on the semantic feature map and the feature label sequence map, through a hierarchical upsampling decoder, includes: The semantic feature map is upsampled by the upsampling module of the hierarchical upsampling decoder to obtain the upsampled feature map corresponding to the semantic feature map, and the feature layer level of the upsampled feature map is identified. Based on the feature layer level of the upsampled feature map, in each downsampled feature map and the feature label sequence map, the target feature map corresponding to the upsampled feature map is selected, and based on the upsampled feature map and the target feature map, a new upsampled map is generated through the alignment module of the hierarchical upsampled decoder; If the preset second iteration condition is not met, the new upsampled image is used as the semantic feature image, and the process is returned to execute the upsampling module through the hierarchical upsampling decoder to perform upsampling operation on the semantic feature image to obtain the upsampled feature image corresponding to the semantic feature image, until the second preset iteration condition is met. The newly upsampled image obtained from the last iteration is used as the lesion mask identifier for the lesion region.

6. The method according to claim 5, characterized in that, The step of generating a new upsampled map based on the upsampled feature map and the target feature map through the alignment module of the hierarchical upsampled decoder includes: The upsampled feature map is projected into a first self-attention matrix, and the target feature map is projected into a second self-attention matrix; Based on the first self-attention matrix and the second self-attention matrix, a first feature concatenation matrix corresponding to the first self-attention matrix and a second feature concatenation matrix corresponding to the second self-attention matrix are generated through a dual cross-attention strategy. Based on the first feature concatenation matrix and the second feature concatenation matrix, the upsampled feature map and the target feature map are subjected to image fusion processing to obtain a new upsampled map.

7. An auxiliary recognition device for peritoneal pseudomyxoma images, characterized in that, The device includes: The acquisition module is used to acquire scanned images of the lesion area and, based on the scanned images, extract lesion feature maps through a lesion feature extraction network. The hierarchical downsampling module is used to generate a feature marker sequence map based on the lesion feature map, and extract a semantic feature map based on the feature marker sequence map through the hierarchical downsampling encoder; The hierarchical upsampling module is used to generate a lesion mask map of the lesion region based on the semantic feature map and the feature label sequence map through the hierarchical upsampling decoder.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images