Neural image analysis method and system based on deep learning

Abstract

The invention provides a neural image analysis method and system based on deep learning. The method comprises the following steps: receiving neural image original data; carrying out multi-modal feature collaborative extraction on the original data of the neural image to generate an initial feature map; inputting the initial characteristic spectrum into a deep characteristic fusion network to carry out enhanced recognition on the lesion area to obtain a lesion area characteristic spectrum; and performing three-dimensional topological structure analysis on the lesion area characteristic spectrum, and constructing a lesion area three-dimensional structure model. According to the method, the spatial form, the volume and the surface topological relation of the lesion area can be accurately represented, a concrete structural basis is provided for surgical path planning, and the reliability of surgical path planning is improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning, and more specifically, to a deep learning-based method and system for neural image analysis. Background Technology

[0002] With the development of medical imaging technology, neuroimaging analysis, through the processing and analysis of neuroimaging data such as MRI and computed tomography (CT) scans, provides crucial information support for the diagnosis and treatment of brain lesions, playing a vital role in lesion area identification and surgical path planning. Currently, the common approach is to extract features independently from single-modality neuroimaging data, then identify lesion areas using traditional image processing algorithms or basic machine learning models, and finally construct a three-dimensional structure based on the identification results using a simple layer stacking method for surgical path planning. However, this approach of processing only single-modality data independently makes it difficult to fully utilize the complementary information of different modalities of imaging data, easily leading to incomplete feature representation. Furthermore, traditional algorithms have limited accuracy when the differences between complex lesion areas and normal tissue features are small. The three-dimensional structure constructed based on simple layer stacking cannot accurately reflect the spatial topological relationship of the lesion area, thus affecting the accuracy and reliability of surgical path planning. Summary of the Invention

[0003] This invention provides a method and system for neuroimaging analysis based on deep learning.

[0004] In a first aspect, embodiments of the present invention provide a deep learning-based neuroimage analysis method, comprising: Receive raw neuroimaging data; Multimodal feature co-extraction is performed on the raw neuroimaging data to generate an initial feature map; The initial feature map is input into a deep feature fusion network to enhance and identify the lesion region, thereby obtaining the lesion region feature map. A three-dimensional topological structure analysis is performed on the feature map of the lesion area to construct a three-dimensional structural model of the lesion area.

[0005] In a second aspect, embodiments of the present invention provide a computer system, comprising: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the deep learning-based neuroimaging analysis method described above.

[0006] The deep learning-based neuroimaging analysis method provided by this invention generates an initial feature map by performing multimodal feature collaborative extraction processing on the raw neuroimaging data. This integrates complementary features from different modalities of neuroimaging data, avoiding the problem of one-sided feature representation caused by insufficient feature information from a single modality. The initial feature map is then input into a deep feature fusion network for lesion region enhancement and recognition processing to obtain a lesion region feature map. This enhances the feature differences between lesion regions and normal tissues through multi-scale feature fusion and attention mechanisms, improving the accuracy of lesion region recognition. Finally, a three-dimensional topological structure analysis processing is performed on the lesion region feature map to construct a three-dimensional structural model of the lesion region. This model accurately represents the spatial morphology, volume, and surface topological relationships of the lesion region, providing a concrete structural basis for surgical path planning and improving the reliability of surgical path planning. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0008] Figure 1 This is a flowchart of a deep learning-based neuroimage analysis method provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0011] Please see Figure 1 , Figure 1 A flowchart of a deep learning-based neuroimaging analysis method provided in this embodiment of the invention, which can be executed by a computer system, includes the following steps S100~S400: Step S100: Receive raw neuroimaging data.

[0012] Raw neuroimaging data refers to various imaging data about the nervous system acquired through medical imaging equipment, encompassing multiple imaging modalities. For example, magnetic resonance imaging (MRI) data uses the principle of nuclear magnetic resonance to image the human nervous system, clearly displaying the anatomical structure and physiological information of nerve tissue. Computed tomography (CT) imaging data uses X-rays to perform a tomographic scan of a specific part of the body, which is then processed by a computer to reconstruct cross-sectional images of the body's interior, providing information such as nerve tissue density. Receiving raw neuroimaging data is the starting point of the entire analysis process. Through data interfaces or transmission protocols, the data acquired by medical imaging equipment is transmitted and stored in the analysis system. For example, in a hospital's radiology department, MRI and CT scanners transmit the acquired image data to a data storage server via the hospital's internal network. The analysis system then receives this raw neuroimaging data from the server, providing the foundational data for subsequent feature extraction and analysis.

[0013] Step S200: Perform multimodal feature co-extraction on the raw neuroimaging data to generate an initial feature map.

[0014] Multimodal feature collaborative extraction refers to the comprehensive consideration of the correlations and interactions between different imaging modalities in raw neuroimaging data to extract key features that reflect the characteristics of the nervous system. Different imaging modalities provide different information about the nervous system; by collaboratively extracting this information, the characteristics of the nervous system can be described more comprehensively and accurately. The initial feature map is a collection of multiple features extracted from the raw neuroimaging data, integrating and representing features from different modalities.

[0015] As one implementation method, step S200 involves multimodal feature co-extraction of the raw neuroimaging data to generate an initial feature map, which can be specifically implemented as follows: steps S210~S270: Step S210: Perform spatial resolution standardization on the magnetic resonance imaging data in the original neuroimaging data. By resampling the images, the magnetic resonance images under different acquisition conditions are unified to a preset spatial size to obtain standardized magnetic resonance images.

[0016] Spatial resolution standardization aims to eliminate differences in spatial resolution between magnetic resonance images acquired under different conditions, ensuring that images from different sources have a uniform spatial size and resolution. Image resampling is the process of achieving spatial resolution standardization. It involves interpolating or rearranging pixels in the original magnetic resonance image to adjust its spatial resolution to a preset size. The preset spatial size is a uniform spatial resolution pre-defined based on the analytical requirements and standards; for example, it can be set to a fixed pixel size and image dimensions. In practice, the original magnetic resonance images may have varying spatial resolutions due to different acquisition parameters, with some images having larger pixels and others smaller. Image resampling techniques, such as bilinear interpolation, recalculate the grayscale value of each pixel, unifying the image to the preset spatial size, thus obtaining a standardized magnetic resonance image. This ensures that subsequent analyses are performed at a uniform spatial scale, improving the accuracy and comparability of the analytical results.

[0017] Step S220: Perform grayscale characteristic normalization on the computed tomography image data in the original neuroimaging data, and map the grayscale distribution of the tomography image to a preset distribution range through dynamic range compression to obtain normalized tomography image.

[0018] Gray-scale normalization aims to eliminate differences in gray-scale distribution between computed tomography (CT) images acquired under different conditions, ensuring that CT images from different sources have similar gray-scale distribution characteristics. Dynamic range compression is the process of achieving gray-scale normalization. By adjusting the gray-scale values ​​of the CT image, the wide dynamic range of the original image is compressed into a preset distribution interval. The preset distribution interval is a range of gray-scale values ​​pre-defined according to the analysis requirements and standards, for example, [0, 255]. In practice, due to different scanning parameters and individual patient differences, the gray-scale distribution of CT images can vary significantly. Dynamic range compression techniques, such as linear mapping algorithms, linearly transform the gray-scale values ​​of the original image, mapping them to the preset distribution interval, thus obtaining a normalized CT image. This makes different CT images comparable in grayscale, facilitating feature extraction and analysis.

[0019] Step S230: Perform intermodal spatial registration between the standardized magnetic resonance image and the normalized tomographic image, optimize the spatial transformation relationship between the two modal images based on mutual information, and generate the registered multimodal image.

[0020] Intermodal spatial registration aims to align images from different imaging modalities in space, ensuring accurate correspondence within the same coordinate system. This facilitates comprehensive analysis of the features of different modalities. Mutual information, a metric for measuring the correlation between two random variables, is used in image registration to measure the similarity between normalized magnetic resonance imaging (MRI) images and normalized computed tomography (CT) images. By optimizing the mutual information calculation, a spatial transformation relationship that maximizes the mutual information between two modalities can be found, enabling accurate spatial alignment. In practice, a spatial transformation model, such as an affine transformation model, is first defined to describe image transformations involving translation, rotation, and scaling. Then, iterative optimization algorithms, such as gradient descent, are used to continuously adjust the spatial transformation parameters to maximize the mutual information between the two modalities. Finally, based on the optimal spatial transformation parameters, the normalized MRI and CT images are transformed to generate the registered multimodal images. This ensures accurate spatial correspondence between images of different modalities.

[0021] Step S240: Perform multi-scale feature decomposition on the registered multimodal image, and generate modal features at different scale levels through the Gaussian pyramid decomposition algorithm to obtain a multi-scale modal feature set.

[0022] Multiscale feature decomposition refers to extracting features from registered multimodal images at different scale levels to obtain feature information at different scales. Different scale levels can reflect the features of the image at different resolutions; large-scale levels can capture the macroscopic features of the image, while small-scale levels can capture the detailed features. The Gaussian pyramid decomposition algorithm generates a series of images at different scales by performing multiple Gaussian filtering and downsampling operations on the image, with each scale corresponding to a scale level of modal features.

[0023] As one implementation method, step S240 involves performing multi-scale feature decomposition on the registered multimodal image, generating modal features at different scale levels using a Gaussian pyramid decomposition algorithm to obtain a multi-scale modal feature set. Specifically, this can be implemented as follows: steps S241~S245: Step S241: Set the first set of filtering parameters with different bandwidths to perform Gaussian filtering on the registered multimodal images, and complete multi-scale coverage by adjusting the filtering range in sequence to generate the first-scale modal features.

[0024] Gaussian filtering reduces noise and detail in an image by weighted averaging of each pixel and its neighborhood, while preserving the image's main features. Different bandwidths of filtering parameters determine the smoothness of the Gaussian filter and its responsiveness to features at different scales. By setting a first set of filtering parameters with different bandwidths, Gaussian filtering is applied to the registered multimodal image, smoothing the image at different scales and extracting modal features at different scales. In practice, the first set of filtering parameters can include different standard deviations, for example, [1, 2, 3]. For each standard deviation, Gaussian filtering is applied to the registered multimodal image, resulting in images with different degrees of smoothness; these images are the different sub-features of the first-scale modal features. By sequentially adjusting the filtering range, i.e., changing the standard deviation values, coverage of different scales can be achieved, thereby generating the first-scale modal features.

[0025] Step S242: Set a second set of filtering parameters with different bandwidths to perform Gaussian filtering on the registered multimodal images, and complete multi-scale coverage by adjusting the filtering range in sequence to generate second-scale modal features.

[0026] Similar to step S241, a second set of filtering parameters with different bandwidths is used to perform Gaussian filtering on the registered multimodal image in order to extract modal features within another scale range. The second set of filtering parameters differs from the first set; they can cover different scale ranges, thereby extracting second-scale modal features that differ from the first-scale modal features. In practice, the standard deviation of the second set of filtering parameters can be set to [4, 5, 6]. By performing Gaussian filtering on the registered multimodal image and sequentially adjusting the filtering range, second-scale modal features are generated.

[0027] Step S243: Set a third set of filtering parameters with different bandwidths to perform Gaussian filtering on the registered multimodal images. By adjusting the filtering range in sequence, multi-scale coverage is completed, and third-scale modal features are generated.

[0028] Similarly, a third set of filtering parameters with different bandwidths is used to apply Gaussian filtering to the registered multimodal images to extract modal features at the third scale. This third set of filtering parameters can further expand the scale coverage or cover scale intervals different from the first two sets, thus providing richer information for subsequent feature analysis. By applying Gaussian filtering to the registered multimodal images and sequentially adjusting the filtering range, third-scale modal features are generated.

[0029] Step S244: Perform a Laplace transform on the first-scale modal features, and extract edge features by calculating the second derivative of the Gaussian-filtered image to obtain the first-scale edge features; perform a Laplace transform on the second-scale modal features, and extract edge features by calculating the second derivative of the Gaussian-filtered image to obtain the second-scale edge features; perform a Laplace transform on the third-scale modal features, and extract edge features by calculating the second derivative of the Gaussian-filtered image to obtain the third-scale edge features.

[0030] The Laplacian transform detects edge information in an image by calculating its second derivative. In a Gaussian-filtered image, edges typically appear as rapid changes in grayscale values, and the Laplacian transform highlights these changes, thus extracting edge features. Performing the Laplacian transform on modal features at the first, second, and third scales yields the edge features for each scale. In practice, the Laplacian transform can be implemented using convolution. A Laplacian kernel is convolved with the Gaussian-filtered image, and the second derivative of the image at each pixel is calculated to obtain the edge features. By performing the Laplacian transform on the first, second, and third scale modal features respectively, the first-scale edge features, second-scale edge features, and third-scale edge features are obtained.

[0031] Step S245: Stack the first-scale modal features, the second-scale modal features, the third-scale modal features, the first-scale edge features, the second-scale edge features, and the third-scale edge features, and concatenate the modal features and edge features of different scales in the channel dimension to generate a multi-scale modal feature set.

[0032] Feature stacking is an operation that integrates modal features and edge features at different scales. By concatenating them along the channel dimension, these features can be combined into a larger feature matrix, thus forming a multi-scale modal feature set. In practice, assuming the number of channels for the first-scale modal features, second-scale modal features, third-scale modal features, first-scale edge features, second-scale edge features, and third-scale edge features are C1, C2, C3, C4, C5, and C6, respectively, after concatenating them along the channel dimension, the number of channels in the multi-scale modal feature set is C1+C2+C3+C4+C5+C6. This effectively integrates features of different scales and types, providing richer feature information for subsequent cross-modal fusion.

[0033] Step S250: Perform cross-attention fusion on the magnetic resonance modal features and tomographic scan modal features in the multi-scale modal feature set, and complete the dynamic feature weight allocation by calculating the feature correlation matrix between modalities to generate cross-modal fusion features.

[0034] Cross-attention fusion dynamically assigns weights to different modal features by calculating the intermodal feature correlation matrix, thereby achieving effective fusion of features from different modalities. The intermodal feature correlation matrix reflects the similarity and correlation between magnetic resonance imaging (MRI) modal features and computed tomography (CT) modal features. By analyzing and processing this matrix, the importance weight of each modal feature can be obtained, and then the features from different modalities can be fused based on these weights.

[0035] As one implementation method, step S250 involves performing cross-attention fusion on the magnetic resonance modal features and tomographic scan modal features in the multi-scale modal feature set, and completing dynamic feature weight allocation by calculating the inter-modal feature correlation matrix to generate cross-modal fusion features. Specifically, this can be implemented as follows: steps S251~S258: Step S251: Separate the magnetic resonance modal feature subset and the tomographic scan modal feature subset from the multi-scale modal feature set.

[0036] Since the multi-scale modal feature set includes both magnetic resonance (MRI) and computed tomography (CT) modal features, it is necessary to separate these two modalities before performing cross-attention fusion. In practice, the MRI and CT modal features can be extracted separately from the multi-scale modal feature set based on their source or channel information, forming MRI and CT modal feature subsets respectively. For example, if the multi-scale modal feature set is a four-dimensional tensor with dimensions [batch size, height, width, number of channels], the channels belonging to the MRI mode and the channels belonging to the CT mode can be extracted separately based on the channel information, forming two independent feature subsets.

[0037] Step S252: Flatten the feature subset of the magnetic resonance mode, convert the three-dimensional feature tensor into a two-dimensional feature matrix, and obtain the flattened magnetic resonance features.

[0038] Feature flattening is an operation that converts a three-dimensional feature tensor into a two-dimensional feature matrix. It combines the spatial dimensions of the features, facilitating matrix operations. In practice, assuming a subset of magnetic resonance modal features is a three-dimensional tensor with dimensions [height, width, number of channels], by merging the height and width dimensions, it is converted into a two-dimensional matrix with dimensions [height × width, number of channels]. This two-dimensional matrix is ​​the magnetic resonance flattened feature.

[0039] Step S253: Flatten the feature subset of the tomographic scan mode, convert the three-dimensional feature tensor into a two-dimensional feature matrix, and obtain the flattened features of the tomographic scan.

[0040] Similar to step S252, the subset of tomographic modal features is flattened, converting it from a three-dimensional feature tensor into a two-dimensional feature matrix to obtain the flattened tomographic features. This allows the spatial dimensions of the tomographic modal features to be merged, facilitating subsequent matrix operations with the flattened magnetic resonance features.

[0041] Step S254: Calculate the correlation matrix between the magnetic resonance flattening features and the tomographic scan flattening features through feature similarity analysis. Each element of the correlation matrix represents the degree of similarity between the magnetic resonance modal features and the tomographic scan modal features.

[0042] As one implementation method, step S254, calculating the correlation matrix between the magnetic resonance flattening features and the computed tomography flattening features through feature similarity analysis, can be specifically implemented as follows: steps S2541~S2544: Step S2541: Perform matrix transpose and multiplication operations on the magnetic resonance flattening feature and the tomographic flattening feature. This is done by multiplying the magnetic resonance flattening feature matrix with the transpose of the tomographic flattening feature matrix to obtain the initial correlation matrix.

[0043] Matrix transpose is the operation of exchanging the rows and columns of a matrix, while multiplication is the operation of multiplying corresponding elements of two matrices and summing them. In practice, assuming the dimensions of the MRI flattening feature matrix are [M, C] and the dimensions of the tomography flattening feature matrix are [N, C], transposing the tomography flattening feature matrix changes its dimensions to [C, N]. Then, multiplying the MRI flattening feature matrix with the transposed tomography flattening feature matrix yields an initial correlation matrix of dimension [M, N]. Each element of this initial correlation matrix represents the inner product between an eigenvector of the MRI flattening feature matrix and an eigenvector of the tomography flattening feature matrix, reflecting the degree of similarity between them.

[0044] Step S2542: Normalize the row-direction standard deviation of the initial correlation matrix by subtracting the row mean from each row element and then dividing by the row standard deviation to obtain the standardized correlation matrix.

[0045] Row-directed standard deviation normalization is a method used to normalize matrices, ensuring that each row of elements has the same mean and standard deviation, thus eliminating scale differences between rows. In practice, for each row of the initial correlation matrix, the mean and standard deviation of the row's elements are first calculated. Then, the row mean is subtracted from each element, and the result is divided by the row standard deviation to obtain the standardized row elements. By performing this process on each row of the initial correlation matrix, a standardized correlation matrix is ​​obtained. This ensures that the elements in each row of the standardized correlation matrix are numerically comparable, facilitating subsequent processing.

[0046] Step S2543: Perform an exponential function mapping on each element in the standardized correlation matrix to enhance the weight of large numerical elements through the natural exponential function, thereby strengthening the strong correlation between features.

[0047] Exponential function mapping can non-linearly map the elements in a standardized association matrix, thereby enhancing the weight of large numerical elements. In practice, for each element in the standardized association matrix, the natural exponential function is used for mapping, i.e., the natural exponent is taken for each element. Since the natural exponential function is a monotonically increasing function, and its growth rate accelerates with the increase of the input value, it can strengthen the strong correlation between features and highlight feature pairs with high similarity.

[0048] Step S2544: Perform global maximum and minimum value normalization on the processed standardized association matrix, compress the matrix elements to a preset numerical range, and generate a feature similarity association matrix.

[0049] Global maximum-minimum normalization is a method to compress matrix elements into a preset numerical range, typically [0,1]. In practice, the maximum and minimum values ​​in the processed standardized association matrix are first identified. Then, for each element in the matrix, the formula (element value - minimum value) / (maximum value - minimum value) is used to map it to the [0,1] interval, resulting in a feature similarity association matrix. This ensures that the elements in the feature similarity association matrix have a uniform scale, facilitating subsequent weight allocation operations.

[0050] Step S255: Normalize the row direction of the correlation matrix and adjust the weight distribution using a normalization function to obtain the magnetic resonance attention weights; normalize the column direction of the correlation matrix and adjust the weight distribution using a normalization function to obtain the tomographic scan attention weights.

[0051] Row and column normalization are used to convert the correlation matrix into an attention weight matrix. Normalization functions ensure that the sum of the matrix elements is 1, thus achieving a reasonable allocation of weights. In practice, the softmax function can be used for row and column normalization of the correlation matrix. For row normalization, each row element of the correlation matrix is ​​input into the softmax function, and the output is the MRI attention weight, representing the relative importance of each tomographic feature vector in the association between each MRI feature vector and the tomographic feature vector. For column normalization, each column element of the correlation matrix is ​​input into the softmax function, and the output is the tomographic attention weight, representing the relative importance of each MRI feature vector in the association between each tomographic feature vector and the MRI feature vector.

[0052] Step S256: Perform matrix multiplication between the magnetic resonance attention weight and the tomographic flattening feature to obtain the magnetic resonance guided tomographic feature.

[0053] Matrix multiplication is the operation of multiplying corresponding elements of two matrices and summing them. In practice, multiplying the magnetic resonance imaging (MRI) attention weight matrix with the tomographic flattening feature matrix yields the MRI-guided tomographic features, representing the weighted features of the tomographic scan guided by the MRI features. Similarly, multiplying the tomographic attention weight matrix with the MRI flattening feature matrix yields the tomographic-guided MRI features, representing the weighted features of the MRI features guided by the tomographic scan features.

[0054] Step S257: Add the magnetic resonance-guided tomography features to the tomography-guided magnetic resonance features element by element to obtain the fused flattened features.

[0055] Element-wise addition involves adding corresponding elements from two matrices. In practice, the corresponding elements of the MRI-guided tomographic feature matrix and the tomographic-guided MRI feature matrix are added to obtain the fused and flattened features. This effectively fuses MRI and tomographic features, fully utilizing the information from both modalities.

[0056] Step S258: Reshape the fused flattened features to restore the two-dimensional fused flattened features to a three-dimensional feature tensor structure, and generate cross-modal fused features.

[0057] Feature reshaping is the operation of restoring two-dimensional fused and flattened features into a three-dimensional feature tensor structure. It restores the spatial dimensions of the features, facilitating subsequent processing and analysis. In practice, assuming the dimensions of the fused and flattened feature are [height × width, number of channels], based on the previous feature flattening process, it is restored to a three-dimensional tensor with dimensions [height, width, number of channels]. This three-dimensional tensor is the cross-modal fusion feature.

[0058] Step S260: Suppress noise in the cross-modal fusion features by eliminating artifact interference in the fusion features through a non-local similarity filtering algorithm to obtain denoised fusion features.

[0059] Noise suppression aims to reduce noise and artifact interference in cross-modal fusion features, improving feature quality and reliability. Nonlocal similarity filtering algorithms are image patch similarity-based filtering methods. They eliminate noise and artifacts by finding similar image patches in the image and performing a weighted average of these patches. In practice, for each image patch in the cross-modal fusion feature, similar image patches are searched throughout the entire image. Then, these patches are weighted and averaged according to their similarity to obtain the filtered image patch. By performing this processing on the entire cross-modal fusion feature, artifact interference is eliminated, resulting in a denoised fusion feature.

[0060] Step S270: The denoised and fused features are regularized in terms of feature dimensions. The multi-scale features are unified to the same dimensional space through the feature mapping function to generate the initial feature map.

[0061] Feature dimension normalization aims to unify multi-scale features in the denoising and fusion features into a common dimensional space, facilitating subsequent analysis and processing. The feature mapping function can be a linear or non-linear transformation, mapping features of different scales to the same dimensional space. In practice, assuming the denoising and fusion features contain features of different scales and dimensions, the feature mapping function transforms these features to unify their dimensions. For example, a fully connected layer can be used as the feature mapping function. The denoising and fusion features are input into the fully connected layer, and by learning the weight matrix, the features are mapped to the same dimensional space, generating an initial feature map.

[0062] Step S300: Input the initial feature map into the deep feature fusion network to enhance and identify the lesion area, and obtain the lesion area feature map.

[0063] Deep feature fusion networks are a type of neural network based on deep learning. They can further extract and fuse features from the initial input feature map, enhancing the feature representation of lesion areas and thus achieving accurate identification of lesion areas. A lesion area feature map is a map containing feature information of the lesion area, helping doctors to more intuitively observe and analyze the characteristics and location of lesions.

[0064] As one implementation method, step S300 involves inputting the initial feature map into a deep feature fusion network for enhanced lesion region recognition to obtain a lesion region feature map. Specifically, this can be implemented as follows: steps S310~S380: Step S310: Input the initial feature map into the low-level feature encoding module of the deep feature fusion network for primary feature extraction to obtain low-level encoded features. The low-level feature encoding module contains multiple convolutional operation units connected in sequence. Each convolutional operation unit consists of a convolutional layer and a batch normalization layer.

[0065] The low-level feature encoding module is the foundation of the deep feature fusion network, performing basic feature extraction on the initial input feature map. Convolutional operation units (COUs) are the basic components of this module. Convolutional layers extract local features from the input feature map by performing convolution operations with kernels. Batch normalization layers normalize the outputs of the convolutional layers, accelerating network training and convergence. In practice, the initial feature map is used as input and sequentially passed through multiple COUs. Each COU performs convolution and batch normalization on the input to obtain the low-level encoded features. For example, the low-level feature encoding module can contain three COUs, each using a 3×3 kernel. The batch normalization layer normalizes the outputs of the convolutional layers, ultimately yielding the low-level encoded features.

[0066] Step S320: Input the bottom-level encoded features into the middle-level feature encoding module for semantic feature enhancement to obtain the middle-level encoded features. The middle-level feature encoding module completes feature reuse and gradient transfer through the residual connection structure.

[0067] The intermediate feature encoding module is the middle part of the deep feature fusion network. It can further extract features and enhance semantics from the lower-level encoded features. The residual connection structure is a technique used to solve the vanishing and exploding gradient problems in deep neural network training. It can directly add input features to the output features, making it easier for the network to learn residual information, while simultaneously achieving feature reuse and efficient gradient propagation. In practice, the lower-level encoded features are input to the intermediate feature encoding module. Through the residual connection structure, the input features are added to the features processed by convolution and non-linear activation functions to obtain the intermediate-level encoded features. This enhances the semantic information of the features and improves the network's ability to identify lesion regions.

[0068] Step S330: Input the mid-level coding features into the high-level feature coding module for global feature integration to obtain high-level coding features. The high-level coding module extracts cross-region feature association information through global context aggregation operation.

[0069] The high-level feature encoding module is a high-level part of the deep feature fusion network. It can perform global feature integration on mid-level encoded features to extract cross-regional feature correlation information. Global context aggregation is a method for capturing global information in the feature map, which can comprehensively consider features from different regions to extract global features related to the lesion area. In practice, mid-level encoded features are input to the high-level feature encoding module. Through global context aggregation operations, such as global average pooling or attention mechanisms, information from different regions in the feature map is aggregated to obtain high-level encoded features. This enhances the network's overall understanding and recognition ability of lesion areas.

[0070] Step S340: Perform feature pyramid fusion on the bottom-level coding features, middle-level coding features and high-level coding features. After scaling the high-level features with the low-level features through upsampling, the features are superimposed to generate multi-scale fused features.

[0071] Feature pyramid fusion aims to combine encoded features from different levels, fully utilizing feature information at different scales to improve the accuracy of lesion region recognition. Upsampling involves enlarging the scale of high-level features to align with that of low-level features, facilitating subsequent stacking operations. In practice, firstly, high-level encoded features are upsampled to the same scale as mid-level encoded features. Then, the upsampled high-level and mid-level features are aligned and stacked to obtain the mid-level fused features. Next, the mid-level fused features are upsampled to the same scale as the bottom-level encoded features, and again aligned and stacked to obtain the bottom-level fused features. Finally, the bottom-level fused features are convolved and integrated to generate multi-scale fused features.

[0072] As one implementation method, step S340 involves performing feature pyramid fusion on the bottom-level coding features, middle-level coding features, and high-level coding features. This is achieved by upsampling the high-level features and low-level features, aligning their scales, and then superimposing them to generate multi-scale fused features. Specifically, this can be implemented as follows: steps S341~S347: Step S341: Perform the first upsampling on the high-level coding features and complete the feature scale amplification through interpolation algorithm to obtain the first upsampled features.

[0073] Interpolation algorithms can generate more pixels by interpolating pixels in high-level coded features, thereby amplifying the feature scale. In practice, interpolation algorithms can be bilinear interpolation or nearest-neighbor interpolation. For example, bilinear interpolation can be used to upsample high-level coded features by linearly interpolating the gray values ​​of neighboring pixels to calculate the gray value of the newly added pixels, resulting in the first upsampled feature. This can amplify the scale of the high-level coded features to be the same as that of the mid-level coded features, facilitating subsequent feature alignment and overlay operations.

[0074] Step S342: Align the first upsampled feature with the middle-layer encoded feature by adjusting the spatial size and position of the feature to complete feature matching and obtain the middle-layer aligned feature.

[0075] Feature alignment ensures that the first upsampled feature and the intermediate encoded feature correspond accurately in space, enabling effective overlay operations. By adjusting the spatial size and position of the features, the first upsampled feature and the intermediate encoded feature can be matched in the same coordinate system. In practice, transformation operations such as translation, rotation, and scaling can be used to adjust the first upsampled feature so that its spatial size and position are consistent with the intermediate encoded feature, resulting in an aligned intermediate feature.

[0076] Step S343: Add the mid-layer aligned features to the first upsampled features element by element to obtain the mid-layer fused features.

[0077] As one implementation method, step S343 involves element-wise addition of the intermediate alignment features and the first upsampled features to obtain the intermediate fusion features. Specifically, this can be implemented as follows: steps S3431~S3434: Step S3431: Perform element-wise multiplication on the mid-level aligned features and the first upsampled features to capture the local correlation information between the two features and obtain the interaction feature matrix.

[0078] Element-wise multiplication is an operation that multiplies the corresponding elements of the mid-level aligned feature and the first upsampled feature, capturing the local correlation information between the two features. In practice, assuming that the mid-level aligned feature and the first upsampled feature have the same dimensions (height, width, number of channels), multiplying their corresponding elements yields an interaction feature matrix of the same dimensions. This interaction feature matrix reflects the interaction and correlation between the mid-level aligned feature and the first upsampled feature at each pixel.

[0079] Step S3432: Perform global average pooling on the interaction feature matrix, calculate the spatial mean of the interaction feature matrix to obtain the overall interaction intensity, and obtain the interaction intensity value.

[0080] Global average pooling calculates the mean of the feature matrix across its spatial dimensions, yielding a scalar value. In practice, for the interaction feature matrix, average pooling is performed across its height and width dimensions to calculate the spatial mean for each channel. The means of all channels are then summed to obtain the overall interaction strength, i.e., the interaction strength value. This interaction strength value reflects the overall correlation between the mid-layer aligned features and the first upsampled features.

[0081] Step S3433: Dynamically adjust the fusion weight coefficients based on the interaction intensity value, which is accomplished through a preset nonlinear mapping function. The higher the interaction intensity value, the more the weight coefficient allocation tends to retain the original feature details.

[0082] The preset nonlinear mapping function can dynamically adjust the fusion weight coefficients based on the interaction strength value. In practice, assuming the preset nonlinear mapping function is a sigmoid function, the interaction strength value is input into the sigmoid function, resulting in a weight coefficient between 0 and 1. The higher the interaction strength value, the closer the weight coefficient is to 1, indicating a stronger correlation between the two features. In this case, the weight coefficient allocation tends to preserve the original feature details. Conversely, the lower the interaction strength value, the closer the weight coefficient is to 0, indicating a weaker correlation between the two features. In this case, the weight coefficient allocation tends to perform feature fusion.

[0083] Step S3434: The mid-layer aligned features and the first upsampled features are weighted and summed according to the dynamically adjusted weight coefficients. Feature fusion is completed through element-wise multiplication and addition operations to obtain the mid-layer fused features.

[0084] Weighted summation is an operation that adds the intermediate-level aligned features and the first upsampled features according to dynamically adjusted weight coefficients. In practice, the intermediate-level aligned features are multiplied by the weight coefficients, and the first upsampled features are multiplied by (1 - the weight coefficients). The two results are then added together to obtain the intermediate-level fused features. This allows for dynamic adjustment of their weights in the fusion process based on the degree of correlation between the two features, achieving more effective feature fusion.

[0085] Step S344: Perform a second upsampling on the mid-layer fusion features, and complete the feature scale amplification through an interpolation algorithm to obtain the second upsampled features.

[0086] Similar to step S341, an interpolation algorithm is used to perform a second upsampling on the middle-layer fused features, scaling them up to the same scale as the bottom-layer encoded features, resulting in the second upsampled features. For example, a bilinear interpolation algorithm can also be used to interpolate based on the gray values ​​of adjacent pixels, generating more pixels and thus scaling up the feature scale.

[0087] Step S345: Align the second upsampled feature with the bottom-level encoded feature by adjusting the spatial size and position of the feature to complete feature matching and obtain the bottom-level aligned feature.

[0088] Similar to step S342, by adjusting the spatial size and position of the features, the second upsampled features are made to correspond accurately with the underlying encoded features in space, thus obtaining the underlying aligned features. Transformation operations such as translation, rotation, and scaling can be used to adjust the second upsampled features to make them consistent with the spatial size and position of the underlying encoded features.

[0089] Step S346: Add the bottom alignment features and the second upsampled features element by element to obtain the bottom fusion features.

[0090] The elements at corresponding positions of the bottom-level aligned features and the second upsampled features are added together to obtain the bottom-level fused features. This effectively superimposes the bottom-level encoded features and the upsampled and fused mid-level features, further integrating feature information from different levels.

[0091] Step S347: Perform convolutional integration on the underlying fusion features, and adjust the channel dimension of the fusion features through the convolution kernel to generate multi-scale fusion features.

[0092] Convolutional integration is used to further extract features and adjust channel dimensions of the underlying fused features. A convolutional kernel performs convolution operations on the underlying fused features, extracting feature information and adjusting channel dimensions. In practice, a convolutional kernel is used with the underlying fused features to adjust the number of channels to a preset number, resulting in multi-scale fused features. This allows for the integration of feature information from different levels, generating a feature map containing multi-scale features.

[0093] Step S350: Input the multi-scale fused features into the spatial attention module to allocate regional attention, and adjust the pixel-level weights of the feature map by generating a two-dimensional spatial attention map to obtain spatially weighted features.

[0094] The spatial attention module generates a two-dimensional spatial attention map, where each pixel value represents the importance weight of the feature at that location within the overall feature map. By multiplying the spatial attention map by the multi-scale fused feature at the pixel level, pixel-level weighting of the feature map can be adjusted, highlighting features in important regions and suppressing features in unimportant regions, resulting in spatially weighted features. In practice, the spatial attention module can be implemented using a convolutional neural network. Taking the multi-scale fused feature as input, it passes through a series of convolutional layers and non-linear activation functions to generate a two-dimensional spatial attention map of the same size as the multi-scale fused feature. Then, the spatial attention map is multiplied by the corresponding pixel values ​​of the multi-scale fused feature to obtain the spatially weighted features.

[0095] Step S360: Input the spatially weighted features into the channel attention module to filter the importance of the feature channels, calculate the channel weights through the feature channel importance evaluation mechanism, and obtain the channel-weighted features.

[0096] As one implementation method, step S360 involves inputting the spatially weighted features into the channel attention module for feature channel importance filtering, calculating channel weights through a feature channel importance evaluation mechanism, and obtaining channel-weighted features. Specifically, this can be implemented as follows: steps S361~S365: Step S361: Perform global feature aggregation on the spatially weighted features, and statistically aggregate the spatial dimensions of each feature channel to obtain the channel description vector.

[0097] Global feature aggregation aims to statistically aggregate spatially weighted features across spatial dimensions to obtain an overall description of each feature channel. In practice, methods such as global average pooling or global max pooling can be used to statistically aggregate each feature channel of the spatially weighted features along the height and width dimensions, resulting in a one-dimensional channel description vector. For example, using global average pooling, the mean of each feature channel in the spatial dimension is calculated, yielding each element of the channel description vector.

[0098] Step S362: Perform the first nonlinear mapping on the channel description vector to map the channel description vector to a low-dimensional space to obtain the intermediate feature vector; activate the intermediate feature vector to obtain the activated feature vector.

[0099] The first nonlinear mapping aims to map the channel description vectors to a lower-dimensional space, reducing the feature dimensionality and introducing a nonlinear transformation to enhance the model's expressive power. In practice, a fully connected layer can be used as the nonlinear mapping function. The channel description vectors are input into the fully connected layer, and by learning the weight matrix, they are mapped to a lower-dimensional space, obtaining intermediate feature vectors. Then, activation operations, such as using the ReLU activation function, are performed on these intermediate feature vectors to enhance the nonlinear expression of the features, resulting in activated feature vectors.

[0100] Step S363: Perform a second nonlinear mapping on the activation feature vector to map the activation feature vector back to the original channel dimension, and obtain the channel weight vector.

[0101] The second nonlinear mapping is to map the activation feature vector back to the original channel dimension, obtaining the weight of each feature channel. Similarly, a fully connected layer can be used as the nonlinear mapping function. The activation feature vector is input into the fully connected layer, and by learning the weight matrix, it is mapped back to the original channel dimension to obtain the channel weight vector.

[0102] Step S364: Normalize the channel weight vector and adjust the channel weights to a preset distribution range to obtain normalized channel weights.

[0103] Normalization is used to adjust the elements of the channel weight vector to a preset distribution range, typically [0, 1]. In practice, the softmax function can be used to normalize the channel weight vector, converting each element of the channel weight vector into a probability value between 0 and 1, with the sum of all elements being 1, thus obtaining the normalized channel weights.

[0104] Step S365: Multiply the normalized channel weights and spatial weighted features by channel dimension weighting, and multiply the feature map of each channel by the corresponding channel weight to obtain the channel weighted features.

[0105] Channel-dimensional weighted multiplication is an operation that multiplies the normalized channel weights and spatially weighted features along the channel dimension. In practice, assuming the dimensions of the spatially weighted features are [height, width, number of channels] and the dimension of the normalized channel weights is [number of channels], the feature map of each channel of the spatially weighted features is multiplied by its corresponding channel weight to obtain the channel-weighted features. This allows for channel-dimensional weighting adjustments to the spatially weighted features based on the importance of each channel, highlighting features from important channels and suppressing features from less important channels.

[0106] Step S370: Input the channel-weighted features into the lesion region classification module for pixel-level classification. The lesion region classification module uses a convolution classifier to predict the lesion probability of each pixel and obtain the initial lesion region features.

[0107] The lesion region classification module performs pixel-level classification on channel-weighted features, predicting the probability that each pixel belongs to a lesion region. The convolution classifier is the core component of this module, extracting and classifying features from the channel-weighted features through convolution operations. In practice, the channel-weighted features are input into the convolution classifier, which processes them through a series of convolutional layers and non-linear activation functions, outputting the probability that each pixel belongs to a lesion region, thus obtaining the initial lesion region features. For example, the convolution classifier can use a single convolutional layer with one output channel, using a sigmoid activation function to convert the output value into a probability value between 0 and 1, representing the probability that each pixel belongs to a lesion region.

[0108] Step S380: Refine the edges of the initial lesion area features, correct the boundary contour of the lesion area through morphological optimization operations, and generate a lesion area feature map.

[0109] Edge refinement aims to correct the boundary contours of the lesion region in the initial lesion region features, making them more accurate and clear. Morphological optimization operations, such as dilation, erosion, opening, and closing operations, can correct and optimize the boundaries of the lesion region. In practice, the initial lesion region features are first binarized to obtain a binary image, where 1 represents the lesion region and 0 represents the non-lesion region. Then, morphological optimization operations are applied, such as erosion to remove small noise points at the edges of the lesion region, and dilation to fill small holes inside the lesion region. Finally, a binary image of the lesion region with more accurate and clear boundaries is obtained. This binary image is then combined with the initial lesion region features to generate a lesion region feature map.

[0110] Step S400: Perform three-dimensional topological structure analysis on the feature map of the lesion area to construct a three-dimensional structural model of the lesion area.

[0111] Three-dimensional topological analysis aims to extract the three-dimensional topological structure information of a lesion region from a two-dimensional feature map, and to construct a three-dimensional structural model of the lesion region. This three-dimensional structural model can more intuitively display the morphology, size, and spatial location of the lesion region, helping doctors better understand the characteristics of the lesion and formulate treatment plans.

[0112] As one implementation method, step S400 involves performing three-dimensional topological structure analysis on the feature map of the lesion region to construct a three-dimensional structural model of the lesion region. Specifically, this can be implemented as follows: steps S410~S480: Step S410: Perform inter-layer spatial alignment on the feature maps of the lesion area. Spatial alignment is completed by calculating the geometric transformation parameters of adjacent feature maps to obtain the aligned feature map sequence.

[0113] Inter-slice spatial alignment aims to eliminate spatial discrepancies between feature maps of adjacent lesion regions, ensuring accurate alignment in three-dimensional space. Geometric transformation parameters describe the translation, rotation, and scaling relationships between adjacent feature maps. In practice, a reference feature map is first selected, and then the geometric transformation parameters between other adjacent feature maps and the reference feature map are calculated. Feature matching algorithms, such as SIFT or SURF, can be used to find matching feature points in adjacent feature maps, and then the geometric transformation parameters are calculated based on these matching points. Finally, based on the calculated geometric transformation parameters, the adjacent feature maps are transformed to spatially align with the reference feature map, resulting in a calibrated feature map sequence.

[0114] Step S420: Perform voxelization transformation on the calibrated feature map sequence to map the two-dimensional feature map sequence to the three-dimensional voxel grid space and generate three-dimensional voxel data.

[0115] As one implementation method, step S420 involves voxelizing the calibrated feature map sequence to map the two-dimensional feature map sequence to a three-dimensional voxel grid space, generating three-dimensional voxel data. Specifically, this can be implemented as follows: steps S421~S427: Step S421: Determine the interlayer spacing parameter and plane resolution parameter of each feature map in the calibrated feature map sequence.

[0116] Interslice spacing refers to the distance between adjacent feature maps, while planar resolution refers to the pixel resolution of the feature maps on a plane. In practice, the interslice spacing and planar resolution of each feature map can be determined based on the acquisition parameters of the medical imaging equipment and the actual situation of the calibrated feature map sequence. For example, the interslice spacing can be obtained by measuring the physical distance between adjacent feature maps, and the planar resolution can be calculated based on the pixel size and actual physical dimensions of the feature maps.

[0117] Step S422: Construct a three-dimensional voxel grid coordinate system based on the interlayer spacing parameter and the planar resolution parameter. The origin of the three-dimensional voxel grid coordinate system is set at the reference position of the starting slice of the feature map sequence.

[0118] The 3D voxel grid coordinate system is a coordinate system used to represent 3D voxel data, which can map 2D feature map sequences into 3D space. The grid size and spacing of the 3D voxel grid coordinate system can be determined based on the interlayer spacing parameter and the planar resolution parameter. In practice, the origin of the 3D voxel grid coordinate system is set at the reference position of the starting slice of the feature map sequence, and then the position of each voxel in 3D space is determined according to the interlayer spacing parameter and the planar resolution parameter.

[0119] Step S423: Map each feature map in the calibrated feature map sequence to the corresponding slice position in the three-dimensional voxel grid coordinate system.

[0120] Each feature map in the calibrated feature map sequence is mapped to its corresponding slice position in the 3D voxel grid coordinate system according to its order in the sequence and the interlayer spacing parameter. In practice, the slice position of each feature map in 3D space is calculated based on the interlayer spacing parameter of the feature map and the origin position of the 3D voxel grid coordinate system, and then the pixel values ​​in the feature map are mapped to the corresponding voxels.

[0121] Step S424: Perform interpolation calculations on the voxel values ​​between adjacent slices, and use a trilinear interpolation algorithm to achieve a smooth transition of voxel values, filling the blank voxels in the three-dimensional voxel grid.

[0122] The trilinear interpolation algorithm can calculate the voxel value of a blank voxel based on the voxel values ​​of adjacent slices, achieving a smooth transition of voxel values. In practice, for blank voxels between adjacent slices, the trilinear interpolation algorithm is used to calculate the voxel value of the blank voxel based on the voxel values ​​of its eight surrounding voxels. This fills the blank voxels in the 3D voxel grid, making the 3D voxel data more complete and continuous.

[0123] Step S425: Threshold the interpolated three-dimensional voxel grid, determine the segmentation threshold based on the feature value distribution of the lesion area, and obtain binarized three-dimensional voxel data.

[0124] Thresholding can determine a segmentation threshold based on the feature value distribution of the lesion region. Voxels with voxel values ​​greater than the threshold are marked as lesion regions, while voxels with voxel values ​​less than the threshold are marked as non-lesion regions. In practice, the distribution of voxel values ​​in the interpolated 3D voxel grid is first statistically analyzed. Based on the feature value distribution of the lesion region, a suitable segmentation threshold is determined. Then, each voxel value in the interpolated 3D voxel grid is compared with the segmentation threshold. Voxels with values ​​greater than the threshold are assigned a value of 1, and voxels with values ​​less than the threshold are assigned a value of 0, resulting in binarized 3D voxel data.

[0125] Step S426: Morphological closure is performed on the binarized 3D voxel data. The voxel data is morphologically repaired using spherical structural elements to fill the hole areas in the voxel data.

[0126] Morphological closure is a morphological operation used to process binary images. It involves dilation followed by erosion to fill holes in binary 3D voxel data, thus restoring the data's structure. A spherical structuring element is a structuring element used in morphological operations; it simulates the shape of a sphere to process voxel data. In practice, a spherical structuring element is used to dilate the binary 3D voxel data, expanding the boundaries of lesion areas outward and filling small holes. Then, a spherical structuring element is used to erode the dilated voxel data, shrinking the boundaries of the lesion areas back to their original positions and removing noise. This morphological closure operation fills the holes in the voxel data, restoring its structure.

[0127] Step S427: Convert the processed binarized 3D voxel data into 3D voxel data.

[0128] The processed binarized 3D voxel data is converted back to 3D voxel data, that is, the binarized voxel values ​​(0 and 1) are converted back to the original voxel values. In practice, based on the original 3D voxel mesh and the binarized 3D voxel data, the voxel position corresponding to 1 in the binarized voxel data is assigned the voxel value of that position in the original 3D voxel mesh, and the voxel position corresponding to 0 in the binarized voxel data is assigned 0, thus obtaining the final 3D voxel data.

[0129] Step S430: Extract connected components from the three-dimensional voxel data, and segment the lesion region from the three-dimensional voxel data using a region growing algorithm to obtain a voxel subset of the lesion region.

[0130] Region growing algorithms can start from a seed point and continuously merge adjacent voxels with similar features into the same region until certain stopping conditions are met. In practice, one or more seed points are first selected; these seed points are usually located within the lesion region. Then, starting from the seed point, adjacent voxels are merged into the same region according to preset growth criteria, such as the similarity of voxel values. This process is repeated until no more voxels can be merged, resulting in a connected region. By processing all voxels in the 3D voxel data using a region growing algorithm, all lesion regions can be segmented, obtaining a subset of voxels representing the lesion region.

[0131] Step S440: Construct a surface mesh for a subset of voxels in the lesion area. Generate a triangular mesh from the voxel data using the moving cube algorithm to generate an initial surface mesh model.

[0132] The Moving Cube algorithm converts voxel data into a triangular mesh, thereby constructing a surface model of the lesion area. In practice, the Moving Cube algorithm divides the 3D voxel data into small cubes. Based on the distribution of voxel values ​​within each cube, it determines the intersection between the cube and the surface, and then generates corresponding triangular patches based on the intersection. By processing all the cubes and connecting the generated triangular patches, an initial surface mesh model is obtained. This initial surface mesh model can visually represent the surface morphology of the lesion area.

[0133] Step S450: Simplify the initial surface mesh model by reducing the number of faces in the mesh through an edge folding algorithm to obtain a simplified surface mesh.

[0134] As one implementation method, step S450 involves simplifying the initial surface mesh model by reducing the number of faces in the mesh using an edge folding algorithm to obtain a simplified surface mesh. Specifically, this can be implemented as follows: steps S451 to S457: Step S451: Calculate the normal vector of each triangular facet in the initial surface mesh model. The facet normal vector is calculated from the normal vector of the vertex of the triangular facet using the area-weighted average method.

[0135] A normal vector is a vector perpendicular to the surface of a triangular facet. It represents the orientation of the facet and the direction of the surface's normal. The area-weighted average method is used to calculate the normal vector of a triangular facet. It calculates the normal vector based on the normal vectors of the facet's vertices and the facet's area. In practice, the normal vector of each vertex is first calculated, and then a weighted average is taken based on the facet's area to obtain the facet's normal vector. For example, assuming the normal vectors of the three vertices of a triangular facet are N1, N2, and N3, and the facet's area is S, the facet's normal vector N can be calculated using the formula N = (S1 × N1 + S2 × N2 + S3 × N3) / S, where S1, S2, and S3 are the areas of the sub-triangles corresponding to the three vertices.

[0136] Step S452: Calculate the curvature characteristics of each triangular facet based on the normal vector, and characterize the degree of curvature of the mesh surface through the average curvature algorithm.

[0137] The average curvature algorithm is used to calculate the curvature of a mesh surface. It calculates the degree of curvature of the mesh surface at a given triangular facet based on the normal vector of that facet. In practice, for each triangular facet, the average curvature at that facet is calculated based on the normal vectors of its neighboring facets and its own normal vector. A larger average curvature indicates a greater degree of curvature of the mesh surface at that facet; a smaller average curvature indicates a flatter mesh surface at that facet.

[0138] Step S453: Sort the triangular facets according to their curvature characteristics, and mark the triangular facets with significant changes in curvature characteristics as high-priority facets.

[0139] Based on the calculated curvature characteristics of each triangular facet, the facets are ranked by importance. Facets with significant changes in curvature characteristics typically contain more detailed information and have a significant impact on the shape and surface features of the mesh; therefore, they are marked as high-priority facets. In practice, facets can be categorized into different priority levels based on the magnitude of the curvature characteristic variation: facets with larger variations in curvature characteristics are marked as high-priority facets, while facets with smaller variations in curvature characteristics are marked as low-priority facets.

[0140] Step S454: Set the mesh simplification scale conditions and determine the number of target patches that need to be retained.

[0141] The mesh simplification ratio is a parameter used to control the degree of mesh simplification, determining the number of target faces to be retained. In practice, an appropriate mesh simplification ratio is set based on actual needs and application scenarios. For example, setting the mesh simplification ratio to 50% means that the number of faces in the initial surface mesh model needs to be reduced to 50% of its original number, i.e., the number of target faces to be retained is 50% of the initial number of faces.

[0142] Step S455: Iteratively simplify the initial surface mesh model using the edge folding algorithm. Mesh simplification is achieved by shrinking an edge into a vertex. In each iteration, edges of the facet with a gentle curvature change are selected for folding.

[0143] In one implementation, in step S455, each iteration selects the edge of the surface with a gentle change in curvature characteristics for folding, specifically implemented as follows: steps S4551~S4555: Step S4551: Calculate the side length parameters of all edges contained in the surface with a gentle curvature change, which is obtained by calculating the spatial Euclidean distance between the two vertices of the edge.

[0144] The edge length parameter refers to the length of an edge, which can be obtained by calculating the spatial Euclidean distance between the two vertices of the edge. In practice, for each edge contained in a surface with gently changing curvature, the length of the edge is calculated using the Euclidean distance formula based on the three-dimensional coordinates of its two vertices.

[0145] Step S4552: Compare the side length parameter with the preset side length threshold condition. The preset side length threshold condition is dynamically determined based on the average side length of the initial grid, and a set of candidate edges that meet the condition is selected.

[0146] The preset edge length threshold is a condition used to filter candidate edges, which can be dynamically determined based on the average edge length of the initial grid. In practice, the average edge length of all edges in the initial grid is first calculated, and then a preset edge length threshold is determined based on the average edge length, for example, the preset edge length threshold is 1.2 times the average edge length. The calculated edge length parameter of each edge is compared with the preset edge length threshold, and edges with a length less than or equal to the preset edge length threshold are selected to form a candidate edge set.

[0147] Step S4553: Calculate the folding cost function value for each edge in the candidate edge set. The folding cost function takes into account the difference in normal vectors between the two vertices of the edge and the area weight of the face to which the edge belongs. The difference in normal vectors is represented by the cosine similarity of the normal vectors of the two vertices. The area weight is allocated according to the area ratio of the face to which the edge belongs.

[0148] The folding cost function comprehensively considers the difference in normal vectors between the two vertices of an edge and the area weight of the face to which the edge belongs. The difference in normal vectors can be characterized by calculating the cosine similarity between the normal vectors of the two vertices; the closer the cosine similarity is to 1, the more similar the normal vectors of the two vertices are; the closer the cosine similarity is to -1, the more opposite the normal vectors of the two vertices are. The area weight is allocated according to the area proportion of the face to which the edge belongs; the larger the area of ​​the face, the greater its corresponding area weight. In practice, for each edge in the candidate edge set, first, the cosine similarity of the normal vectors of the two vertices of the edge is calculated; then, the area weight is allocated according to the area proportion of the face to which the edge belongs; finally, the folding cost function value of each edge is calculated according to the formula for calculating the folding cost function.

[0149] Step S4554: Sort the candidate edge set in ascending order of fold cost function values, and calculate the comprehensive optimization objective by weighted summation of the fold cost function values.

[0150] The candidate edge set is sorted in ascending order of fold cost function values. The overall optimization objective is calculated by weighted summation of the fold cost function values. In practice, the fold cost function values ​​of each edge in the candidate edge set are compared and then sorted in ascending order. The purpose of this sorting is to prioritize edges with smaller fold cost function values ​​for folding operations, thus reducing the number of patches while preserving the mesh shape and surface features as much as possible.

[0151] Step S4555: Select the candidate edge that ranks first in the sort and perform the edge folding operation.

[0152] The candidate edge that ranks first in the sorting is selected for edge folding, which shrinks the edge into a single vertex while updating the mesh's topology and vertex information. In practice, according to the edge folding algorithm, the two vertices of the candidate edge that ranks first are merged into one vertex, and then the information of the faces and edges adjacent to that edge is updated to complete the edge folding operation.

[0153] Step S456: During the iterative simplification process, the volume deviation of the simplified mesh is calculated in real time, and the simplification quality is evaluated by comparing the volume difference of the mesh before and after simplification.

[0154] During the iterative simplification process, the volume deviation of the simplified mesh is calculated in real time, and the simplification quality is evaluated by comparing the volume difference between the meshes before and after simplification. Volume deviation refers to the difference between the volume of the simplified mesh and the volume of the mesh before simplification. In practice, a volume calculation algorithm is used to calculate the volumes of the meshes before and after simplification, and then the difference between the two is calculated to obtain the volume deviation. The smaller the volume deviation, the less impact the simplification process has on the mesh volume, and the higher the simplification quality.

[0155] Step S457: When the volume deviation exceeds the preset deviation condition, terminate the current edge folding operation and select the next edge for processing; when the target number of facets is reached, stop the iterative simplification and obtain the simplified surface mesh.

[0156] The preset deviation condition is a threshold used to control volume deviation. When the volume deviation exceeds the preset deviation condition, it indicates that the current edge folding operation has too great an impact on the mesh volume, and the current edge folding operation needs to be terminated and the next edge selected for processing. When the target number of faces is reached, it means that the mesh simplification goal has been achieved, and iterative simplification stops, resulting in the final simplified surface mesh. In practice, after each edge folding operation, the volume deviation of the simplified mesh is calculated in real time and compared with the preset deviation condition. If the volume deviation exceeds the preset deviation condition, the current edge folding operation is terminated, and the next edge is selected from the candidate edge set for processing; if the target number of faces is reached, iterative simplification stops, and the simplified surface mesh is output.

[0157] Step S460: Perform topology repair on the simplified surface mesh by adding or deleting mesh edges to repair non-manifold connections and eliminate non-manifold structures in the mesh.

[0158] Non-manifold connections refer to connections that do not conform to the rules of manifold topology. Examples include multiple edges sharing a vertex without forming normal face connections, or edge connections that result in abnormal local topological structures in the mesh. Such non-manifold structures can affect the subsequent analysis and application of the 3D structural model of the affected area, and therefore require repair.

[0159] In practice, the simplified surface mesh is first subjected to a topological check to identify locations with non-manifold connections. This can be done by traversing all vertices, edges, and faces in the mesh and checking whether their connections conform to manifold topology rules. For example, for each vertex, the number and manner of its connected edges and faces are checked; if any anomalies are found, they are marked as non-manifold connections.

[0160] For detected non-manifold connections, repairs are performed by adding or deleting mesh edges, depending on the specific situation. If the non-manifold connection is due to missing edges, adding appropriate edges can improve the connection and make the local topology conform to manifold rules. For example, if some adjacent faces lack connecting edges, adding connecting edges can connect these faces into a normal whole. If the non-manifold connection is due to redundant edges, these redundant edges are deleted. For example, if there are too many edge connections at some vertices, causing a chaotic topology, deleting some unnecessary edges can restore a normal topology.

[0161] During the process of adding or deleting edges, it is necessary to ensure that no new non-manifold structures are introduced. After each operation, the topology of the mesh must be re-examined to ensure that the repaired mesh conforms to manifold topology rules both locally and globally. Through such repair operations, non-manifold structures in the simplified surface mesh are eliminated, resulting in a surface mesh with a good topology, providing a reliable foundation for subsequent geometric feature quantization and model construction.

[0162] Step S470: Perform geometric feature quantization on the repaired surface mesh, calculate the geometric parameters of the lesion area using a three-dimensional morphological measurement algorithm, and extract the volume features, surface area features, and spatial location features of the lesion area.

[0163] 3D morphological measurement algorithms are a set of algorithms used to quantify the geometric features of 3D objects. They can extract various geometric parameters of the lesion region from the repaired surface mesh. Volumetric features refer to the size of the lesion region in 3D space, reflecting the scale of the lesion. Surface area features are the total surface area of ​​the lesion region, reflecting the surface complexity of the lesion region. Spatial location features describe the specific location of the lesion region in 3D space.

[0164] In practice, a mesh-based volumetric calculation method can be used to calculate volumetric features. The repaired surface mesh is viewed as a closed volume composed of multiple triangular facets. By processing the information of each triangular facet and its corresponding vertex, the volume contained within the closed volume can be calculated. For example, the divergence theorem can be used to transform the volume calculation into an integral operation on the surface; the volume of the lesion region is obtained by summing the integrals over all triangular facets.

[0165] To calculate the surface area, simply add up the areas of all triangular facets in the repaired surface mesh. The area of ​​each triangular facet can be calculated using the cross product of its three vertices. By traversing all triangular facets and summing their areas, the surface area of ​​the lesion region can be obtained.

[0166] For the extraction of spatial location features, a predefined coordinate system is typically used as a reference. The centroid coordinates of the lesion region can be calculated; the centroid is the geometric center of the lesion region in three-dimensional space. The centroid coordinates are obtained by weighted averaging the coordinates of all vertices in the repaired surface mesh, and these coordinates are one of the spatial location features of the lesion region. The positional relationship of the lesion region relative to a fixed point or reference plane can also be calculated to further refine its spatial location.

[0167] These three-dimensional morphological measurement algorithms are used to comprehensively quantify the geometric features of the repaired surface mesh, accurately extract the volume features, surface area features, and spatial location features of the lesion area, and provide important parameter support for the subsequent construction of a three-dimensional structural model of the lesion area.

[0168] Step S480: Construct a three-dimensional structural model of the lesion area based on volume features, surface area features, and spatial location features.

[0169] A three-dimensional structural model of a lesion area is a complete description of the lesion area in three-dimensional space. It integrates important information such as volume features, surface area features, and spatial location features, and can intuitively display the shape, size, and location of the lesion.

[0170] In practice, the approximate size of the lesion region is first determined based on the extracted volumetric features. These volumetric features reflect the three-dimensional space occupied by the lesion region, and a preliminary three-dimensional framework matching the volume of the lesion region can be constructed based on this. For example, if the volumetric features indicate a large lesion region, then the size of the constructed three-dimensional framework will also be correspondingly large.

[0171] Then, surface area features are used to refine the surface morphology of the lesion area. Surface area features reflect the complexity of the lesion area's surface. Based on the size and distribution of the surface area, the surface of the 3D framework can be adjusted and refined. If the surface area is large and complex in distribution, it indicates that the surface of the lesion area may have many wrinkles and unevenness, and more details need to be added to simulate these surface features when building the model.

[0172] Finally, the specific location of the lesion region in three-dimensional space is determined by combining spatial location features. Spatial location features describe the position of the lesion region relative to a certain reference coordinate system. The constructed model with volume and surface features is placed at the corresponding spatial location to complete the construction of the three-dimensional structural model of the lesion region.

[0173] The completed three-dimensional structural model of the lesion area can be presented in a visual form. Doctors can observe the model to intuitively understand the shape, size, and location of the lesion area, providing strong support for developing surgical plans and assessing the condition. At the same time, the model can also serve as the basis for subsequent analysis and research, such as conducting simulated surgeries and studying the growth and development of lesions.

[0174] Step S500: Generate surgical path planning data based on the three-dimensional structural model of the lesion area.

[0175] Surgical path planning data is a series of data generated to guide surgical procedures. Based on a three-dimensional structural model of the lesion area, it comprehensively considers the location, size, shape of the lesion, and the condition of surrounding tissues to plan a safe and effective surgical path.

[0176] In practice, the first step is to obtain detailed information about the lesion area from a three-dimensional structural model, including its spatial location, volume, surface area, and relative relationship with surrounding tissues. This information is crucial for determining the surgical incision point and the path to reach the lesion area.

[0177] Then, taking into account the patient's physiological structure and the specific requirements of the surgery, factors such as the operating range and flexibility of the surgical instruments are considered. For example, different surgical instruments have different lengths, curvatures, and operating angles, and the surgical path needs to be planned according to these characteristics to ensure that the surgical instruments can smoothly reach the lesion area and perform effective operations.

[0178] Simultaneously, it is crucial to avoid important nerves, blood vessels, and other critical tissues. Analyzing the patient's medical imaging data helps identify surrounding structures and minimize conflicts with them when planning the surgical path. Computer simulation technology can be used to model the surgical path in three-dimensional space, assessing its distance from and potential impact on surrounding vital tissues.

[0179] When planning a surgical approach, the safety and feasibility of the procedure must also be considered. The potential risks associated with the surgical approach, such as bleeding and infection, should be assessed, and appropriate optimization measures should be taken. For example, if a particular approach might cross numerous blood vessels, the approach may be adjusted to enter from a relatively safer area.

[0180] By comprehensively considering the above factors, a series of surgical path planning data is generated. This data may include the coordinates of the surgical incision point, the direction and length of the path, and information on important tissues to be avoided. Surgeons can use this data to make detailed plans and preparations before surgery, and operate according to the planned path during the procedure, thereby improving the success rate and safety of the surgery.

[0181] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention, such as Gaussian pyramid decomposition algorithm, region growing algorithm, edge folding algorithm, etc., can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solution of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set the threshold based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select the activation function, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0182] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0183] In one embodiment, the processor 101 executes the deep learning-based neuroimaging analysis method provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A deep learning-based neuroimaging analysis method, characterized in that, include: Receive raw neuroimaging data; Multimodal feature co-extraction is performed on the raw neuroimaging data to generate an initial feature map; The initial feature map is input into a deep feature fusion network to enhance and identify the lesion region, thereby obtaining the lesion region feature map. A three-dimensional topological structure analysis is performed on the feature map of the lesion area to construct a three-dimensional structural model of the lesion area.

2. The method as described in claim 1, characterized in that, The step of performing multimodal feature co-extraction on the raw neuroimaging data to generate an initial feature map includes: Spatial resolution standardization is performed on the magnetic resonance imaging data in the original neuroimaging data. By resampling the images, the magnetic resonance images under different acquisition conditions are unified to a preset spatial size to obtain standardized magnetic resonance images. The grayscale characteristics of the computed tomography image data in the original neuroimaging data are normalized, and the grayscale distribution of the tomography image is mapped to a preset distribution range through dynamic range compression to obtain a normalized tomography image. The standardized magnetic resonance image and the normalized tomographic image are spatially registered between modes. The spatial transformation relationship between the two modal images is optimized based on mutual information to generate a registered multimodal image. The registered multimodal image is subjected to multi-scale feature decomposition, and modal features at different scale levels are generated by Gaussian pyramid decomposition algorithm to obtain a multi-scale modal feature set. Cross-attention fusion is performed on the magnetic resonance modal features and tomographic scan modal features in the multi-scale modal feature set. Dynamic feature weight allocation is completed by calculating the inter-modal feature correlation matrix to generate cross-modal fusion features. The cross-modal fusion features are subjected to noise suppression, and the artifact interference in the fusion features is eliminated by a nonlocal similarity filtering algorithm to obtain denoised fusion features; The denoised fusion features are regularized in terms of feature dimensions, and the multi-scale features are unified into the same dimensional space through a feature mapping function to generate an initial feature map.

3. The method as described in claim 1, characterized in that, The step of inputting the initial feature map into a deep feature fusion network for lesion region enhancement and recognition to obtain a lesion region feature map includes: The initial feature map is input into the low-level feature encoding module of the deep feature fusion network for primary feature extraction to obtain the low-level encoded features. The low-level feature encoding module contains multiple convolution operation units connected in sequence. Each convolution operation unit consists of a convolutional layer and a batch normalization layer. The low-level encoded features are input into the mid-level feature encoding module for semantic feature enhancement to obtain mid-level encoded features. The mid-level feature encoding module completes feature reuse and gradient transfer through a residual connection structure. The middle-level coding features are input into the high-level feature coding module for global feature integration to obtain high-level coding features. The high-level feature coding module extracts cross-region feature association information through global context aggregation operation. The low-level coding features, middle-level coding features and high-level coding features are fused using a feature pyramid. After upsampling, the high-level features and low-level features are scale-aligned and superimposed to generate multi-scale fused features. The multi-scale fusion features are input into the spatial attention module for region attention allocation. The feature map is then adjusted at the pixel level by generating a two-dimensional spatial attention map to obtain spatially weighted features. The spatially weighted features are input into the channel attention module for feature channel importance filtering, and the channel weights are calculated through the feature channel importance evaluation mechanism to obtain the channel weighted features; The channel-weighted features are input into the lesion region classification module for pixel-level classification. The lesion region classification module uses a convolution classifier to predict the lesion probability of each pixel and obtain the initial lesion region features. The initial lesion region features are refined at the edges, and the boundary contour of the lesion region is corrected through morphological optimization operations to generate a lesion region feature map.

4. The method as described in claim 1, characterized in that, The step of performing three-dimensional topological structure analysis on the feature map of the lesion region to construct a three-dimensional structural model of the lesion region includes: Inter-layer spatial position calibration is performed on the feature map of the lesion area. Spatial alignment is completed by calculating the geometric transformation parameters of adjacent feature maps to obtain a calibrated feature map sequence. The calibrated feature map sequence is voxelized to map the two-dimensional feature map sequence to a three-dimensional voxel grid space, generating three-dimensional voxel data. Connected region extraction is performed on the three-dimensional voxel data, and the lesion region is segmented from the three-dimensional voxel data using a region growing algorithm to obtain a voxel subset of the lesion region; A surface mesh is constructed on a subset of voxels in the lesion region. A triangular mesh is generated from the voxel data using the moving cube algorithm to generate an initial surface mesh model. The initial surface mesh model is simplified by reducing the number of faces in the mesh through an edge folding algorithm to obtain a simplified surface mesh. The simplified surface mesh is topologically repaired by adding or deleting mesh edges to repair non-manifold connections and eliminate non-manifold structures in the mesh. The geometric features of the repaired surface mesh are quantized, and the geometric parameters of the lesion area are calculated by a three-dimensional morphological measurement algorithm. The volume features, surface area features and spatial location features of the lesion area are extracted. A three-dimensional structural model of the lesion area is constructed based on the volume characteristics, surface area characteristics, and spatial location characteristics.

5. The method as described in claim 2, characterized in that, The registered multimodal image undergoes multi-scale feature decomposition, generating modal features at different scale levels using a Gaussian pyramid decomposition algorithm, resulting in a multi-scale modal feature set, including: The first set of filtering parameters with different bandwidths is used to perform Gaussian filtering on the registered multimodal image. Multi-scale coverage is completed by adjusting the filtering range in sequence, and the first-scale modal features are generated. A second set of filtering parameters with different bandwidths is set to perform Gaussian filtering on the registered multimodal images. Multi-scale coverage is completed by adjusting the filtering range in sequence, and second-scale modal features are generated. A third set of filtering parameters with different bandwidths is set to perform Gaussian filtering on the registered multimodal images. Multi-scale coverage is completed by adjusting the filtering range in sequence, and third-scale modal features are generated. A Laplace transform is performed on the first-scale modal features, and edge features are extracted by calculating the second derivative of the Gaussian filtered image to obtain the first-scale edge features. A Laplace transform is performed on the second-scale modal features, and edge features are extracted by calculating the second derivative of the Gaussian filtered image to obtain the second-scale edge features. A Laplace transform is performed on the third-scale modal features, and edge features are extracted by calculating the second derivative of the Gaussian filtered image to obtain the third-scale edge features. The first-scale modal features, second-scale modal features, third-scale modal features, first-scale edge features, second-scale edge features, and third-scale edge features are stacked. Modal features and edge features of different scales are then concatenated along the channel dimension to generate a multi-scale modal feature set.

6. The method as described in claim 2, characterized in that, The process of performing cross-attention fusion on the magnetic resonance modal features and tomographic modal features in the multi-scale modal feature set, and generating cross-modal fused features by calculating the inter-modal feature correlation matrix to complete dynamic feature weight allocation, includes: Separate the magnetic resonance modal feature subset and the tomographic scan modal feature subset from the multi-scale modal feature set; The magnetic resonance modal feature subset is flattened to convert the three-dimensional feature tensor into a two-dimensional feature matrix, thus obtaining the flattened magnetic resonance features. The feature subset of the tomographic scan modal features is flattened to convert the three-dimensional feature tensor into a two-dimensional feature matrix, thus obtaining the tomographic scan flattened features. The correlation matrix between the magnetic resonance flattening features and the tomographic scan flattening features is calculated by feature similarity analysis. Each element of the correlation matrix represents the degree of similarity between the magnetic resonance modal features and the tomographic scan modal features. The correlation matrix is ​​normalized in the row direction, and the weight distribution is adjusted by the normalization function to obtain the magnetic resonance attention weights. The column direction of the correlation matrix is ​​normalized, and the weight distribution is adjusted by a normalization function to obtain the tomographic scan attention weights. The magnetic resonance attention weights and the tomographic flattening features are multiplied by a matrix to obtain the magnetic resonance-guided tomographic features. The tomographic scan attention weights and magnetic resonance flattening features are multiplied by a matrix to obtain the tomographic scan-guided magnetic resonance features. The magnetic resonance-guided tomography features are added element-wise to the tomography-guided magnetic resonance features to obtain the fused flattened features. The fused flattened features are reshaped to restore the two-dimensional fused flattened features to a three-dimensional feature tensor structure, thereby generating cross-modal fused features.

7. The method as described in claim 3, characterized in that, The feature pyramid fusion of the bottom-level, middle-level, and high-level coding features involves upsampling the high-level and low-level features to align their scales and then superimpose them to generate multi-scale fused features. The high-level coded features are upsampled for the first time, and the feature scale is enlarged by an interpolation algorithm to obtain the first upsampled features; The first upsampled feature is aligned with the middle-layer encoded feature. Feature matching is completed by adjusting the spatial size and position of the feature to obtain the middle-layer aligned feature. The mid-layer alignment feature is added element-wise to the first upsampled feature to obtain the mid-layer fusion feature; The mid-layer fusion features are upsampled a second time, and the feature scale is enlarged by an interpolation algorithm to obtain the second upsampled features; The second upsampled feature is aligned with the bottom-level encoded feature. Feature matching is completed by adjusting the spatial size and position of the feature to obtain the bottom-level aligned feature. The bottom alignment feature is added element-wise to the second upsampled feature to obtain the bottom fusion feature; The underlying fusion features are convolutionally integrated, and the channel dimension of the fusion features is adjusted by the convolution kernel to generate multi-scale fusion features.

8. The method as described in claim 3, characterized in that, The step of inputting the spatially weighted features into the channel attention module for feature channel importance filtering, and calculating channel weights through a feature channel importance evaluation mechanism to obtain channel-weighted features includes: Global feature aggregation is performed on the spatially weighted features, and statistical aggregation is performed on the spatial dimension of each feature channel to obtain the channel description vector; The channel description vector is first nonlinearly mapped to a low-dimensional space to obtain an intermediate feature vector; the intermediate feature vector is then activated to obtain an activated feature vector. A second nonlinear mapping is performed on the activation feature vector to map the activation feature vector back to the original channel dimension, resulting in the channel weight vector. The channel weight vector is normalized to adjust the channel weights to a preset distribution range, thus obtaining normalized channel weights. The normalized channel weights and spatial weighted features are multiplied by channel dimension weighting, and the feature map of each channel is multiplied by the corresponding channel weight to obtain the channel weighted features.

9. The method as described in claim 4, characterized in that, The step of voxelizing the calibrated feature map sequence to map the two-dimensional feature map sequence to a three-dimensional voxel grid space to generate three-dimensional voxel data includes: Determine the interlayer spacing parameter and plane resolution parameter of each feature map in the calibrated feature map sequence; A three-dimensional voxel grid coordinate system is constructed based on the interlayer spacing parameter and the planar resolution parameter, and the origin of the three-dimensional voxel grid coordinate system is set at the reference position of the starting slice of the feature map sequence. Each feature map in the calibrated feature map sequence is mapped to the corresponding slice position in the three-dimensional voxel grid coordinate system; Interpolation calculations are performed on the voxel values ​​between adjacent slices, and a trilinear interpolation algorithm is used to achieve a smooth transition of voxel values, filling the blank voxels in the three-dimensional voxel mesh; Thresholding is performed on the interpolated three-dimensional voxel grid, and the segmentation threshold is determined based on the feature value distribution of the lesion area to obtain binary three-dimensional voxel data; Morphological closure is performed on the binarized three-dimensional voxel data, and morphological repair is performed on the voxel data using spherical structural elements to fill the hole areas in the voxel data. The processed binary 3D voxel data is converted into 3D voxel data.

10. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the deep learning-based neuroimaging analysis method as described in any one of claims 1-9.

Images