Medical image texture analysis method, system and equipment based on artificial intelligence and medium

Abstract

The invention relates to the technical field of image analysis and processing, provides a medical image texture analysis method, system and device based on artificial intelligence and a medium, and solves the problems that texture features are mixed, generalization is insufficient, and texture evolution and internal structure modeling are insufficient. The method comprises the steps that a to-be-analyzed medical image and a standard medical image are acquired, and a texture residual image is generated through differential operation; arranging to form an image sequence and inputting the image sequence to a time sequence encoder to generate a first feature tensor; the first feature tensor is decoupled, and a first texture feature and a second texture feature are separated; recombining the first texture feature and the second texture feature to generate a recombined feature; constructing a graph structure according to the recombined features, and carrying out message passing and aggregation to extract a topology embedding vector; and inputting the topology embedded vector into a preset classifier, and outputting a texture classification result. According to the technical scheme provided by the invention, the universality of the analysis method among different devices and the comparability of results are improved, and the distinction degree of atypical texture modes is also improved.

Description

Technical Field

[0001] This application relates to the field of image analysis and processing technology, and in particular to a method, system, device and medium for medical image texture analysis based on artificial intelligence. Background Technology

[0002] With the development of precision medicine, the role of medical image analysis in disease diagnosis and assessment has become increasingly prominent, especially the need to decode the heterogeneity of lesion texture.

[0003] The current approach is an end-to-end model based on deep convolutional neural networks. This model inputs images into the network, extracts features through convolutional layers, and outputs classification results, avoiding the limitations of manual feature design and improving the automation of texture recognition.

[0004] However, this method still has shortcomings. The features it learns are usually the result of coupling multiple factors such as imaging parameters, equipment differences and lesion biological characteristics, which are difficult to separate. This leads to insufficient feature stability. When faced with image data from different sources, the model performance is prone to decline. At the same time, this method fails to explicitly model the dynamic changes of texture in the sequence and lacks effective characterization of the internal topological relationships of texture. Summary of the Invention

[0005] This application provides a medical image texture analysis method, system, device, and medium based on artificial intelligence, to solve the problems of mixed texture features, insufficient generalization, and inadequate modeling of texture evolution and internal structure in the prior art.

[0006] Firstly, this application provides an artificial intelligence-based method for medical image texture analysis, including:

[0007] Acquire a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed;

[0008] The medical image to be analyzed is compared with each image in the standard medical image to generate a set of texture residual images.

[0009] The set of texture residual images are arranged in a preset subtype order to form an image sequence. The image sequence is then input into a temporal encoder to determine the dependency and evolution rules between different residual images in the image sequence and generate a first feature tensor. The first feature tensor is used to characterize the temporal features of the image sequence.

[0010] The first feature tensor is decoupled to separate the first texture feature and the second texture feature, wherein the first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion itself.

[0011] Within a preset potential space, the first texture feature and the second texture feature are recombined to generate a recombined feature;

[0012] Based on the recombined features, a graph structure is constructed, and a hierarchical graph neural network is used to perform message passing and aggregation on the graph structure to extract a topological embedding vector. The topological embedding vector is used to characterize the internal texture structure and relationships of the lesion region of the medical image to be analyzed.

[0013] The topological embedding vector is input into a preset classifier to output the texture classification result for the medical image to be analyzed.

[0014] Preferably, the set of texture residual images is arranged according to a preset subtype order to form an image sequence, including:

[0015] Obtain the subtype identifier of the standard medical image associated with the texture residual image, wherein the subtype identifier is used to indicate the specific category to which the standard medical image belongs;

[0016] According to the preset subtype order, the arrangement position value corresponding to each subtype identifier is determined, wherein the preset subtype order defines the order between different subtype identifiers;

[0017] Based on the arrangement position values, the texture residual images are reorganized so that the texture residual images are arranged in ascending order of the arrangement position values ​​to form an image sequence.

[0018] Preferably, the image sequence is input to a temporal encoder, which determines the dependencies and evolution patterns between different residual images in the image sequence, generating a first feature tensor, including:

[0019] Each frame of the texture residual image in the image sequence is input into the processing unit of the temporal encoder, and the current input texture residual image is used to extract features to obtain the current image features.

[0020] The current image features are fused with the initial internal state vector of the processing unit to update the internal state vector;

[0021] The updated internal state vector is combined with the position indication information corresponding to the texture residual image to be processed in the next step to generate intermediate features;

[0022] Based on the intermediate features, the correlation strength distribution between the currently processed texture residual image and all other texture residual images in the image sequence is calculated;

[0023] Based on the correlation strength distribution, the image features corresponding to all texture residual images in the image sequence are weighted and combined to generate the context features of the current frame;

[0024] The current image features are aggregated with the context features of the current frame to output the fused features of the current frame, and the fused features of all frames are collected to form the first feature tensor.

[0025] Preferably, the first feature tensor is decoupled to separate a first texture feature and a second texture feature, wherein the first texture feature is used to characterize device and protocol information originating from the imaging process, and the second texture feature is used to characterize biological characteristic information originating from the lesion, including:

[0026] The first feature tensor is simultaneously input into the pre-trained first feature extraction unit and the pre-trained second feature extraction unit;

[0027] The first feature tensor is processed by the first feature extraction unit to obtain an initial first feature, and the first feature tensor is processed by the second feature extraction unit to obtain an initial second feature.

[0028] Calculate the matching degree between the initial first feature and the known imaging pattern feature, and calculate the statistical independence measure between the initial second feature and the initial first feature, wherein the known imaging pattern feature is used to represent different medical imaging devices and imaging protocols;

[0029] Based on the matching degree and the statistical independence measure, an optimization objective function is constructed;

[0030] By iteratively adjusting the processing parameters of the first feature extraction unit and the second feature extraction unit, the value of the optimization objective function is increased;

[0031] When the optimization objective function increases to meet the preset optimization conditions, the first texture feature is output by the first feature extraction unit, and the second texture feature is output by the second feature extraction unit.

[0032] Preferably, within a preset latent space, the first texture feature and the second texture feature are recombined to generate a recombined feature, including:

[0033] The first texture feature and the second texture feature are respectively mapped to the preset latent space to obtain the first latent feature representation and the second latent feature representation;

[0034] Based on the second texture feature, calculate the reconstructed weight vector;

[0035] Using the recombined weight vector, a weighted fusion operation is performed on the first latent feature representation and the second latent feature representation to generate a combined latent space representation;

[0036] The combined latent space representation is mapped to output features with different dimensions than the latent space, which are then used as the recombined features.

[0037] Preferably, based on the recombined features, a graph structure is constructed, and a hierarchical graph neural network is used to perform message passing and aggregation on the graph structure to extract topological embedding vectors, including:

[0038] From the recombined features, feature representations of multiple lesion sub-regions are extracted;

[0039] Each lesion sub-region is used as a node. The connection strength between each node is calculated based on the degree of feature correlation and positional correlation between each lesion sub-region. The connection strength is then assigned to the edges between the corresponding nodes to construct the initial graph structure.

[0040] Obtain the first feature information of other nodes connected to any node in the initial graph structure, combine the first feature information with the first feature information of the node itself, update the first feature information of the node, and generate the updated graph structure.

[0041] In the updated graph structure, multiple nodes that meet the preset merging conditions in space are aggregated into a target node, and aggregation feature information is generated for the target node. Then, a high-level graph structure is constructed based on the target node and the aggregation feature information corresponding to the target node.

[0042] The aggregated feature information of other target nodes connected to any target node in the high-level graph structure is obtained, and the aggregated feature information is combined with the aggregated feature information of the target node itself to update the aggregated feature information of the target node.

[0043] The aggregated feature information of all updated target nodes is integrated to generate a topological embedding vector.

[0044] Preferably, the topological embedding vector is input into a preset classifier to output the texture classification result of the medical image to be analyzed, including:

[0045] The topology embedding vector is input into the first data processing layer of the preset classifier;

[0046] In the first data processing layer, a first linear transformation operation is performed on the topological embedding vector, and the result of the first linear transformation operation is input to the first nonlinear processing unit for processing to generate a first intermediate feature.

[0047] The first intermediate feature is input to the second data processing layer of the preset classifier. In the second data processing layer, a second linear transformation operation is performed on the first intermediate feature, and the result of the second linear transformation operation is input to the second nonlinear processing unit for processing to generate the second intermediate feature.

[0048] The second intermediate feature is input to the output layer of the preset classifier. In the output layer, a third linear transformation operation is performed on the second intermediate feature to generate a classification score vector corresponding to a plurality of preset candidate texture categories. Each element value in the classification score vector corresponds to the confidence level of a candidate texture category.

[0049] The target element value with the largest value is determined from the classification score vector, and the candidate texture category corresponding to the target element value is output as the texture classification result for the medical image to be analyzed.

[0050] Secondly, this application provides an artificial intelligence-based medical image texture analysis system, comprising:

[0051] The acquisition module is used to acquire a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed;

[0052] The calculation module is used to perform a difference operation between the medical image to be analyzed and each image in the standard medical image to generate a set of texture residual images;

[0053] The determination module is used to arrange the set of texture residual images according to a preset subtype order to form an image sequence, and input the image sequence to a temporal encoder to determine the dependency relationship and evolution law between different residual images in the image sequence, and generate a first feature tensor, which is used to characterize the temporal features of the image sequence.

[0054] The separation module is used to decouple the first feature tensor to separate the first texture feature and the second texture feature, wherein the first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion itself.

[0055] The recombination module is used to recombine the first texture feature and the second texture feature within a preset potential space to generate a recombined feature;

[0056] The construction module is used to construct a graph structure based on the recombined features, and to use a hierarchical graph neural network to perform message passing and aggregation on the graph structure to extract a topological embedding vector. The topological embedding vector is used to characterize the internal texture structure and relationships of the lesion region of the medical image to be analyzed.

[0057] The output module is used to input the topological embedding vector into a preset classifier to output the texture classification result for the medical image to be analyzed.

[0058] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement an artificial intelligence-based medical image texture analysis method as described in the first aspect above.

[0059] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements an artificial intelligence-based medical image texture analysis method as described in the first aspect.

[0060] The beneficial effects of this application are:

[0061] This application utilizes a full-chain processing approach, including constructing texture residual image sequences, temporal coding, feature decoupling and recombination, graph structure construction, and hierarchical aggregation, to systematically decode the heterogeneous textures of lesion regions in medical images. This method effectively integrates the dynamic evolution patterns and static topological structure information of textures, and explicitly separates imaging interference from biological characteristics, thereby extracting lesion feature expressions with stronger discriminative power and better interpretability. This makes the texture classification results based on the topological embedding vector more accurate and robust, providing a more reliable quantitative basis for assisted diagnosis.

[0062] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 A flowchart of an artificial intelligence-based medical image texture analysis method provided in this application is shown;

[0065] Figure 2A schematic diagram of the structure of a medical image texture analysis system based on artificial intelligence provided in this application is shown;

[0066] Figure 3 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0067] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0068] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0069] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] Figure 1 A flowchart of an artificial intelligence-based medical image texture analysis method is provided for this application, such as... Figure 1 As shown, the method includes:

[0071] Figure 1 A flowchart of an artificial intelligence-based medical image texture analysis method is provided for this application, such as... Figure 1 As shown, the method includes:

[0072] Step 101: Obtain a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed.

[0073] In this step, the medical images to be analyzed refer to digital images containing suspected lesion areas obtained after scanning specific parts of a patient using medical imaging equipment such as CT, MRI, or PET. These images are used as the core analysis object to extract texture features and classify the images. They are usually obtained directly from the hospital database through an image archiving and communication system.

[0074] Standard medical images refer to a set of reference images that have undergone standardized image processing (such as grayscale normalization and resolution unification) and have clear pathological diagnostic labels. The lesion contours are aligned with the images to be analyzed through image registration and segmentation techniques for subsequent texture difference analysis. In practice, they can be obtained from public medical image databases (such as TCIA, LIDC-IDRI) or institutional standardized image databases, and their lesion contours can be automatically segmented and confirmed by algorithms.

[0075] In this step, the patient's specific anatomical region is first scanned using medical imaging equipment to collect raw data, which is then processed by an image reconstruction algorithm to generate a medical image containing the target region to be analyzed.

[0076] Secondly, from the pre-built standard image database, a set of relevant candidate standard medical images are retrieved based on conditions such as anatomical location and imaging modality. Then, through the image registration algorithm, each candidate standard medical image is precisely aligned with the medical image to be analyzed in space to ensure that the two depict the same anatomical structure.

[0077] Then, using image segmentation algorithms, such as the U-Net network based on deep learning or region growing algorithms, the precise contours of lesion regions are automatically identified and delineated on the medical images to be analyzed. Finally, the lesion contours obtained in the above steps on the medical images to be analyzed are applied to each candidate standard medical image that has been spatially aligned through contour mapping or the same segmentation algorithm, thereby extracting image blocks corresponding to the same contour region in each standard image. These image blocks are then integrated to form a standard medical image.

[0078] For example, the following is a specific embodiment:

[0079] A radiologist at Hospital A needs to perform texture analysis on a lung nodule in patient B to assess its nature. First, the CT scan image of patient B is retrieved, which is the medical image to be analyzed. Then, from Hospital A's image database, based on tags such as lung, CT plain scan, and lung nodule, 1,000 historical patients' chest CT images are automatically retrieved as initial candidates.

[0080] Then, an elastic registration algorithm based on mutual information is used to spatially align these 1,000 candidate images with the CT images of patient B one by one, so that their anatomical structures match. After alignment, a pre-trained U-Net neural network model is called to automatically segment the precise boundaries of the lung nodules on the CT images of patient B.

[0081] Then, using the boundary of the patient's B lung nodule as a template, the same U-Net model was used to automatically search and segment the nodule regions with the most similar shape and location on 999 other aligned CT images. Finally, 200 images were selected from these images whose nodule contours and template contours had a shape similarity exceeding a preset threshold. The nodule region image blocks of these 200 images constituted standard medical images.

[0082] Step 102: Perform a difference operation between the medical image to be analyzed and each image in the standard medical image to generate a set of texture residual images.

[0083] In this step, the texture residual image refers to a set of two-dimensional images that highlight the subtle texture differences between the two by subtracting the pixel values ​​of the corresponding area of ​​the standard medical image from the pixel values ​​of the lesion area of ​​the medical image to be analyzed. It is used to amplify and capture the specific patterns of the lesion to be analyzed in grayscale or signal intensity distribution relative to the standard reference image. It is obtained by pixel-by-pixel difference operation and standardization of the results.

[0084] In this step, the acquired medical image to be analyzed and its lesion outline are first read, along with a set of standard medical images with the same lesion outline. Then, the following operations are performed on each standard medical image:

[0085] First, using the array operation function in the image processing library, the intensity value of each pixel in the lesion area of ​​the medical image to be analyzed is subtracted from the intensity value of the pixel at the same spatial position in the corresponding area of ​​the current standard medical image to obtain an initial pixel value difference map.

[0086] Then, all pixel values ​​in this initial difference image are standardized using the Z-score standardization method. First, the mean and standard deviation of all pixel values ​​in the entire difference image are calculated. Then, each pixel value is subtracted from the mean and divided by the standard deviation to transform the difference values ​​to a distribution centered at zero with a unit standard deviation.

[0087] Finally, the standardized image is output as a texture residual image, and the above difference and standardization operations are repeated for each image in the standard medical image group. When all standard images have been processed, a set of texture residual images is obtained.

[0088] For example, following the embodiment of step 101, after acquiring the CT image of the lung nodule of patient B and 200 standard medical images with the same lesion outline, the first standard medical image is processed. The nodule region image block of patient B and the nodule region image block of the first standard image are loaded from memory simultaneously. A preset difference calculation function is called. The difference calculation function traverses each pixel in the image block. For example, the pixel located in the i-th row and j-th column is subtracted from the CT value of the standard image at the same position (e.g., -110 HU) by the CT value of patient B's image at that position (e.g., -120 HU) to obtain the initial residual value (-10 HU) of that pixel. After traversing all pixels, an initial difference map is generated.

[0089] Next, the mean and standard deviation of all pixel values ​​in this initial difference image are calculated. Assuming the mean is 5 and the standard deviation is 15, each pixel value in the image is then transformed. For example, the pixel value of -10 HU is transformed into (-10-5) / 15 = -1. This transformation process is applied to all pixels, and the first texture residual image is finally obtained. Then, this process is repeated, and the same pixel-level differencing and standardization operations are performed on the remaining 199 standard medical images in turn, finally generating a set of 200 texture residual images.

[0090] Step 103: Arrange the set of texture residual images according to a preset subtype order to form an image sequence, and input the image sequence into a temporal encoder to determine the dependency relationship and evolution law between different residual images in the image sequence, and generate a first feature tensor. The first feature tensor is used to characterize the temporal features of the image sequence.

[0091] Optionally, step 103 may specifically include:

[0092] Step 1031: Obtain the subtype identifier of the standard medical image associated with the texture residual image, wherein the subtype identifier is used to indicate the specific category to which the standard medical image belongs.

[0093] Step 1032: Determine the arrangement position value corresponding to each subtype identifier according to the preset subtype order, wherein the preset subtype order defines the order between different subtype identifiers.

[0094] Step 1033: Based on the arrangement position value, the texture residual image is reorganized so that the texture residual image is arranged in ascending order according to the arrangement position value to form an image sequence.

[0095] Step 1034: Input each frame of texture residual image in the image sequence into the processing unit of the temporal encoder to extract features from the currently input texture residual image and obtain the current image features.

[0096] Step 1035: The current image features are fused with the initial internal state vector of the processing unit to update the internal state vector.

[0097] Step 1036: Combine the updated internal state vector with the position indication information corresponding to the texture residual image to be processed in the next step to generate intermediate features.

[0098] Step 1037: Based on the intermediate features, calculate the correlation strength distribution between the currently processed texture residual image and all other texture residual images in the image sequence.

[0099] Step 1038: Based on the correlation intensity distribution, the image features corresponding to all texture residual images in the image sequence are weighted and combined to generate the context features of the current frame.

[0100] Step 1039: Aggregate the current image features with the context features of the current frame, output the fused features of the current frame, and gather the fused features of all frames to form the first feature tensor.

[0101] In this step, the preset subtype order refers to a predefined list that is driven by data. This list specifies the logical or biological progression order between different disease subtypes or pathological states, and is used to guide the organization of discrete texture residual images into a directional sequence, such as: benign → dysplasia → carcinoma in situ → invasive carcinoma. The order between different subtype identifiers is defined and is determined by statistical analysis of disease development patterns in historical data.

[0102] An image sequence refers to a list of logically ordered images formed by reordering a group of originally independent texture residual images according to a preset subtype order. This sequence is used to simulate or reflect a potential texture evolution path so that its dynamic patterns can be analyzed later.

[0103] A subtype identifier is a label attached to each standard medical image to indicate the specific category to which the standard medical image belongs, such as benign, carcinoma in situ, invasive carcinoma, etc., obtained through the pathological diagnosis report or expert annotation corresponding to the acquisition of the standard image.

[0104] The arrangement position value refers to a numerical sequence number assigned to each specific subtype identifier according to the preset subtype order, indicating its position in that order. It is used to convert non-numeric category labels into numerical values ​​that can be compared in size. The subtype identifier is mapped to the corresponding sequence number by querying the preset subtype order.

[0105] A temporal encoder is a neural network module specifically designed for processing sequential data. It contains processing units that can remember historical information and analyze the interrelationships and dependencies between data at different positions in the sequence. It is used to extract features that characterize the dynamic changes of the entire sequence from the image sequence. This is achieved by constructing a network structure that includes recurrent connections or self-attention mechanisms. In this step, sequence modeling structures such as recurrent neural networks, long short-term memory networks, gated recurrent units, or Transformer encoders can be used to capture the temporal dependencies between images within the sequence. Specifically, the processing unit of the temporal encoder can be an LSTM unit, whose internal state vector is updated in each frame. Combined with position encoding, it calculates inter-frame attention weights and generates a first feature tensor that fuses the temporal context.

[0106] The current image feature refers to a mathematical vector representing the core information of the current frame image obtained by the processing unit of the temporal encoder after performing a preliminary analysis on the texture residual image of the currently input frame. It is used as the basic unit for constructing more complex temporal relationships.

[0107] The initial internal state vector of the processing unit refers to a small mathematical vector, either all zero or randomly initialized, that is preset inside the processing unit before the timing encoder starts processing the first frame of the image sequence. It serves as the starting point for memory, does not carry any historical information, and is obtained through system initialization settings.

[0108] Intermediate features refer to a transitional mathematical vector formed by combining the currently updated internal state vector with information representing the position of subsequent frames in the sequence. This vector is used to calculate the degree of association between the current frame and all other frames in the sequence.

[0109] The correlation strength distribution refers to a list of values, where each value represents the strength of the correlation between the currently processed texture residual image and another specific image in the image sequence. The higher the value, the stronger the correlation, and it is used to determine how much attention should be given to other frames when constructing the context information of the current frame.

[0110] Image features refer to the mathematical vectors obtained after preliminary feature extraction of each frame of texture residual image.

[0111] The contextual features of the current frame refer to a comprehensive mathematical vector obtained by weighting and summing the image features of all frames in the image sequence using the correlation strength distribution as weights. It aggregates information related to other frames in the entire sequence and is used to provide a global perspective for understanding the current frame.

[0112] The fused feature of the current frame refers to a richer mathematical vector generated by combining the current image features representing the information of the current frame itself with the context features of the current frame representing the global information of the sequence. It contains both local details and global correlations.

[0113] The first feature tensor refers to a multidimensional mathematical array formed by stacking the fused features corresponding to each frame of texture residual image in the image sequence in sequence order. It is used to comprehensively represent the temporal evolution law and dependency relationship of the entire image sequence.

[0114] In this step, firstly, the generated texture residual image is read, and by tracing its source, the original standard medical image used to generate it is found. Then, the pre-stored subtype identifier is retrieved from the metadata or associated database of the standard medical image.

[0115] Secondly, load the preset subtype order list, use each subtype identifier obtained in the previous step as the query key, and search in the subtype order list to determine a unique arrangement position value for each identifier that represents its order, for example, benign corresponds to 1, high-grade intraepithelial neoplasia corresponds to 2, and adenocarcinoma corresponds to 3.

[0116] Next, based on the previously determined arrangement position values, all texture residual images are sorted. All images and their corresponding arrangement position values ​​are placed into a list. Then, a sorting algorithm is used to reorganize this list in ascending order of arrangement position values. After sorting, an image sequence is obtained.

[0117] The image sequence is then input into the temporal encoder. Starting from the first frame of the image sequence, the texture residual image of the first frame is sent to the processing unit of the temporal encoder. The processing unit is usually composed of a convolutional neural network. The processing unit performs multi-layer convolution and pooling operations on the texture residual image of the first frame, converting it into a fixed-dimensional mathematical vector, i.e., the current image feature. At the same time, the internal state vector of the processing unit of the temporal encoder is fused with the current image feature. The two vectors are input into a fully connected layer and non-linearly activated to generate a new state vector to replace the original internal state vector, thereby realizing the state update.

[0118] Next, a mathematical vector representing the position of the next frame in the sequence is generated, i.e., the position indication information. The updated internal state vector is combined with this position indication information. By adding the vectors, an intermediate feature is generated. This intermediate feature is used to calculate the association strength distribution. The intermediate feature is then multiplied by a transformed feature corresponding to each frame in the sequence, including past and future frames, to obtain a series of original association scores. These scores are then normalized using the Softmax function so that the sum of all scores is 1. This set of normalized scores is the association strength distribution.

[0119] Then, using this set of correlation strength distributions as weight coefficients, the image features corresponding to each frame in the image sequence are weighted and summed to calculate the context features of the current frame.

[0120] Then, the current image features and the context features of the current frame are aggregated. The vectors are concatenated and then fused through a fully connected layer to output a fused feature that represents the meaning of the current frame in the sequence context. The above process is repeated to process the texture residual image of each frame in the image sequence according to the preset subtype order. After all frames in the image sequence have been processed, the fused features output by each frame are stacked together in the order of the frames to form a three-dimensional mathematical array as the first feature tensor.

[0121] For example, after generating 200 texture residual images, the source of these 200 images is traced to find the 200 standard medical images used to generate them, and the pathological diagnosis corresponding to each standard image is read from the pathology report of Hospital A as a subtype identifier, such as atypical adenomatous hyperplasia, minimally invasive adenocarcinoma, invasive adenocarcinoma, etc.

[0122] The system then invokes a pre-defined subtype order list. Based on this list, atypical adenomatous hyperplasia is mapped to position 1, microinvasive adenocarcinoma to position 2, and invasive adenocarcinoma to position 3. Using position 1, 2, and 3, a fast sorting algorithm is employed to reorganize all 200 texture residual images. After sorting, all residual images identified as atypical adenomatous hyperplasia are placed first, followed by microinvasive adenocarcinoma, and finally invasive adenocarcinoma, thus forming an image sequence that simulates the direction of disease progression.

[0123] The ordered image sequence is then input into a pre-constructed temporal encoder, where the processing unit of the temporal encoder is a small convolutional neural network. The first frame residual image of the image sequence is read in through the temporal encoder, and after several layers of convolution and ReLU activation function processing, a 128-dimensional vector is extracted as the current image feature. This feature is then fused with a 256-dimensional internal state vector that is initially set to all zeros, and a new 256-dimensional internal state vector is generated through a fully connected layer.

[0124] Next, the new state vector is added to the position encoding vector representing the second frame's location to generate intermediate features. These intermediate features are then multiplied by the pre-calculated feature keys of all 200 frames in the image sequence and subjected to Softmax normalization to obtain a correlation strength distribution containing 200 values. The first frame itself may have the highest weight.

[0125] Then, the feature values ​​of each of the 200 frames are weighted and summed according to the correlation strength distribution to obtain the context features of the first frame. The current image features of the first frame are then concatenated with its context features and fused through a fully connected layer to output the fused features of the first frame. This process is repeated until the 200th frame of the sequence is processed. Finally, the 200 fused features are stacked in order to form a 200x256 first feature tensor.

[0126] Step 104: Decouple the first feature tensor to separate the first texture feature and the second texture feature, wherein the first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion itself.

[0127] Optionally, step 104 may specifically include:

[0128] Step 1041: Input the first feature tensor into both the pre-trained first feature extraction unit and the pre-trained second feature extraction unit.

[0129] Step 1042: The first feature tensor is processed by the first feature extraction unit to extract the initial first feature, and the first feature tensor is processed by the second feature extraction unit to extract the initial second feature.

[0130] Step 1043: Calculate the matching degree between the initial first feature and the known imaging pattern feature, and calculate the statistical independence measure between the initial second feature and the initial first feature, wherein the known imaging pattern feature is used to represent different medical imaging devices and imaging protocols.

[0131] Step 1044: Based on the matching degree and the statistical independence measure, construct an optimization objective function.

[0132] Step 1045: The processing parameters of the first feature extraction unit and the second feature extraction unit are iteratively adjusted to increase the value of the optimization objective function.

[0133] Step 1046: When the optimization objective function increases to meet the preset optimization conditions, the first texture feature is output by the first feature extraction unit, and the second texture feature is output by the second feature extraction unit.

[0134] In this step, the first texture feature refers to the mathematical vector obtained after decoupling to characterize the device and protocol information originating from the imaging process, which is used to isolate and eliminate interference caused by differences in imaging conditions in subsequent analysis.

[0135] The second texture feature refers to the mathematical vector obtained after decoupling, which is used to characterize the biological characteristics of the lesion itself and serves as the core basis for subsequent disease classification and diagnosis.

[0136] The pre-trained first feature extraction unit refers to a small neural network module that has been trained on a large amount of historical medical image data and whose internal parameters have been fixed. Its training goal is to make its output features strongly correlated with the imaging conditions. It is used to specifically extract patterns related to imaging equipment and protocols from the first feature tensor. It is obtained by training the model on historical data with the goal of maximizing the matching degree with known imaging pattern features.

[0137] The pre-trained second feature extraction unit refers to a small neural network module that has been trained on a large amount of historical medical image data and whose internal parameters have been fixed. Its training objective is to make its output features as unrelated as possible to the output of the first feature extraction unit. It is used to extract patterns related to the biological characteristics of lesions from the first feature tensor. At the same time, the model is trained on historical data with minimizing the correlation with the output of the first feature extraction unit as one of the training objectives.

[0138] Known imaging pattern features refer to a pre-constructed set of feature vectors, where each vector represents a specific imaging condition pattern, such as "thin-slice bone algorithm reconstructed image from CT device A" or "T1-weighted image from MRI device B". It is used as a benchmark to measure whether the output of the first feature extraction unit matches the imaging information. It is obtained by collecting a large number of images that are biologically as similar as possible but only different in imaging conditions and extracting their common feature patterns.

[0139] Statistical independence measures are numerical indicators used to quantify the degree of weak correlation between two random variables or vectors. The lower the value, the more independent and weaker the correlation between the two. They are used to measure the degree of separation or decorrelation between the initial second feature and the initial first feature.

[0140] The decoupling process is achieved through two parallel feature extraction networks, with the optimization objectives of maximizing the matching degree with known imaging protocol features and minimizing the correlation between the two types of features, respectively. Specifically, mutual information minimization loss or adversarial training strategies can be adopted to ensure that the first texture feature mainly reflects device and protocol information, and the second texture feature mainly reflects biological characteristics.

[0141] The optimization objective function is a single scalar function that takes the matching degree and statistical independence measure as input variables and combines them through a mathematical formula. The value of the function is used to comprehensively evaluate the decoupling performance of the two feature extraction units.

[0142] Preset optimization conditions refer to the convergence criteria set during the iterative optimization process to determine when to stop optimization and to decide when to output the final first and second texture features.

[0143] In this step, the first feature tensor is first loaded, and the pre-trained first feature extraction unit and the pre-trained second feature extraction unit are called from storage. The first feature tensor is input into the two feature extraction units at the same time. The first feature extraction unit performs linear transformation and nonlinear mapping operations on the input first feature tensor and outputs the initial first feature. At the same time, the second feature extraction unit performs linear transformation and nonlinear mapping operations on the same first feature tensor and outputs the initial second feature.

[0144] Next, the matching degree between the initial first feature and the known imaging pattern features is calculated. The set of known imaging pattern features is loaded from the database, and then the cosine similarity between the initial first feature and each known imaging pattern feature in the set is calculated. The maximum value of the cosine similarity is taken as the matching degree. At the same time, the statistical independence measure between the initial second feature and the initial first feature is calculated. The absolute value of the Pearson correlation coefficient is used as the measure. First, the correlation coefficient between each corresponding dimension of the initial first feature and the initial second feature is calculated. After taking the absolute value, the absolute values ​​of all dimensions are averaged to obtain a statistical independence measure between 0 and 1. The lower this value, the higher the independence.

[0145] Next, based on the calculated matching degree and statistical independence measure, an optimization objective function is constructed. This function is usually designed as: Optimization objective function value = matching degree - λ * statistical independence measure, where λ is a preset positive coefficient used to balance the weights of the two terms. This function encourages high matching degree and low statistical independence measure.

[0146] Then, the objective function is optimized through iterative adjustments. Using the idea of ​​gradient descent, the gradient of the current objective function value with respect to all adjustable processing parameters within the first and second feature extraction units is calculated. Then, these processing parameters are updated with a very small step size along the direction that increases the objective function value, i.e., the opposite direction of the gradient. After that, it is determined whether the objective function value meets the preset optimization conditions after the parameter update. The preset optimization conditions are usually that the function value increases by less than a very small threshold in multiple consecutive iterations. If not, the first feature tensor is reprocessed using the updated feature extraction unit, and the initial first feature, initial second feature, matching degree, statistical independence measure, and new objective function value are recalculated. The iteration continues. When the preset optimization conditions are met, the iteration stops and the updated first and second feature extraction units output the first texture feature and the second texture feature.

[0147] For example, following the previous step, after obtaining a 200x256 first feature tensor, the first feature tensor is loaded, and the first feature extraction unit and the second feature extraction unit are called simultaneously. The first feature extraction unit and the second feature extraction unit each have three layers, and the output dimension is 64. The first feature tensor is sent to the two feature extraction units at the same time. The first feature extraction unit performs operations on the first feature tensor and outputs a 64-dimensional vector as the initial first feature. At the same time, the second feature extraction unit outputs another 64-dimensional vector as the initial second feature.

[0148] The matching degree between the initial first feature and a database containing features of five known CT scan protocols was then calculated. Cosine similarity was used to determine that it was most similar to the high-resolution bone algorithm protocol features, with a similarity of 0.85. Next, the Pearson correlation coefficients between the initial second feature and the initial first feature across all dimensions were calculated. The absolute values ​​were then averaged to obtain a statistical independence measure of 0.6. A balance coefficient λ was set to 0.5, and an optimization objective function was constructed: 0.85 - 0.5 * 0.6 = 0.55. The gradient of this objective function with respect to the parameters of the two feature extraction units was calculated and slightly updated.

[0149] The first feature tensor is then reprocessed using the updated feature extraction unit to obtain a new initial first feature and a new initial second feature. The new matching degree may increase to 0.88, and the new independence measure may decrease to 0.4, thus the objective function value becomes 0.68.

[0150] This iterative process continues until, after multiple iterations, the objective function value increases to 0.92 and stabilizes, satisfying the preset convergence condition. The iteration then stops, and the final 64-dimensional vector output by the first feature extraction unit is taken as the first texture feature, which contains imaging protocol information such as the high-resolution bone algorithm. At the same time, the final 64-dimensional vector output by the second feature extraction unit is taken as the second texture feature.

[0151] Step 105: Within a preset potential space, the first texture feature and the second texture feature are recombined to generate a recombined feature.

[0152] Optionally, step 105 may specifically include:

[0153] Step 1051: Map the first texture feature and the second texture feature into the preset latent space respectively to obtain the first latent feature representation and the second latent feature representation.

[0154] Step 1052: Calculate the recombination weight vector based on the second texture feature.

[0155] Step 1053: Using the recombined weight vector, perform a weighted fusion operation on the first latent feature representation and the second latent feature representation to generate a combined latent space representation.

[0156] Step 1054: Map the combined latent space representation to output features with different dimensions than the latent space, and use them as recombined features.

[0157] In this step, the pre-defined latent space refers to the low-dimensional feature space that is pre-learned by an autoencoder or variational autoencoder, and its dimensions are fixed (e.g., 128 dimensions).

[0158] The recombined feature refers to a comprehensive mathematical vector formed by adaptively fusing the first texture feature representing imaging information and the second texture feature representing biological information within a predefined latent space. This vector serves as the input basis for constructing the texture relationship map structure inside the lesion.

[0159] The first latent feature representation refers to the coordinate vector obtained by projecting the first texture feature onto a preset latent space through a mathematical transformation function, which is used to represent imaging-related information within a unified framework.

[0160] The second latent feature representation refers to the coordinate vector obtained by projecting the second texture feature onto a predefined latent space through a mathematical transformation function. This vector is used to represent biologically relevant information within a unified framework.

[0161] The recombined weight vector is a numerical vector whose dimension, calculated from the second texture feature, matches the dimension of the latent space feature. The value of each element in the vector is between 0 and 1. It is used to dynamically determine the proportion that should be adopted in each feature dimension when fusing the first latent feature representation and the second latent feature representation. It is dynamically generated based on the second texture feature through a lightweight fully connected network and is used to weight and fuse the representations of the two types of features in the latent space.

[0162] The combined latent space representation refers to a transitional feature vector obtained within a preset latent space after element-wise weighted summation of the first and second latent feature representations using the recombined weight vector, which is used to temporarily store the fusion result.

[0163] The recombined feature refers to the final mathematical vector obtained by mapping the combined latent space representation from the dimension of the latent space to another target space that usually has a different dimension.

[0164] In this step, firstly, the first texture feature and second texture feature are loaded, and the first texture feature is input into the first fully connected layer of the preset latent space. The fully connected layer is equipped with a non-linear activation function such as ReLU. The input features are linearly combined and non-linearly transformed through the fully connected layer to output the first latent feature representation. At the same time, the second texture feature is input into the second fully connected layer of the preset latent space. The second fully connected layer linearly combines the second texture feature to output the second latent feature representation.

[0165] Next, a recombined weight vector is calculated based on the second texture feature. The second texture feature is input into a specially designed weight generation network, and the Sigmoid activation function is used to compress each element value of the second texture feature to between 0 and 1, so as to output a recombined weight vector with the same dimension as the first latent feature representation and the second latent feature representation.

[0166] Next, the calculated recombination weight vector is used to perform weighted fusion of the two latent feature representations. The recombination weight vector is multiplied element-wise with the first latent feature representation to obtain a weighted first component. At the same time, a complement weight vector is calculated, where each element is equal to 1 minus the element at the corresponding position of the recombination weight vector. This complement weight vector is then multiplied element-wise with the second latent feature representation to obtain a weighted second component. Finally, the two weighted components are added element-wise to generate the combined latent space representation.

[0167] Then, a final dimensionality transformation is performed on the combined latent space representation. The combined latent space representation is input into the third fully connected layer without changing the batch and sequence dimensions of the features, but the feature dimension of the combined latent space representation is transformed from the dimension of the latent space to the final desired output dimension to obtain the recombined features.

[0168] For example, following the previous step, after obtaining a 64-dimensional first texture feature and a 64-dimensional second texture feature, the preset latent space dimension is 128 dimensions. The first texture feature is input into a first fully connected layer with 128 output neurons in the preset latent space dimension, followed by a ReLU activation function, to output a 128-dimensional first latent feature representation;

[0169] Simultaneously, the second texture feature is input into another independent 128-dimensional fully connected layer, also connected to a ReLU activation function, outputting a 128-dimensional second latent feature representation. Then, the second texture feature is input into a weight generation network. This weight generation network first passes through a 64-dimensional fully connected layer, then through a 128-dimensional fully connected layer, and finally uses a Sigmoid function to output a 128-dimensional recombined weight vector, assuming its first few values ​​are [0.2, 0.9, 0.05, ...]. This weight vector is then used for fusion. For example, for the first dimension, the complement weight 1 - 0.2 = 0.8 is calculated. 0.2 is multiplied by the value of the first latent feature representation in that dimension, and 0.8 is multiplied by the value of the second latent feature representation in that dimension. The two products are then added together to obtain the value of the combined latent space representation in that dimension. This process is repeated for all 128 dimensions to generate the combined latent space representation.

[0170] The combined latent space representation is then input into the third fully connected layer, and after linear transformation, a 256-dimensional vector is obtained. This vector is the final recombined feature.

[0171] Step 106: Based on the recombined features, construct a graph structure and use a hierarchical graph neural network to perform message passing and aggregation on the graph structure to extract a topological embedding vector. The topological embedding vector is used to characterize the internal texture structure and relationships of the lesion region of the medical image to be analyzed.

[0172] Optionally, step 106 may specifically include:

[0173] Step 1061: Extract feature representations of multiple lesion sub-regions from the recombined features.

[0174] Step 1062: Using the feature representation of each lesion sub-region as a node, calculate the connection strength between each node based on the feature correlation degree and position correlation degree between each lesion sub-region, and assign the connection strength to the edges between the corresponding nodes to construct the initial graph structure.

[0175] Step 1063: Obtain the first feature information of other nodes connected to any node in the initial graph structure, and combine the first feature information with the first feature information of the node itself to update the first feature information of the node, so as to generate the updated graph structure.

[0176] Step 1064: On the updated graph structure, multiple nodes that meet the preset merging conditions in space are aggregated into a target node, and aggregation feature information is generated for the target node. Then, a high-level graph structure is constructed based on the target node and the aggregation feature information corresponding to the target node.

[0177] Step 1065: Obtain the aggregated feature information of other target nodes connected to any target node in the high-level graph structure, and combine the aggregated feature information with the aggregated feature information of the target node itself to update the aggregated feature information of the target node.

[0178] Step 1066: Integrate the aggregated feature information of all updated target nodes to generate a topological embedding vector.

[0179] In this step, the feature representation of the lesion sub-region refers to the feature vector corresponding to each local small region divided from the reconstructed features according to the spatial position of the lesion outline, which is used to characterize the independent texture attributes of different micro-regions inside the lesion.

[0180] The nodes of the graph structure are the feature vectors of the lesion sub-regions, and the edge connection strength is calculated based on the weighted average of feature similarity and spatial distance.

[0181] Hierarchical graph neural networks employ graph attention networks or graph convolutional networks for multi-layer message passing, and construct high-level graph structures through node clustering, ultimately extracting embedding vectors that integrate multi-scale topological information.

[0182] Feature correlation degree refers to a numerical index used to measure the similarity in texture patterns between the feature expressions of any two lesion sub-regions. The higher the value, the more similar the texture features of the two regions are. It is obtained by calculating the cosine similarity between two feature vectors.

[0183] Location correlation refers to a numerical index used to measure the physical distance between any two lesion sub-regions in the original medical image space. The higher the value, the closer the two regions are in space. It is obtained by calculating the Euclidean distance between the coordinates of the center points of the two lesion sub-regions and performing a negative correlation transformation.

[0184] The connection strength between nodes refers to the weight value assigned to the edge connecting two nodes when constructing the graph structure. This weight comprehensively reflects the similarity in features and the spatial proximity of the lesion sub-regions represented by the two nodes, and is used to quantify the tightness of the relationship between nodes in the graph.

[0185] The initial graph structure refers to a graph data model consisting of a set of nodes representing lesion sub-regions and a set of edges with associated connection strength weights, used to initially model the local texture relationships within the lesion.

[0186] The first feature information is a synonym for the feature expression carried by the node in the initial graph structure. It refers to the feature vector of the lesion sub-region corresponding to each node during initialization, and is used as the initial value of the node feature during the message passing process of the graph.

[0187] Preset merging conditions refer to a set of pre-defined logical rules used to determine which nodes in the updated graph structure are spatially close enough to qualify for merging into a larger node. For example, the distance between the center points of nodes is required to be less than a certain threshold, which is defined by setting a spatial distance threshold and a maximum number of nodes to be merged.

[0188] The target node refers to an abstract node that represents a larger lesion area, formed by aggregating multiple original nodes that meet the preset merging conditions in the updated graph structure. It is used to construct a higher-level graph structure to achieve multi-scale analysis.

[0189] Aggregated feature information refers to the feature vector generated for the target node, which represents the comprehensive texture attributes of its corresponding larger region. It is used as the initial feature for the target node to pass messages in the high-level graph structure.

[0190] High-level graph structures refer to graph data models consisting of target nodes and the connecting edges between these target nodes, describing the texture structure relationships of lesion areas at a more macroscopic scale.

[0191] Topological embedding vector refers to a comprehensive mathematical vector used to characterize the complex structure and relationships of the texture within the entire lesion area, integrating multi-level topological information from local details to macroscopic regions.

[0192] In this step, the reconstructed features are loaded first. Since the reconstructed features usually retain the spatial structure information of the lesion, a uniform grid division method is used to divide the feature map in the spatial dimension, thereby extracting multiple non-overlapping local blocks. The feature vector corresponding to each local block is the feature expression of a lesion sub-region.

[0193] Secondly, the feature representation of each lesion sub-region is used as a graph node, and the feature association degree is calculated. The cosine similarity between the feature representation vectors of any two graph nodes is calculated to obtain a value between 0 and 1. Then, the Euclidean distance is calculated based on the center coordinates of the two sub-regions in the original image. The distance is then converted into a value that decreases as the distance increases by taking the reciprocal or by using a Gaussian kernel function, which is also mapped to a value between 0 and 1 as the position association degree.

[0194] The two values ​​are then added together with preset weights to obtain the final connection strength. An edge is created for all graph node pairs with connection strength greater than a certain threshold, and the calculated connection strength value is assigned to this edge, thereby constructing an initial graph structure containing nodes and weighted edges.

[0195] Next, the first round of message passing and aggregation is performed on the initial graph structure to update the node features. For each node in the graph, the first feature information of all its direct neighbors is collected first. Then, a graph neural network layer is used to calculate the attention weight of the current graph node to each neighbor graph node, taking the first feature information of the current node itself, the first feature information of all its neighbors, and the connection strength of the corresponding edges from the current node to each neighbor graph node as input. Then, the features of the neighbor graph nodes are weighted and summed according to these weights. Finally, the weighted summation result is combined with the features of the current graph node itself and passed through a fully connected layer and activation function to output the updated first feature information of the node. After the first feature information of all nodes in the graph has been updated, the updated graph structure is generated.

[0196] Then, graph node aggregation is performed on the updated graph structure to build a higher level of abstraction. All graph nodes are traversed, and the spatial distance between any two or more graph nodes is checked. If the distance between the center points of a group of nodes is less than the preset merging distance threshold, and the number of nodes in this group does not exceed the preset maximum number of merges, then they meet the preset merging conditions. This group of graph nodes is merged into a new target node. The position of this new target node can be set as the average value of the center points of this group of nodes.

[0197] Simultaneously, aggregated feature information is generated for this target node. The updated first feature information of this group of merged graph nodes is input into the average pooling layer, and their feature vectors are averaged element-wise along the node dimension to obtain a new feature vector, which is used as the aggregated feature information of the target node. After all the graph node groups that meet the conditions are merged, all target nodes and their aggregated feature information are obtained. Then, based on the new spatial positions of these target nodes, the spatial proximity relationship between them is recalculated, and new edges are established based on this to construct a high-level graph structure.

[0198] Then, a second round of message passing and aggregation is performed on the high-level graph structure. For each target node in the high-level graph structure, the aggregation feature information of other target nodes directly connected to it is collected. Similarly, the graph neural network layer is used to fuse and calculate the aggregation feature information of the current target node itself, the aggregation feature information of its neighboring target nodes, and the connection strength of the current edge, and output the updated aggregation feature information of the target node.

[0199] Finally, the final features of all target nodes are integrated to generate a topological embedding vector representing the entire lesion.

[0200] For example, following the previous step, after obtaining a 256-dimensional recombined feature, this recombined feature is spatially divided into 16 4x4 sub-regions. A 256-dimensional feature vector is extracted from each sub-region, resulting in feature representations of 16 lesion sub-regions. Using these 16 feature vectors as nodes, the connection strength between node 1 and node 2 is calculated. First, the feature association degree is calculated, with a cosine similarity of 0.8. Then, the position association degree is calculated, as their center points are relatively close, resulting in a Gaussian kernel transformation of 0.9. With feature and position weights each set to 0.5, the connection strength is 0.5 * 0.8 + 0.5 * 0.9 = 0.85. Edges are then constructed for all node pairs with a connection strength > 0.3, forming the initial graph structure.

[0201] Next, the graph attention network layer is used to update the node features. For example, if node 1 has three neighbors, the graph attention network calculates the attention weights of node 1 to the features of the three neighbors as 0.4, 0.4, and 0.2, respectively. Based on this, the neighbor features are weighted and summed, and then combined with the features of node 1 itself. The updated 256-dimensional first feature information of node 1 is output through a fully connected layer. This step is repeated until all nodes are updated, and the updated graph structure is obtained. At the same time, the merging distance threshold is set to 2 pixels. When nodes 1, 2, 5, and 6 are found to be spatially adjacent and the distance between each pair is less than the threshold, they are merged into a target node A. The updated features of these four nodes are then averaged and pooled to obtain the 256-dimensional aggregated feature information of the target node A.

[0202] Similarly, other target nodes B, C, and D are generated, and a high-level graph structure is constructed based on their new positions. On the high-level graph structure, for target node A, another graph attention network layer is used to update its own aggregated feature information by combining the aggregated feature information of its neighboring target nodes B and C. Global average pooling is then performed on the updated aggregated feature information of the final target nodes A, B, C, and D to obtain a 256-dimensional vector, namely the topological embedding vector.

[0203] Step 107: Input the topological embedding vector into a preset classifier to output the texture classification result for the medical image to be analyzed.

[0204] Optionally, step 107 may specifically include:

[0205] Step 1071: Input the topological embedding vector into the first data processing layer of the preset classifier.

[0206] Step 1072: In the first data processing layer, a first linear transformation operation is performed on the topological embedding vector, and the result of the first linear transformation operation is input to the first nonlinear processing unit for processing to generate a first intermediate feature.

[0207] Step 1073: The first intermediate feature is input to the second data processing layer of the preset classifier. In the second data processing layer, a second linear transformation operation is performed on the first intermediate feature, and the result of the second linear transformation operation is input to the second nonlinear processing unit for processing to generate the second intermediate feature.

[0208] Step 1074: The second intermediate feature is input to the output layer of the preset classifier. In the output layer, a third linear transformation operation is performed on the second intermediate feature to generate a classification score vector corresponding to a plurality of preset candidate texture categories. Each element value in the classification score vector corresponds to the confidence level of a candidate texture category.

[0209] Step 1075: Determine the target element value with the largest value from the classification score vector, and output the candidate texture category corresponding to the target element value as the texture classification result for the medical image to be analyzed.

[0210] In this step, the preset classifier refers to a neural network model consisting of multiple sequentially connected data processing layers, which has been trained on a large amount of labeled medical image data. It is used to determine the texture category of the lesion based on the input topological embedding vector. The model is trained on labeled historical data with the goal of minimizing classification error. In this step, the preset classifier is a multilayer perceptron, which contains at least two fully connected layers and a non-linear activation function (such as ReLU). The output layer uses the Softmax function to generate the category probability distribution and finally outputs the texture category with the highest confidence.

[0211] The first data processing layer refers to the first neural network layer in the preset classifier model. It is used to perform preliminary feature transformation and information extraction on the input topological embedding vector. It is constructed by a fully connected layer to achieve linear transformation, followed by a non-linear activation function.

[0212] The first linear transformation operation refers to the mathematical calculation process in the first data processing layer of mapping the input vector to another dimensional space through matrix multiplication and adding a bias vector, which is used to realize the linear combination and dimensional transformation of features.

[0213] The first nonlinear processing unit refers to a computational module that follows the first linear transformation operation. It applies a nonlinear mathematical function to each value after the linear transformation to introduce the nonlinear expressive power necessary for the model to handle complex patterns. This is usually achieved through the ReLU activation function.

[0214] The first intermediate feature refers to a new feature vector output after the topological embedding vector has undergone complete processing by the first data processing layer, that is, after sequentially undergoing the first linear transformation operation and the first nonlinear processing unit, which is used as the input of the next data processing layer.

[0215] The second data processing layer refers to the second neural network layer in the preset classifier model that follows the first data processing layer. It is used to further abstract the features and condense the information of the first intermediate features. Its structure is similar to that of the first data processing layer, consisting of a fully connected layer and a non-linear activation function.

[0216] The second linear transformation operation refers to the linear mapping and bias addition calculation performed on the first intermediate feature in the second data processing layer. Its mathematical form is the same as that of the first linear transformation operation, but the parameters are different.

[0217] The second nonlinear processing unit refers to the nonlinear activation calculation module that follows the second linear transformation operation in the second data processing layer, and its function is the same as that of the first nonlinear processing unit.

[0218] The second intermediate feature refers to the feature vector output after the first intermediate feature has been fully processed by the second data processing layer. Its dimension is usually lower than that of the first intermediate feature, and it is used as the input of the output layer.

[0219] The output layer refers to the last neural network layer in the predefined classifier model. It is used to map the second intermediate features to a dimension corresponding to the number of candidate texture categories and generate the raw score for each category. It is usually composed of a fully connected layer without a non-linear activation function.

[0220] The third linear transformation operation refers to the final linear mapping and bias addition calculation performed on the second intermediate feature in the output layer, which transforms it into a vector with a length equal to the number of candidate categories.

[0221] The classification score vector refers to the original numerical vector obtained after the third linear transformation operation of the output layer. The value of each element in this vector represents the relative probability or confidence level of the texture features of the input image belonging to a specific candidate texture category.

[0222] Texture classification results refer to the clear category labels of the lesion areas in the medical images to be analyzed, which are finally output by the system and used to provide doctors or users with the final auxiliary diagnostic conclusions.

[0223] In this step, the topological embedding vector is first loaded and fed into the first data processing layer of the preset classifier as input. The first linear transformation operation is then performed. The input vector is multiplied by the preset weight matrix, and then the preset bias vector is added to obtain the vector result after preliminary transformation. This vector result is then input into the first nonlinear processing unit. The first nonlinear processing unit applies the ReLU activation function to each input value, that is, sets all values ​​less than zero to zero, while keeping values ​​greater than or equal to zero unchanged, thereby obtaining the first intermediate feature.

[0224] Next, the generated first intermediate feature is fed into the second data processing layer of the preset classifier and a second linear transformation operation is performed. Another preset weight matrix is ​​used to perform matrix multiplication on the first intermediate feature, and another preset bias vector is added to obtain a new vector. This new vector is then input into the second nonlinear processing unit, and the ReLU activation function is used to process each value in the vector to finally output the second intermediate feature.

[0225] Next, the second intermediate feature is input into the output layer of the preset classifier, and the third linear transformation operation is performed. The second intermediate feature is multiplied by the last preset weight matrix, and the last preset bias vector is added. The purpose of this transformation is to map the feature dimension to the number of candidate texture categories in order to obtain the classification score vector.

[0226] Finally, the classification score vector is parsed, all element values ​​in the classification score vector are traversed, the target element value with the largest value is found, and the index position of the target element value in the classification score vector is used to query the preset candidate texture category list to find the specific category name corresponding to the index position, so as to obtain the final texture classification result output for the lesion area of ​​the medical image to be analyzed.

[0227] For example, following the previous embodiment, in the preset classifier, the weight matrix of the first data processing layer is 128x256 in size, and the bias vector is 128-dimensional. The first linear transformation operation is performed, multiplying the 256-dimensional input with the matrix and adding the bias to obtain a 128-dimensional vector. Assume that the first value of this vector is -0.1 and the second value is 0.5. Subsequently, the first nonlinear processing unit processes it, changing -0.1 to 0 and keeping 0.5 unchanged, and outputting the first intermediate feature of 128 dimensions.

[0228] Next, the second linear transformation operation is performed through the second data processing layer to output a 64-dimensional vector, which is then processed by the ReLU function of the second nonlinear processing unit to obtain the 64-dimensional second intermediate feature.

[0229] Then, a third linear transformation operation is performed through the output layer to obtain a 3-dimensional classification score vector, for example, with a value of [0.1, 0.7, 0.2]. The three values ​​are then traversed to determine the target element with the largest value, 0.7, which is located in the second position of the vector. The category list is then queried, and the second position corresponds to the "boundary" category. Therefore, the final output texture classification result is boundary.

[0230] Figure 2 This application provides a schematic diagram of the structure of an artificial intelligence-based medical image texture analysis system, such as... Figure 2 As shown, the system includes:

[0231] The acquisition module 21 is used to acquire a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed.

[0232] The calculation module 22 is used to perform a difference operation between the medical image to be analyzed and each image in the standard medical image to generate a set of texture residual images;

[0233] The determination module 23 is used to arrange the set of texture residual images according to a preset subtype order to form an image sequence, and input the image sequence to a temporal encoder to determine the dependency relationship and evolution law between different residual images in the image sequence through the temporal encoder, and generate a first feature tensor, which is used to characterize the temporal features of the image sequence.

[0234] The separation module 24 is used to decouple the first feature tensor to separate the first texture feature and the second texture feature, wherein the first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion itself.

[0235] The recombination module 25 is used to recombine the first texture feature and the second texture feature within a preset potential space to generate a recombined feature;

[0236] The construction module 26 is used to construct a graph structure based on the recombined features, and to use a hierarchical graph neural network to perform message passing and aggregation on the graph structure to extract a topological embedding vector. The topological embedding vector is used to characterize the internal texture structure and relationship of the lesion region of the medical image to be analyzed.

[0237] Output module 27 is used to input the topological embedding vector into a preset classifier to output the texture classification result for the medical image to be analyzed.

[0238] Figure 2 The aforementioned AI-based medical image texture analysis system can perform... Figure 1 The implementation principle and technical effects of the AI-based medical image texture analysis method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the AI-based medical image texture analysis system in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0239] In one possible design, Figure 2 The illustrated embodiment of an artificial intelligence-based medical image texture analysis system can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0240] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0241] The processing component 32 is used for the above Figure 1 The embodiment describes an artificial intelligence-based medical image texture analysis method.

[0242] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A medical image texture analysis method based on artificial intelligence, characterized in that, include: Acquire a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed; The medical image to be analyzed is compared with each image in the standard medical image to generate a set of texture residual images. The set of texture residual images are arranged according to a preset subtype order to form an image sequence. The image sequence is then input into a temporal encoder. The temporal encoder determines the dependency relationship and evolution law between different residual images in the image sequence and generates a first feature tensor. The first feature tensor is used to characterize the temporal features of the image sequence. The first feature tensor is decoupled to separate the first texture feature and the second texture feature. The first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion. Within a preset potential space, the first texture feature and the second texture feature are recombined to generate a recombined feature; Based on the recombined features, a graph structure is constructed, and a hierarchical graph neural network is used to perform message passing and aggregation on the graph structure to extract topological embedding vectors. The topological embedding vectors are used to characterize the internal texture structure and relationships of the lesion region of the medical image to be analyzed. The topological embedding vector is input into a preset classifier, and the texture classification result of the medical image to be analyzed is output.

2. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, The set of texture residual images are arranged according to a preset subtype order to form an image sequence, including: Obtain the subtype identifier of the standard medical image associated with the texture residual image, wherein the subtype identifier is used to indicate the specific category to which the standard medical image belongs; According to the preset subtype order, the arrangement position value corresponding to each subtype identifier is determined, wherein the preset subtype order defines the order between different subtype identifiers; Based on the arrangement position values, the texture residual images are reorganized so that the texture residual images are arranged in ascending order of the arrangement position values ​​to form an image sequence.

3. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, The image sequence is input to a temporal encoder, which determines the dependencies and evolution patterns between different residual images in the image sequence, generating a first feature tensor, including: Each frame of the texture residual image in the image sequence is input into the processing unit of the temporal encoder, and the current input texture residual image is used to extract features to obtain the current image features. The current image features are fused with the initial internal state vector of the processing unit to update the internal state vector; The updated internal state vector is combined with the position indication information corresponding to the texture residual image to be processed in the next step to generate intermediate features; Based on the intermediate features, the correlation strength distribution between the currently processed texture residual image and all other texture residual images in the image sequence is calculated; Based on the correlation strength distribution, the image features corresponding to all texture residual images in the image sequence are weighted and combined to generate the context features of the current frame; The current image features are aggregated with the context features of the current frame to output the fused features of the current frame, and the fused features of all frames are collected to form the first feature tensor.

4. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, The first feature tensor is decoupled to separate a first texture feature and a second texture feature. The first texture feature characterizes device and protocol information originating from the imaging process, while the second texture feature characterizes biological characteristics originating from the lesion, including: The first feature tensor is simultaneously input into the pre-trained first feature extraction unit and the pre-trained second feature extraction unit; The first feature tensor is processed by the first feature extraction unit to obtain an initial first feature, and the first feature tensor is processed by the second feature extraction unit to obtain an initial second feature. Calculate the matching degree between the initial first feature and the known imaging pattern feature, and calculate the statistical independence measure between the initial second feature and the initial first feature, wherein the known imaging pattern feature is used to represent different medical imaging devices and imaging protocols; Based on the matching degree and the statistical independence measure, an optimization objective function is constructed; By iteratively adjusting the processing parameters of the first feature extraction unit and the second feature extraction unit, the value of the optimization objective function is increased; When the optimization objective function increases to meet the preset optimization conditions, the first texture feature is output by the first feature extraction unit, and the second texture feature is output by the second feature extraction unit.

5. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, Within a preset latent space, the first texture feature and the second texture feature are recombined to generate a recombined feature, including: The first texture feature and the second texture feature are respectively mapped to the preset latent space to obtain the first latent feature representation and the second latent feature representation; Based on the second texture feature, calculate the reconstructed weight vector; Using the recombined weight vector, a weighted fusion operation is performed on the first latent feature representation and the second latent feature representation to generate a combined latent space representation; The combined latent space representation is mapped to output features with different dimensions than the latent space, which are then used as the recombined features.

6. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, Based on the recombined features, a graph structure is constructed, and a hierarchical graph neural network is used to perform message passing and aggregation on the graph structure to extract topological embedding vectors, including: From the recombined features, feature representations of multiple lesion sub-regions are extracted; Each lesion sub-region is used as a node. The connection strength between each node is calculated based on the degree of feature correlation and positional correlation between each lesion sub-region. The connection strength is then assigned to the edges between the corresponding nodes to construct the initial graph structure. Obtain the first feature information of other nodes connected to any node in the initial graph structure, combine the first feature information with the first feature information of the node itself, update the first feature information of the node, and generate the updated graph structure. In the updated graph structure, multiple nodes that meet the preset merging conditions in space are aggregated into a target node, and aggregation feature information is generated for the target node. Then, a high-level graph structure is constructed based on the target node and the aggregation feature information corresponding to the target node. The aggregated feature information of other target nodes connected to any target node in the high-level graph structure is obtained, and the aggregated feature information is combined with the aggregated feature information of the target node itself to update the aggregated feature information of the target node. The aggregated feature information of all updated target nodes is integrated to generate a topological embedding vector.

7. The medical image texture analysis method based on artificial intelligence according to claim 1, characterized in that, The topological embedding vector is input into a preset classifier, and the texture classification result of the medical image to be analyzed is output, including: The topology embedding vector is input into the first data processing layer of the preset classifier; In the first data processing layer, a first linear transformation operation is performed on the topological embedding vector, and the result of the first linear transformation operation is input to the first nonlinear processing unit for processing to generate a first intermediate feature. The first intermediate feature is input to the second data processing layer of the preset classifier. In the second data processing layer, a second linear transformation operation is performed on the first intermediate feature, and the result of the second linear transformation operation is input to the second nonlinear processing unit for processing to generate the second intermediate feature. The second intermediate feature is input to the output layer of the preset classifier. In the output layer, a third linear transformation operation is performed on the second intermediate feature to generate a classification score vector corresponding to a plurality of preset candidate texture categories. Each element value in the classification score vector corresponds to the confidence level of a candidate texture category. The target element value with the largest value is determined from the classification score vector, and the candidate texture category corresponding to the target element value is output as the texture classification result for the medical image to be analyzed.

8. A medical image texture analysis system based on artificial intelligence, characterized in that, include: The acquisition module is used to acquire a medical image to be analyzed and a set of standard medical images with the same lesion contour as the medical image to be analyzed; The calculation module is used to perform a difference operation between the medical image to be analyzed and each image in the standard medical image to generate a set of texture residual images; The determination module is used to arrange the set of texture residual images according to a preset subtype order to form an image sequence, and input the image sequence to a temporal encoder. The temporal encoder determines the dependency relationship and evolution law between different residual images in the image sequence and generates a first feature tensor. The first feature tensor is used to characterize the temporal features of the image sequence. The separation module is used to decouple the first feature tensor and separate the first texture feature and the second texture feature, wherein the first texture feature is used to characterize the device and protocol information originating from the imaging process, and the second texture feature is used to characterize the biological characteristic information originating from the lesion itself. The recombination module is used to recombine the first texture feature and the second texture feature within a preset potential space to generate a recombined feature; The construction module is used to construct a graph structure based on the recombined features, and to use a hierarchical graph neural network to perform message passing and aggregation on the graph structure to extract a topological embedding vector. The topological embedding vector is used to characterize the internal texture structure and relationships of the lesion region of the medical image to be analyzed. The output module is used to input the topological embedding vector into a preset classifier and output the texture classification result of the medical image to be analyzed.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement an artificial intelligence-based medical image texture analysis method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements an artificial intelligence-based medical image texture analysis method as described in any one of claims 1 to 7.

Images