Method for early evaluation of therapeutic efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging

Abstract

The application relates to a method for early evaluation of curative effect of hepatocellular carcinoma transformation treatment based on magnetic resonance imaging. The method maps multi-phase dynamic contrast-enhanced images and multi-sequence structure images into statistically independent enhancement embedding and texture embedding by a double-branch decoupling encoder network, with mutual information lower bound minimization as a constraint, and actively separates enhancement dynamic features and tissue texture features; then, the enhancement sub-regions and the texture sub-regions are divided, a time series graph is constructed based on spatial adjacency and cross-time voxel transformation relationship; the graph sequence encoder is used to respectively output enhancement and texture sub-region transformation probability matrices; finally, the difference between the volume-weighted enhancement transformation rate and the texture transformation rate is taken as an early curative effect index. The scheme models the evolution of decoupled representation and independent sub-regions, so that the vascular normalization effect and the tissue necrosis effect are tracked in their respective subspaces, the sensitive detection of early response is realized, and the problems of signal dilution and unknown response source caused by feature entanglement are solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and deep learning, and in particular relates to an early assessment method for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging. Background Technology

[0002] Hepatocellular carcinoma (HCC) is one of the most common malignant liver tumors in my country and globally. For patients in the middle to late stages who are initially diagnosed but cannot be resected due to excessive tumor burden, vascular invasion, or insufficient remaining liver volume, the emerging conversion therapy strategy has brought hope for radical surgery. Conversion therapy, through the combination of targeted drugs and immune checkpoint inhibitors, can achieve significant tumor shrinkage, necrosis, or downstaging in some patients, thus creating the opportunity for radical resection. However, the efficacy of conversion therapy exhibits significant individual heterogeneity: some patients experience deep pathological remission early in treatment, while others, even without significant changes in tumor volume, have already experienced microscopic responses such as vascular normalization, indicating a favorable subsequent treatment outcome. Current clinical evaluation systems mainly rely on the Solid Tumor Efficacy Evaluation Standard or its modified version, using changes in maximum tumor diameter or the extent of surviving tumors during arterial phase enhancement as the evaluation criteria. These assessment methods, based on single-dimensional size or single-phase enhancement, are difficult to capture the complex internal remodeling process induced by targeted immunotherapy, especially with severely insufficient sensitivity in detecting early efficacy signals, often leading to delays or errors in treatment decisions.

[0003] To overcome the limitations of purely anatomical assessments, high-throughput quantitative analysis techniques based on radiomics have been introduced into the field of liver cancer treatment efficacy prediction. These techniques construct predictive models by extracting a large number of texture, morphological, and functional features from different magnetic resonance imaging (MRI) sequences. However, existing technologies generally treat features from different sequence sources as independent parallel channels, utilizing them through simple splicing or decision-level fusion. In the specific context of hepatocellular carcinoma conversion therapy, this assumption does not hold: the contrast agent inflow-clearance time dynamics reflected by dynamic contrast-enhanced MRI and the inherent tissue structure and texture features reflected by T2-weighted imaging and diffusion-weighted imaging have a deep statistical coupling relationship in the feature space. The vascular normalization effect induced by targeted drugs often precedes tissue necrosis, manifesting as changes in enhancement patterns preceding changes in texture features. However, these two types of signals are highly entangled in the original feature space; changes in one dimension are easily diluted by fluctuations in the other, causing the weak signals of early efficacy to be submerged in the total variance of feature changes. Existing methods fail to actively decouple enhancement patterns from structural texture within the feature space, and this technical blind spot constitutes the fundamental reason for the limited sensitivity of early efficacy assessment.

[0004] Furthermore, the tumor's internal response during conversion therapy is not uniform, but rather exhibits spatiotemporal heterogeneity composed of different subregional niches. Subregions such as highly enhanced active areas, poorly vascularized necrotic areas, and fibrotic transition areas not only undergo volume changes within themselves, but their spatial topological relationships also dynamically reorganize during treatment: necrotic areas expand outward from the center, active areas are gradually eroded, and new active foci may appear in adjacent areas while existing active areas transform into necrotic areas. This transformation relationship between subregions contains richer efficacy information than single subregion volume changes or overall average characteristics. However, existing methods either only collect static subregion features at a single time point for prediction or compare images before and after treatment as a whole, failing to explicitly model the transformation and evolution of subregions in both spatial and temporal dimensions. Combining dynamic graph neural networks with multi-time-point analysis of subregions to track the evolutionary trajectory of each subregion is currently a gap in the existing literature.

[0005] Therefore, how to achieve decoupled representation of enhancement patterns and structural textures within the feature space, and on this basis capture the spatiotemporal evolution of tumor subregions during treatment, has become a core technical problem that urgently needs to be solved in the field of accurate evaluation of early efficacy of hepatocellular carcinoma conversion therapy. Summary of the Invention

[0006] Therefore, it is necessary to provide a method for early evaluation of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging to address the above-mentioned technical problems.

[0007] In a first aspect, this application provides a method for early evaluation of the efficacy of conversion therapy for hepatocellular carcinoma based on magnetic resonance imaging, including:

[0008] S1. Obtain liver magnetic resonance images of the patient at at least one time point before and after conversion therapy, preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region.

[0009] S2. Input the multi-phase dynamic enhancement image patch and the multi-sequence structure image patch into the pre-trained dual-branch decoupled encoder network to obtain statistically independent enhancement embedding feature map and texture embedding feature map; wherein, the dual-branch decoupled encoder network includes an enhancement coding branch and a texture coding branch. During the training process, the lower bound of the mutual information between the enhancement embedding and the texture embedding is used as a constraint to make the enhancement embedding feature map and the texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features.

[0010] S3. Based on the enhanced embedding feature map, the tumor voxel is divided into multiple enhanced subregions, and the volume of each enhanced subregion is determined; based on the texture embedding feature map, the tumor voxel is divided into multiple texture subregions, and the volume of each texture subregion is determined.

[0011] S4. Construct an enhancement subspace graph with multiple enhancement sub-regions as nodes, and construct a texture subspace graph with multiple texture sub-regions as nodes. Establish spatial adjacency edges within the first time point and cross-time transformation edges between the first time point and the second time point in the enhancement subspace graph and the texture subspace graph to form a time series graph.

[0012] S5. Input the time series graph into the pre-trained graph sequence encoder, and output the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-time transformation relationship of the sub-region;

[0013] S6. Calculate the enhanced conversion rate based on the enhanced conversion probability matrix and the volume of the enhanced subregion, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture subregion, and generate early efficacy assessment results based on the difference between the enhanced conversion rate and the texture conversion rate.

[0014] Secondly, this application also provides an early assessment system for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging, used to implement the method described in the first aspect, the system comprising:

[0015] The multimodal image preprocessing module is used to acquire liver magnetic resonance images of patients at at least one time point before and after conversion therapy, and to preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region.

[0016] The feature decoupling coding module is used to input multi-phase dynamic enhancement image patches and multi-sequence structure image patches into a pre-trained dual-branch decoupling encoder network to obtain statistically independent enhancement embedding feature maps and texture embedding feature maps. The dual-branch decoupling encoder network includes an enhancement coding branch and a texture coding branch. During training, minimizing the lower bound of the mutual information between enhancement embedding and texture embedding is used as a constraint to make the enhancement embedding feature map and texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features.

[0017] The sub-region segmentation and quantization module is used to divide tumor voxels into multiple enhanced sub-regions based on the enhanced embedding feature map and determine the volume of each enhanced sub-region; and to divide tumor voxels into multiple texture sub-regions based on the texture embedding feature map and determine the volume of each texture sub-region.

[0018] The spatiotemporal graph construction module is used to construct an enhanced subspace graph with multiple enhanced subregions as nodes and a texture subspace graph with multiple texture subregions as nodes. It establishes spatial adjacency edges within the first time point and cross-time transformation edges between the first time point and the second time point in the enhanced subspace graph and the texture subspace graph to form a time series graph.

[0019] The graph sequence evolution coding module is used to input the time series graph into a pre-trained graph sequence encoder and output the enhancement transformation probability matrix of the enhancement sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-temporal transformation relationship of the sub-region;

[0020] The efficacy assessment generation module is used to calculate the enhancement conversion rate based on the enhancement conversion probability matrix and the volume of the enhancement subregion, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture subregion, and generate early efficacy assessment results based on the difference between the enhancement conversion rate and the texture conversion rate.

[0021] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an early assessment method for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging as described in the first aspect.

[0022] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for early evaluation of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging as described in the first aspect.

[0023] The aforementioned method for early assessment of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging utilizes a pre-trained bi-branch decoupled encoder network. Under the constraint of minimizing the lower bound of mutual information, it maps multi-phase dynamic enhancement images and multi-sequence structural images into statistically independent enhancement embedding feature maps and texture embedding feature maps, thereby actively separating the entangled contrast agent dynamic features and tissue texture features in the feature space in traditional methods. Based on the decoupled representation, the enhancement embedding and texture embedding are sub-regionally divided, and a time-series graph that simultaneously carries spatial topology and temporal evolution information is constructed using each sub-region as a node and spatial adjacency and cross-temporal voxel transformation relationships as edges. Subsequently, a graph sequence encoder is used to jointly encode the spatial adjacency edges and cross-temporal transformation edges, independently outputting the enhancement sub-region transformation probability matrix and the texture sub-region transformation probability matrix, respectively capturing the enhancement pattern redistribution driven by vascular effects and the texture pattern redistribution driven by tissue necrosis. Finally, the difference between the volume-weighted enhancement conversion rate and the texture conversion rate is used as a quantitative indicator of early efficacy. This scheme uses a technical approach of "feature space decoupling - independent sub-region evolution modeling - conversion rate difference quantification" to enable the vascular normalization effect and tissue necrosis effect to be tracked and compared separately in their respective independent subspaces. This allows for sensitive detection of the early response to hepatocellular carcinoma conversion therapy, overcoming the technical defects of existing methods where the early weak therapeutic signal is diluted due to feature entanglement and the inability to distinguish the source and direction of response diffusion. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating an early assessment method for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging, provided by this invention.

[0026] Figure 2 This is a schematic diagram of the process of constructing a time series graph in one optional embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the structure of an early assessment system for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging, provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] This invention provides a method for early assessment of the efficacy of conversion therapy for hepatocellular carcinoma based on magnetic resonance imaging (MRI). This method can be deployed in an integrated application environment comprising an MRI scanner, a medical image storage and transmission system, a high-performance computing server, and a clinical display terminal. The MRI scanner acquires multi-sequence, multi-phase images of the liver at baseline before conversion therapy and at various follow-up time points after treatment, storing and transmitting them in a medical digital imaging and communication standard format. The medical image storage and transmission system, acting as a data center, receives raw image data from the imaging equipment, provides data archiving, retrieval, and distribution functions, and outputs standardized image data streams to the subsequent computing server. The computing server is equipped with multiple graphics processing units or dedicated neural network acceleration units, responsible for performing all computationally intensive tasks such as image preprocessing, decoupled encoder forward inference, subregion partitioning, graph construction, graph sequence encoding, and efficacy index calculation. The computing server connects to the clinical display terminal via a high-speed hospital network, sending the generated early efficacy assessment results, conversion probability matrix, and three-dimensional efficacy response heatmap to the terminal. The clinical display terminal can be a physician workstation, a portable display device, or a tablet computer, used to present assessment reports and visualized images to assist clinicians in making subsequent treatment decisions. In the entire data processing flow, the raw images generated by the magnetic resonance imaging equipment are first transmitted to the data receiving module of the computing server through the medical image storage and transmission system. Then, the preprocessing module sequentially performs bias field correction, multi-time point multi-sequence registration, tumor segmentation, intensity normalization, and 3D block cropping. Subsequently, the pre-trained decoupled encoder network extracts features from the standardized image blocks. The sub-region clustering module, graph construction module, graph sequence encoding module, and efficacy evaluation module are executed in sequence. Finally, the results are pushed to the clinical display terminal for visualization through the network.

[0030] Example 1:

[0031] refer to Figure 1 The document presents a flowchart illustrating an early assessment method for the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging, as provided in this application. The method includes the following steps:

[0032] S1. Obtain liver magnetic resonance images of the patient at at least one time point before and after conversion therapy, preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region.

[0033] Specifically, for each hepatocellular carcinoma patient receiving conversion therapy, standardized multi-sequence liver magnetic resonance imaging (MRI) was performed at baseline before treatment initiation and at the first early follow-up after treatment. Dynamic contrast-enhanced MRI employed a three-dimensional volumetric interpolation rapid phase-scrambled gradient echo sequence. Following intravenous bolus injection of gadolinium-containing contrast agent, images were continuously acquired in the mask phase, late arterial phase, portal venous phase, delayed phase, and hepatobiliary phase (selectable depending on the contrast agent type). Each phase covered the entire liver, and the reconstructed voxels were at isotropic resolution. Structural imaging sequences included at least fat-suppressed T2-weighted imaging and diffusion-weighted imaging. Diffusion-weighted imaging acquired images with at least two different diffusion sensitivity coefficient values, and an apparent diffusion coefficient map was calculated. All raw image data from the imaging equipment and phases were exported to the image archiving and communication system in a medical digital imaging and communication standard format.

[0034] The first step in the preprocessing stage is to perform N4ITK bias field correction on the 3D images of each sequence and each phase to eliminate low-frequency grayscale variations caused by inhomogeneous surface coil receiver sensitivity. Bias field correction is achieved by alternately estimating the maximum a posteriori probability of the real image and the bias field, iteratively solving the following optimization problem: given an observed image... Under the premise of finding the bias field and real images , making ,in For noise terms, Using spatial coordinates, the bias field is parameterized using B-splines, and alternating least-squares fitting and spatial smoothing constraints are performed in the logarithmic domain to obtain the corrected image. The corrected image has a more uniform overall intensity distribution, and the comparability of different spatial locations is significantly improved.

[0035] The second step is to perform spatial registration of images from multiple phases, sequences, and time points. Using the portal venous phase image at the pre-treatment baseline time point as a unified reference space, images from all other phases, sequences, and follow-up time points are aligned to this reference space. The registration employs a three-level progressive strategy: the first level is rigid registration based on normalized mutual information, finding the optimal rigid transformation matrix by optimizing six degrees of freedom parameters (three translation parameters and three rotation parameters); the second level is affine registration, adding scaling and shearing transformation parameters to the rigid registration result, resulting in twelve degrees of freedom, and optimizing the affine transformation matrix; the third level is deformable registration based on a B-spline free deformation model, setting a uniformly distributed control point grid within the reference space, with the displacement vector of each control point as the parameter to be optimized, and obtaining the displacement of each voxel through B-spline basis function interpolation, forming a dense deformation field. All three levels of registration use mutual information as a similarity measure, and its calculation formula is as follows: .in, A fixed image within the reference space, i.e., the baseline portal venous phase image; The moving image to be registered; Spatial transformation (rigid transformation, affine transformation, or deformable transformation) for the current level registration; This represents the result of spatial transformation of a moving image; The information entropy of a fixed image; The information entropy of the transformed moving image; This represents the joint information entropy of the two images. Each information entropy is estimated from the joint gray-level histogram of the corresponding image regions. Mutual information measures the statistical dependence between the two images. Maximizing mutual information is equivalent to making the transformed moving image as consistent as possible with the fixed image in gray-level statistics. This measure is particularly suitable for multimodal image registration. The transformation matrices and deformation fields generated at each registration stage are fully stored so that the voxel correspondence between any time points can be calculated in subsequent steps.

[0036] After spatial registration, two radiologists with experience in abdominal MRI diagnosis collaboratively delineated the three-dimensional regions of interest (ROIs) of the tumor at each time point on the registered images, layer by layer. The delineation primarily used abnormal enhancement areas on portal venous or delayed phase images as references, supplemented by T2-weighted images and apparent diffusion coefficient maps to confirm tumor boundaries, excluding surrounding normal liver parenchyma, major intrahepatic vessels, and bile duct structures. A semi-automatic segmentation method can also use a pre-trained three-dimensional U-shaped convolutional neural network model to provide initial ROIs, which are then manually corrected by physicians to improve efficiency and consistency. The segmentation results are stored as binary masks, labeling each voxel within the tumor.

[0037] Subsequently, Z-score intensity normalization was performed on the voxels within the tumor region of interest. First, statistics for the liver parenchyma background region needed to be obtained. The liver parenchyma background region was extracted from the registered image using an automatic liver segmentation algorithm, and non-liver parenchyma structures such as tumor regions, large blood vessels, and the gallbladder were excluded through morphological operations and distance transformation. For each sequence channel, the arithmetic mean of the intensities of all voxels within the liver parenchyma background region was calculated. and standard deviation Then, the intensity value of each voxel within the region of interest of the tumor is normalized using the following formula: .in, The original signal intensity of a voxel in a certain sequence channel. This represents the mean intensity of the liver parenchyma background region under this sequence channel. For the corresponding standard deviation, This represents the normalized voxel intensity value. After this transformation, the normalized intensity value has one dimension and its distribution approximates a standard normal distribution, making the image intensity comparable between different patients and different scanning devices.

[0038] Finally, using the centroid coordinates of the tumor's 3D region of interest as the center, the registered and normalized image is expanded outward by a predetermined number of voxels and cropped into fixed-size 3D cubes. If the tumor region exceeds the cropping boundary, it is truncated at the centroid using the largest inscribed cube; if the expanded region exceeds the image range, it is filled with zero values. After cropping, trilinear interpolation is performed to resample to isotropic voxel resolution. After the above preprocessing, a standardized multi-phase dynamic enhancement image block is obtained, with the shape of... ,in Period number, , , The number of voxels in the spatial dimension; and the standardized multi-sequence structure image patch, which is composed of fat-suppressed T2-weighted images and apparent diffusion coefficient maps stitched together in the channel dimension, with a shape of .

[0039] S2. Input the multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks into the pre-trained dual-branch decoupled encoder network to obtain statistically independent enhancement embedding feature maps and texture embedding feature maps. The dual-branch decoupled encoder network includes an enhancement coding branch and a texture coding branch. During training, the lower bound of the mutual information between enhancement embedding and texture embedding is minimized as a constraint to make the enhancement embedding feature map and texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features.

[0040] Specifically, in this embodiment, the dual-branch decoupled encoder network employs mutual information minimization constraints during the training phase, using a computationally efficient single-step alternating update method. The enhancement coding branch dynamically enhances image patches across multiple phases. As input, the data is downsampled and channel-amplified layer by layer by a 3D convolutional neural network, and the output is an enhanced embedding feature map. Texture coding branch uses multi-sequence structured image patches. As input, the texture embedding feature map is output after passing through another set of weighted independent 3D convolutional neural networks. Both branches of the encoder contain three spatial downsampling operations, each halving the spatial size through a convolution or pooling with a stride of 2. Therefore, the spatial size of the final output enhanced embedding feature map and texture embedding feature map is one-eighth of the original input, i.e., from... Shrink to The number of channels increases accordingly. The embedding vectors output by both branches have the same dimension, thus ensuring the spatial correspondence at the voxel level.

[0041] In each training iteration, a fixed number of spatial locations are first randomly sampled from the current batch of data within the common effective spatial region of the augmented embedding feature map and the texture embedding feature map. For the ... Extract the enhanced embedding vector from each sampling location. and texture embedding vector Those from the same spatial location The pairs are positive sample pairs, derived from the joint distribution. Then maintain... The order remains unchanged, for The indexes are randomly arranged to generate negative sample pairs. , used to approximate the product of marginal distributions; where, Indicates index A random arrangement.

[0042] Pre-built discriminator network It consists of multiple stacked fully connected layers, and its input is a concatenation of the enhancement embedding vector and the texture embedding vector. Output a scalar value. This network parameter is denoted as... Based on mutual information neural estimation theory, mutual information... A tight lower bound can be given by the following formula:

[0043]

[0044] in, This represents the lower bound of mutual information to estimate the loss value; The number of spatial locations sampled each time; This indicates that the discriminator network is for the first... The output value of each positive sample pair; To randomly shuffle the order and then... A texture embedding vector paired with an enhanced embedding vector forms a negative sample pair; For the natural constant An exponential function with base 0; It is the natural logarithm function. The first term of the formula... The positive sample pairs are the sample mean output by the discriminator, and their joint distribution expectation is estimated in a Monte Carlo manner. ,in This represents the true joint distribution of the augmented embedding vector and the texture embedding vector. Find the expected value; second term It is the logarithm of the negative sample pairs with the sample mean after exponential transformation, and it estimates the marginal logarithmic expectation. ,in Represents the product of marginal distributions Let's find the expected value. According to the Donsker-Varadhan representation theorem, the difference between these two terms constitutes a variational lower bound on the mutual information.

[0045] During training, for the discriminator network parameters To maximize Gradient ascent is applied to the target to optimize the discriminator, resulting in a tighter lower bound estimate of the mutual information. Subsequently, with the discriminator parameters fixed, the total loss function for both the augmented autoencoder and the textured autoencoder is a weighted combination of the mean squared reconstruction error of each modality and the mutual information penalty term. The reconstruction loss of the augmented autoencoder is:

[0046]

[0047] in, This indicates that the input multi-phase dynamic enhancement image patch is in phase. Spatial coordinates voxel value at the location; These are the voxel values ​​corresponding to the reconstructed image; Total number of phases; Traverse all spatial coordinates; The total number of elements in a multi-phase dynamic enhancement image patch. Reconstruction loss of the texture autoencoder. Defined in exactly the same form on a two-channel structured image patch, its total number of elements is .

[0048] The total loss function is ,in The balancing hyperparameters are positive. Minimize With the goal of improving the parameters of the enhancement coding branch and the texture coding branch, gradient descent is used to update them, thereby driving them to achieve good reconstruction capabilities and statistically independent embedding representations. This single-step alternating update strategy has low computational cost, effectively suppresses the lower bound of mutual information, and promotes the two subspaces to tend towards independence.

[0049] After training, the augmentation embedding feature map mainly captures the temporal intensity curve features of each voxel during the dynamic augmentation process, while the texture embedding feature map mainly captures the tissue-inherent properties of voxels in terms of T2 relaxation and diffusion restriction. The statistical dependence between the two is significantly weakened. This active decoupling within the feature space makes it possible to track augmentation mode changes and texture changes separately, overcoming the information dilution problem caused by the simple concatenation of multiple sequence features in traditional methods.

[0050] S3. Based on the enhanced embedding feature map, the tumor voxel is divided into multiple enhanced sub-regions, and the volume of each enhanced sub-region is determined; based on the texture embedding feature map, the tumor voxel is divided into multiple texture sub-regions, and the volume of each texture sub-region is determined.

[0051] Specifically, for the enhanced embedding feature map of the current patient, the enhanced embedding vectors corresponding to all voxels within the tumor region of interest are extracted. Let voxels be... The enhanced embedding vector is ,in To enhance the embedding dimension of the embedding vectors, we take the dataset containing all the enhanced embedding vectors as input and perform K-means clustering, using the cluster distance metric. Using the squared Euclidean distance, the expression is as follows:

[0052]

[0053] in, For the first The center vectors of the clusters, For vectors The dimensional components, for The Dimensional components. The number of clusters can be pre-determined based on common internal niche types in hepatocellular carcinoma, such as enhanced activity zones, hypovascular necrotic zones, and fibrotic transition zones. Each voxel is assigned to the category label with the smallest distance from each cluster center, forming an enhanced subregion label map. The number of voxels contained in each enhanced subregion category is counted and denoted as the volume of that enhanced subregion. , To enhance the indexing of subregions.

[0054] Texture subregion partitioning and enhancement subregions are performed completely independently. Texture embedding vectors for each voxel are extracted from the texture embedding feature map. Independently run K-means clustering to obtain the texture sub-region label map and the volume of each texture sub-region. , For the index of the texture subregion, This represents the embedding dimension of the texture embedding vector. Since the input features and initialization of the two clusters are independent, there is no one-to-one mapping between the boundaries and labels of the augmented subregion and the texture subregion. This lays the foundation for subsequently modeling the transformation relationship in the two independent subspaces.

[0055] S4. Construct an enhancement subspace graph with multiple enhancement subregions as nodes, and construct a texture subspace graph with multiple texture subregions as nodes. Establish spatial adjacency edges within the first time point and cross-time transformation edges between the first time point and the second time point in the enhancement subspace graph and texture subspace graph to form a time series graph.

[0056] Specifically, the enhanced subspace graph uses enhanced subregions as nodes. For each node... Its initial node feature vector Taken as the arithmetic mean of all voxel enhancement embedding vectors within the enhanced subregion, the calculation formula is: .in, Indicates enhanced subregion The set of voxels contained therein This represents the number of voxels in the set. This mean aggregation method preserves the overall enhanced dynamics of the subregion. Initial feature vectors of nodes in the texture subspace graph. Similarly, the mean of the texture embedding vector within the texture subregion is constructed.

[0057] Spatial adjacency edges are established based on voxel adjacency relationships in three-dimensional space. In the enhanced subregion label map, if the enhanced subregion... and If at least one pair of voxels satisfy a face adjacency relationship (i.e., two voxels share a voxel face), then at the nodes of the enhanced subspace graph... and Add an undirected spatial adjacency edge between them. Edge weight. The value is determined by the ratio of the number of adjacent voxel faces to the geometric mean of the volumes of the two subregions, and the calculation formula is as follows: .in, To enhance sub-regions and Number of adjacent voxel faces; and Enhanced subregions and The volume, expressed in voxels, is defined. The weights are normalized using the geometric mean of the volume to avoid large subregions naturally having higher edge weights, ensuring the balance of the graph structure. Spatial adjacency edges and their weights in the textured subspace graph are established in the same way.

[0058] Transitional edges are used to connect sub-region nodes at the first and second time points, characterizing the dynamic evolution of sub-region types. First, the voxel-level correspondence between the first and second time points needs to be obtained. Using the deformable registration deformation fields of each time point relative to the baseline stored in the preprocessing stage, a forward deformation field from the second time point to the first time point is generated through inverse transformation and composition operations. Let the deformation field mapped from the image coordinates at the first time point to the baseline be... The deformation field mapped from the image coordinates at the second time point to the baseline is: The deformation field mapped from the second time point to the first time point... It can be obtained through composition and inverse transformation, and the expression is: .in, Indicates composite operation of deformation field, superscript This represents the inverse transformation of the deformation field. Since the B-spline deformation field is a continuously differentiable dense mapping, both the inverse transformation and the composite operation can be solved through numerical optimization and iteration, ensuring the invertibility and consistency of the deformation field.

[0059] Utilizing deformation field The tumor region at the second time point is mapped to the space at the first time point. Within the space at the first time point, all voxels located within the tumor region are traversed. For each voxel... Read its enhanced subregion label at the first time point The position of the corresponding second time point is obtained through the deformation field. Read the enhanced subregion label of the location at the second time point. Then the elements of the corresponding enhanced transformation counting matrix will be... Add one. After traversing all voxels, perform row normalization on the counting matrix to obtain the enhanced empirical transformation probability matrix. The calculation formula is: .in, To increase the number of subregions. If the first... The sum of the rows (That is, the subregion does not exist at the first point in time), then set This is done by assigning a uniform, informationless prior. The resulting augmented empirical transformation probability is directly used as the augmented subregion at the first time point. Pointing to the second time point to enhance the subregion The weights of directed transtemporal transformation edges. The transtemporal transformation edges and their empirical transformation probability weights of the textured subspace graph are obtained using a completely symmetric process.

[0060] S5. Input the time series graph into the pre-trained graph sequence encoder and output the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-time transformation relationship of the sub-region.

[0061] Specifically, this embodiment uses a cascaded graph convolutional network and a long short-term memory network as a graph sequence encoder. At each time step, the graph convolutional network is first used to locally aggregate and update node features along spatially adjacent edges. The inter-layer propagation rule of the graph convolutional network is as follows: .in, For the first The node feature matrix of a layer, with rows representing nodes and columns representing feature dimensions; The matrix is ​​a row-stacked matrix of initial node features, consisting of the enhanced node feature vectors or texture node feature vectors constructed in step S4; To add the adjacency matrix after adding self-loops, elements in The corresponding spatial adjacency edge weight or (Weight of spatial adjacency edges in the texture subspace graph). It is the identity matrix; for The degree matrix, whose diagonal elements ,in for Element; The adjacency matrix is ​​symmetric and normalized. This normalization ensures that the feature scale of nodes of different degrees remains stable during message passing. For the first The learnable linear transformation matrix of the layer; The activation function is a non-linear one, using a modified linear unit. By stacking two layers of graph convolutional networks, each node can aggregate information from its second-order neighborhood.

[0062] Subsequently, the graph convolutional update features of each node at two time steps are input into the Long Short-Term Memory (LSTM) network in chronological order for temporal evolution modeling. The hidden state update of the LSM network follows a standard gating mechanism, expressed as: , , , , , .in, For time step The graph convolution outputs the node feature vectors. For time step The hidden state, For time step cellular state; and These represent the hidden state and cellular state of the previous time step, respectively. , , They are time steps The activation values ​​of the forget gate, input gate, and output gate range from 0 to 1; For time step Candidate cell state; It is a sigmoid function; Represents element-wise product; , , , This is the corresponding input weight matrix. , , , This is the corresponding cyclic weight matrix. , , , Here are the corresponding bias vectors, all of which are learnable parameters. The initial hidden state at the first time step. and initial cell state It is obtained by mapping the graph convolution output from the first time step using a single-layer fully connected network, using the hyperbolic tangent activation function.

[0063] After processing by the Long Short-Term Memory (LSTM) network, each node obtains a hidden state that integrates spatial and temporal context information at the first and second time steps, respectively. For the augmented subspace graph, each augmented subregion node at the first time step... Hidden state With each enhanced subregion node in the second time step Hidden state The network is concatenated, taking a single-layer fully connected network as input and outputting a scalar augmentation conversion tendency score. For the first time step node All possible transformation targets Perform softmax normalization, the formula is as follows: .in, It is a natural exponential function. To predict the enhanced subregion transformation probability, nonnegativity and row normalization constraints must be satisfied; To increase the number of subregions; This indicates that the first step is to enhance the sub-region. Enhanced subregions in the second time step The transformation tendency score. The texture transformation probability matrix is ​​predicted by independent networks with the same structure according to a symmetric process, and the parameters are not shared.

[0064] The training of the graph sequence encoder uses the empirical transition probability calculated in step S4 as the supervision target, and the loss function is cross-entropy.

[0065]

[0066] in, Iterate through all consecutive time pairs. The number of subregions in the corresponding subspace (in the enhanced subspace) In the texture subspace ); For time points The probability of empirical transformation under the given conditions; The corresponding predicted conversion probability is given. The optimizer uses adaptive moment estimation, and the weights are initialized using Ho's initialization method. Training continues until the loss converges.

[0067] S6. Calculate the enhanced conversion rate based on the enhanced conversion probability matrix and the volume of the enhanced subregion, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture subregion, and generate early efficacy assessment results based on the difference between the enhanced conversion rate and the texture conversion rate.

[0068] Specifically, enhance conversion rate The purpose is to measure the overall intensity of the redistribution of enhancement patterns throughout the tumor, and its calculation formula is as follows: ,in To increase the number of subregions; For the first The volume of each enhanced subregion; To enhance the first transformation probability matrix Line 1 Diagonal elements represent enhanced subregions. The probability of maintaining its own type. This represents the probability that the enhanced subregion transforms into another type. The numerator is the weighted sum of the volumes of each enhanced subregion and its transformation probability, and the denominator is the sum of the volumes of all enhanced subregions. Therefore, This is the volume-weighted average conversion probability, with a value between 0 and 1. The larger the value, the more drastic the redistribution of enhancement subregion types and the more significant the overall change in enhancement patterns.

[0069] Similarly, texture conversion rate The calculation formula is: ,in The number of texture subregions; For the texture subregion volume; These are the diagonal elements of the texture transformation probability matrix. The physical meaning and Perfectly symmetrical.

[0070] The difference between the enhancement conversion rate and the texture conversion rate is defined as the enhancement-texture conversion difference index. Theoretically, this index can range from -1 to 1. Its biological explanation lies in the fact that during targeted combined immunotherapy for hepatocellular carcinoma, the vascular normalization effect induced by anti-angiogenic drugs appears relatively early, mainly manifested as tumor-feeding artery remodeling and decreased microvascular permeability. These changes are first reflected in dynamic contrast-enhanced MRI images as changes in contrast agent inflow slowing, peak enhancement decreasing, and clearance patterns altering the enhancement dynamics. Tumor cell necrosis, apoptosis, and subsequent fibrosis are relatively delayed, requiring a longer timeframe to cause significant changes in T2 signal and diffusion restriction characteristics. Therefore, if the transformation rate is enhanced... Significantly higher than texture conversion rate A large positive ETDI value indicates that the enhancement pattern has significantly redistributed while tissue texture has not yet had time to respond. This is an early imaging marker of the effectiveness of targeted combined immunotherapy. ETDI serves directly as a continuous early efficacy assessment result; the higher the value, the greater the likelihood of a treatment response. It provides clinicians with a quantitative tool to determine treatment benefit without waiting for significant tumor shrinkage. The introduction of this differential indicator cleverly utilizes the dual-channel information brought about by decoupled characterization and independent transformational modeling to accurately capture the temporal differences between vascular and tissue effects. This is a technical effect that traditional single-modality analysis or simple feature fusion methods cannot achieve.

[0071] The aforementioned method for early assessment of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging utilizes a pre-trained bi-branch decoupled encoder network. Under the constraint of minimizing the lower bound of mutual information, it maps multi-phase dynamic enhancement images and multi-sequence structural images into statistically independent enhancement embedding feature maps and texture embedding feature maps, thereby actively separating the entangled contrast agent dynamic features and tissue texture features in the feature space in traditional methods. Based on the decoupled representation, the enhancement embedding and texture embedding are sub-regionally divided, and a time-series graph that simultaneously carries spatial topology and temporal evolution information is constructed using each sub-region as a node and spatial adjacency and cross-temporal voxel transformation relationships as edges. Subsequently, a graph sequence encoder is used to jointly encode the spatial adjacency edges and cross-temporal transformation edges, independently outputting the enhancement sub-region transformation probability matrix and the texture sub-region transformation probability matrix, respectively capturing the enhancement pattern redistribution driven by vascular effects and the texture pattern redistribution driven by tissue necrosis. Finally, the difference between the volume-weighted enhancement conversion rate and the texture conversion rate is used as a quantitative indicator of early efficacy. This scheme uses a technical approach of "feature space decoupling - independent sub-region evolution modeling - conversion rate difference quantification" to enable the vascular normalization effect and tissue necrosis effect to be tracked and compared separately in their respective independent subspaces. This allows for sensitive detection of the early response to hepatocellular carcinoma conversion therapy, overcoming the technical defects of existing methods where the early weak therapeutic signal is diluted due to feature entanglement and the inability to distinguish the source and direction of response diffusion.

[0072] Example 2:

[0073] In this embodiment, the training process of the dual-branch decoupled encoder network includes the following steps:

[0074] S11. Input the multi-phase dynamic enhancement image block samples of the training samples into the enhancement coding branch, and input the multi-sequence structure image block samples of the training samples into the texture coding branch to obtain the training enhancement embedding feature map and the training texture embedding feature map, respectively.

[0075] Specifically, the training samples consist of multiple MRI scans from a large number of hepatocellular carcinoma patients across multiple medical centers throughout the entire conversion therapy process, covering pre-treatment baseline and multiple post-treatment time points. The enhancement coding branch employs an enhancement dynamic encoder architecture, which adds a temporal self-attention aggregation module on top of the basic 3D convolutional encoder in Example 1. This module first uses a shared-weight 3D residual network to independently extract spatial feature maps of each phase, then stacks all phase feature maps along the temporal dimension, and performs multi-head self-attention encoding along the time axis for each downsampled spatial location. Specifically, at a certain spatial location, the feature vectors of each phase are considered as a time series with a sequence length of [missing information]. Each vector's dimension is the number of channels in the spatial features. A learnable aggregation token vector is added to the beginning of this time series. This vector is a trainable parameter with the same dimension as the phase feature vector, acting similarly to a classification token in a visual Transformer. It interacts with features from all phases through a self-attention mechanism, thereby summarizing global temporal information. The query, key, and value matrices for multi-head self-attention are calculated from the sequence after adding the aggregation token through different linear projections. The attention weights are calculated using scaled dot products and normalized using softmax. After multi-layer self-attention encoding, the output vector corresponding to this aggregation token vector serves as the enhanced embedding vector for that spatial location, with a dimension of [missing information]. The texture encoding branch employs a structured texture encoder, which can specifically utilize a 3D densely connected convolutional network. Within each dense block, layers are densely connected using channel-stitched connections. The growth rate controls the rate of increase in the number of channels in the feature map. Between dense blocks, transition layers containing batch normalization, corrected linear units, pointwise convolutions, and average pooling are used for downsampling and channel compression. The final output is a texture embedding feature map with the same spatial size as the enhanced embedding feature map.

[0076] S12. Extract training augmented embedding vectors at multiple spatial locations from the training augmented embedding feature map, and extract training texture embedding vectors at spatial locations corresponding to the training augmented embedding vectors from the training texture embedding feature map.

[0077] Specifically, in a training mini-batch, let the shape of the enhanced embedding feature map be... The texture embedding feature map shape is ,in For batch size, , , This represents the spatial size after downsampling. Within the common effective spatial region (avoiding filled regions), a fixed number of spatial indices are randomly sampled, and for each index... Extract the enhanced embedding vector and texture embedding vector Both dimensions satisfy .

[0078] S13. Pair training augmentation embedding vectors with training texture embedding vectors at the same spatial location to obtain positive sample pairs; pair training augmentation embedding vectors with training texture embedding vectors at randomly shuffled spatial locations to obtain negative sample pairs.

[0079] Specifically, positive sample pairs are directly derived from the same index. Composition. When generating negative sample pairs, a random permutation of the index is generated. This will enhance the embedding vector With texture embedding vector Pairing, forming Because the arrangement disrupts the original spatial alignment, negative sample pairs approximate the product of marginal distributions.

[0080] S14. Based on the positive and negative sample pairs, calculate the first mutual information lower bound estimation loss value using the pre-built discriminator network, and update the parameters of the discriminator network by maximizing the first mutual information lower bound estimation loss value; wherein, the input of the discriminator network is the concatenation result of the training augmented embedding vector and the training texture embedding vector, and the output is the mutual information estimation value used to estimate the mutual information lower bound between the training augmented embedding vector and the training texture embedding vector.

[0081] Specifically, discriminator network This can be a four-layer fully connected feedforward neural network. The first layer has an input dimension of... Receives the concatenation of the enhanced embedding vector and the texture embedding vector. The second and third layers are hidden layers, each followed by a modified linear unit activation with leakage. The negative slope is fixed at a small value to allow negative gradient flow. The fourth layer is the output layer, containing only one neuron, with no activation function, and outputting a scalar. The first mutual information lower bound estimate of the loss value is calculated using the mutual information neural estimation lower bound formula:

[0082]

[0083] in, This represents the lower bound of mutual information to estimate the loss value; The number of spatial locations sampled each time; This indicates that the discriminator network is for the first... The output value of each positive sample pair; To randomly shuffle the order and then... A texture embedding vector paired with an enhanced embedding vector; For the natural constant An exponential function with base 0; Let be the natural logarithm function. During this discriminator update phase, the parameters of the augmentation coding branch and the texture coding branch remain fixed, and are only maximized... To update the discriminator parameters This embodiment further employs a multi-step discriminator update strategy. Before each encoder update, the discriminator is updated multiple times to ensure that the discriminator always maintains a relatively tight estimate of the lower bound of mutual information, thereby providing the encoder with a more accurate mutual information gradient signal.

[0084] S15. Based on the updated discriminator network, calculate the second mutual information lower bound estimation loss value for positive sample pairs and negative sample pairs, and update the parameters of the enhancement coding branch and the texture coding branch by minimizing the second mutual information lower bound estimation loss value.

[0085] Specifically, after the discriminator completes its multi-step update, the current batch of data is recalculated. As a mutual information penalty term, the reconstruction loss of the augmented autoencoder is used. Reconstruction loss of texture autoencoder All calculations are based on mean squared error, and the formula is the same as in Example 1. The total loss function is... .in, To enhance the reconstruction loss of the autoencoder, For the reconstruction loss of the texture autoencoder, The balancing hyperparameters are positive. At this point, the discriminator parameters are fixed to minimize... Update the parameters of the enhancement coding branch and texture coding branch for the target. Minimize. This causes the lower bound of mutual information between the augmented embedding and the texture embedding to continuously decrease, driving the feature representations learned by the two branches to become statistically independent, while the constraint of the reconstruction loss ensures that the decoupling process does not discard the inherent information structure within each modality.

[0086] S16. Repeat S14 and S15 until the preset convergence condition is met to obtain the trained dual-branch decoupled encoder network; wherein, the first mutual information lower bound estimation loss value and the second mutual information lower bound estimation loss value are calculated according to the following formula:

[0087]

[0088] in, This represents the lower bound of mutual information to estimate the loss value. The number of spatial locations sampled in each batch. This represents the discriminator network. Indicates the first positive sample pair Training augmented embedding vectors for each spatial location, Indicates the first positive sample pair Training texture embedding vectors at spatial locations, Indicates the negative sample pair and the first The training texture embedding vectors corresponding to each spatial location This represents an exponential function with the natural constant as its base. This represents the natural logarithm function.

[0089] Specifically, the discriminator undergoes repeated, alternating multi-step updates and encoder single-step updates until a preset convergence condition is met, such as the mutual information estimate on the validation set changing less than a preset threshold over multiple consecutive training epochs, or the total loss no longer decreasing. Training uses the AdamW optimizer and employs a cosine annealing learning rate decay strategy. The enhanced encoding branch and texture encoding branch obtained after training constitute the final two-branch decoupled encoder network.

[0090] To further ensure the adequacy of decoupling, this embodiment can also introduce cross-modal reconstruction verification as a diagnostic method. Two additional structurally symmetric cross-modal decoders are constructed: the first cross-modal decoder takes the enhanced embedding feature map as input, restores resolution through transposed convolution or interpolation upsampling, and outputs reconstructed multi-sequence structured image patches; the second cross-modal decoder takes the texture embedding feature map as input and outputs reconstructed multi-phase dynamically enhanced image patches. During training, the encoder parameters are frozen on the validation set at regular intervals, and the two cross-modal decoders are trained separately until convergence. Then, the peak signal-to-noise ratio (PSNR) between the reconstructed image and the real image is calculated. The formula for calculating PSNR is... Where MAX is the theoretical maximum value of image intensity, and MSE is the mean square error between the reconstructed image and the real image. If the peak signal-to-noise ratio obtained by any cross-modal decoder is consistently higher than the preset level, it indicates that cross-modal decodeable information still remains in the corresponding embedding, and the decoupling is not sufficient. In this case, the mutual information penalty weight is increased. The alternating training process described above is repeated until the peak signal-to-noise ratio of cross-modal reconstruction decreases below the threshold. This closed-loop feedback mechanism ensures that the statistical independence of the decoupled encoder output embedding is fully guaranteed. This verification module is not a simple evaluation metric, but rather directly intervenes in the optimization process by dynamically adjusting the training hyperparameters, forming a decoupled training paradigm with self-correcting capabilities.

[0091] Example 3:

[0092] Based on the enhancement embedding feature map, the tumor voxels are divided into multiple enhancement subregions, and the volume of each enhancement subregion is determined; based on the texture embedding feature map, the tumor voxels are divided into multiple texture subregions, and the volume of each texture subregion is determined, including the following steps:

[0093] S21. Input the multi-phase dynamic enhancement image patch samples and multi-sequence structure image patch samples of all historical patients in the training set before treatment into the dual-branch decoupled encoder network to obtain the training enhancement embedding feature map and the training texture embedding feature map.

[0094] Specifically, using the trained augmented coding branch and texture coding branch, forward inference is performed on preprocessed image patches of each historical patient in the training set at the pre-treatment baseline time point to obtain the training augmented embedding feature map and training texture embedding feature map for each patient. This training set, like the training samples in Example 2, consists of multiple MRI scans of a large number of hepatocellular carcinoma patients from multiple medical centers throughout the conversion therapy process. The pre-treatment time point was chosen as a reference because it is unaffected by treatment intervention and reflects the original niche distribution of the tumor before disturbance.

[0095] S22. Perform K-means clustering on the voxels in the training augmented embedding feature map to obtain multiple global augmented cluster centers; perform K-means clustering on the voxels in the training texture embedding feature map to obtain multiple global texture cluster centers.

[0096] Specifically, the augmented embedding vectors of all voxels located within the tumor region in the pre-treatment augmented embedding feature maps of all trained patients are aggregated into a global augmented embedding vector set. Let this vector set contain... indivual A dimensional vector. K-means clustering is performed on this vector set, with each cluster run independently and randomly initialized multiple times to reduce the influence of local minima. Number of clusters. The silhouette coefficients are automatically determined by iterating through a finite number of candidate values ​​and calculating the silhouette coefficients for each clustering result. For a given clustering result, the mean silhouette coefficient of all voxels is:

[0097]

[0098] in, voxels The average distance to other voxels within its cluster. voxels The minimum average distance to all voxels in the nearest other cluster. This is the average profile coefficient. Choose to make... Maximize Value (i.e., the number of augmented subregions). After clustering converges, the center vector of each cluster is extracted. ,in These center vectors constitute the global definition template for the enhanced subregions. Similarly, K-means clustering is performed independently on the global texture embedding vector set and automatically determined. (i.e., the number of texture sub-regions) yields the global texture cluster center. ,in Two sets of global cluster centers are persistently stored as a unified reference for the subsequent subregional division of all new patients.

[0099] S23. For the current patient, extract the enhancement embedding vector of each voxel from the enhancement embedding feature map, calculate the first Euclidean distance between the enhancement embedding vector and each global enhancement cluster center; assign the voxel to the global enhancement cluster center with the smallest first Euclidean distance to obtain the enhancement sub-region, and count the volume of each enhancement sub-region.

[0100] Specifically, for the enhanced embedding feature map of the current patient, each voxel within the tumor region Enhanced embedding vector The subregion label is determined by the following formula:

[0101]

[0102] in, voxels The enhanced subregion label, argmin, represents taking the value that minimizes the sum of squares that follows. value; To increase the dimension of the embedding vector; For vectors The Dimensional components; For the first Globally enhanced cluster centers The Dimensional components. This assignment process is a deterministic nearest neighbor classification, requiring no iteration. Due to the fixed global cluster centers, subregions with the same label value across different patients exhibit consistent enhanced pattern semantics. Statistical analysis of each label... The corresponding number of voxels, i.e., the volume of the enhanced subregion. .

[0103] S24. For the current patient, extract the texture embedding vector of each voxel from the texture embedding feature map, calculate the second Euclidean distance between the texture embedding vector and each global texture cluster center; assign the voxel to the global texture cluster center with the smallest second Euclidean distance to obtain the texture sub-region, and count the volume of each texture sub-region.

[0104] Specifically, the division of texture sub-regions is completely symmetrical to S23. For texture embedding vectors... Its texture sub-region labels Determined by the following formula:

[0105]

[0106] in, This represents the number of global texture cluster centers. The dimension of the texture embedding vector; For vectors The Dimensional components; For the first Global texture clustering center The Dimensional components. The texture label map and the corresponding volumes of each texture sub-region are obtained. This global clustering center strategy not only ensures the semantic consistency of the labels but also makes the sub-region divisions generated by the same patient at different follow-up time points directly comparable, providing a stable and transferable node definition foundation for the subsequent construction of cross-time transformation maps. In contrast, the labels of single-case independent clustering in Example 1 may correspond to completely different niches among different patients, which is not conducive to cross-individual generalization. This embodiment effectively improves the standardization of sub-region division by introducing population statistical priors, and its implementation effect is a significant enhancement to the generalizability of this invention.

[0107] Example 4:

[0108] refer to Figure 2 To form a time series graph, spatial adjacency edges within the first time point and cross-time transformation edges between the first and second time points are established in the enhanced subspace graph and the texture subspace graph, including the following steps:

[0109] S31. For each node in the augmented subspace graph, calculate the augmented embedding mean vector based on the augmented embedding vector located in the corresponding augmented subregion in the augmented embedding feature graph; calculate the augmented geometric feature vector based on the geometric properties of the augmented subregion; concatenate the augmented embedding mean vector with the augmented geometric feature vector to obtain the augmented node feature vector.

[0110] Specifically, enhance sub-regions Enhanced node feature vectors Enhanced embedding mean vector and enhanced geometric eigenvectors It is composed of splicing parts. The calculation method is the same as in Example 1, namely, enhancing the sub-region. The arithmetic mean of all voxel-enhanced embedding vectors. It can include the following geometric properties: volume (Volumetric number), firmness (in The values ​​represent the surface area of ​​the subregion (approximated by the number of voxel faces), three normalized coordinates of the centroid relative to the center of the image patch, and the minimum distance from the edge of the subregion to the edge of the overall tumor. All geometric properties are Z-score normalized to eliminate dimensional differences.

[0111] S32. For each node in the texture subspace graph, calculate the texture embedding mean vector based on the texture embedding vector located in the corresponding texture subregion in the texture embedding feature graph; calculate the texture geometric feature vector based on the geometric properties of the texture subregion; concatenate the texture embedding mean vector with the texture geometric feature vector to obtain the texture node feature vector.

[0112] Specifically, texture subregions Texture node feature vectors Constructed in the same manner, where the texture embeds the mean vector. For texture subregions The arithmetic mean of all voxel texture embedding vectors, texture geometric feature vectors Includes geometric properties corresponding to the enhanced side.

[0113] S33. Based on the three-dimensional spatial position of each enhancement sub-region in the enhancement subspace graph, establish a first spatial adjacency edge between any two enhancement sub-regions that share at least one adjacent voxel; based on the three-dimensional spatial position of each texture sub-region in the texture subspace graph, establish a second spatial adjacency edge between any two texture sub-regions that share at least one adjacent voxel.

[0114] Specifically, the establishment of the first spatial adjacency edge is based on the adjacency of the voxel surface (6-neighborhood connection), and the second spatial adjacency edge is established in the same way.

[0115] S34. Based on the number and volume of adjacent voxel faces of the two enhanced subregions connected by each first spatial adjacent edge in the enhanced subspace graph, calculate the weight of the first spatial adjacent edge; based on the number and volume of adjacent voxel faces of the two texture subregions connected by each second spatial adjacent edge in the texture subspace graph, calculate the weight of the second spatial adjacent edge; wherein, the calculation formula for the weights of the first and second spatial adjacent edges is as follows:

[0116]

[0117]

[0118] in, Indicates at a point in time Lower Enhanced Subregion With enhanced subregions The weights of the first spatial adjacent edges between them. Indicates enhanced subregion With enhanced subregions At the point of time The number of adjacent voxel faces, and Representing enhanced subregions and enhance sub-regions At the point of time Volume; Indicates at a point in time Subtexture subregion With texture subregion The weights of the second-space adjacent edges between them. Represents texture subregions With texture subregion At the point of time The number of adjacent voxel faces, and Representing texture subregions respectively and texture subregion At the point of time The volume.

[0119] Specifically, the creativity of this weight definition lies in the design of the exponent in the denominator: simply using the square root of the volume product results in insufficient normalization when volume differences are extremely large; using the volume product itself over-penalizes sub-region pairs with volume matching. The power, as an exponent, originates from the scale relationship between volume and surface area in three-dimensional space—under the assumption of ideal similar shapes, the surface area and volume are related by their scale. The normalization is proportional to the power of the power, so it can effectively eliminate the influence of volume scale on edge weights, so that the weights only reflect the relative size of the contact surfaces, thus ensuring the rationality of the graph topology.

[0120] S35. Perform deformable registration on the multi-phase dynamic enhancement image blocks at the first time point and the multi-phase dynamic enhancement image blocks at the second time point to obtain the voxel correspondence between the tumor region at the first time point and the tumor region at the second time point.

[0121] Specifically, the baseline-referenced deformation field stored during the preprocessing stage is utilized. and The forward deformation field from the second time point to the first time point is generated through inverse transformation and composition operation. The inverse transformation of the deformation field is solved through iterative optimization, with the objective function being to minimize the displacement residuals at each control point under the composite transformation.

[0122] S36. Based on the voxel correspondence, count the number of first voxel transformations from the enhanced subregion at the first time point to the enhanced subregion at the second time point; use the number of first voxel transformations to form a first transformation counting matrix, perform row normalization on the first transformation counting matrix to obtain the enhanced experience transformation probability matrix; extract the enhanced experience transformation probability from the enhanced subregion at the first time point to the enhanced subregion at the second time point from the enhanced experience transformation probability matrix.

[0123] Specifically, within the space at the first time point, all tumor voxels are traversed, and the enhanced subregion label of the voxel at the first time point is set as follows: The enhanced subregion label at the second time point after deformation field mapping is: For counting matrix elements Increment by one. The row normalization expression is: .in, To increase the probability of experience conversion; To increase the number of subregions, if the row sum is zero, assign a uniform value. .

[0124] S37. Based on the voxel correspondence, count the number of second voxel transformations from the texture sub-region at the first time point to the texture sub-region at the second time point; use the number of second voxel transformations to form a second transformation counting matrix, perform row normalization on the second transformation counting matrix to obtain the texture empirical transformation probability matrix; extract the texture empirical transformation probability from the texture sub-region at the first time point to the texture sub-region at the second time point from the texture empirical transformation probability matrix.

[0125] Specifically, the construction process of the texture experience transformation probability matrix is ​​completely symmetrical. Let the texture sub-region label of the voxel at the first time point be... The texture subregion label at the second time point after deformation field mapping is... Texture experience conversion probability The calculation formula is the same as that for the enhanced side, only the formula is changed to... Replace with , Replace with texture conversion counting matrix elements .

[0126] S38. Establish a fully connected first time-transformation edge between the enhanced subregion at the first time point and the enhanced subregion at the second time point in the enhanced subspace graph, and use the corresponding enhanced empirical transformation probability as the weight of the first time-transformation edge; establish a fully connected second time-transformation edge between the texture subregion at the first time point and the texture subregion at the second time point in the texture subspace graph, and use the corresponding texture empirical transformation probability as the weight of the second time-transformation edge.

[0127] Specifically, the first cross-time transformation edge is a set of directed edges pointing from the first time point node to the second time point node, with edge weights... The second time-transformation edge follows the same principle, with different weights. .

[0128] S39. The enhanced subspace graph containing the first spatial adjacent edges and their weights, the first cross-time transformation edges and their weights, and the enhanced node feature vectors is taken as the first time series subspace graph corresponding to the enhanced subspace graph; the texture subspace graph containing the second spatial adjacent edges and their weights, the second cross-time transformation edges and their weights, and the texture node feature vectors is taken as the second time series subspace graph corresponding to the texture subspace graph; the time series graph is composed of the first time series subspace graph and the second time series subspace graph.

[0129] Specifically, the final constructed time-series graph fully includes node features, spatial adjacency relationships, and cross-temporal transformation priors in both the enhancement and texture subspaces, providing rich structured input for the subsequent graph sequence encoder.

[0130] Example 5:

[0131] The time series image is input into a pre-trained graph sequence encoder, which outputs an enhancement transformation probability matrix for the enhanced sub-region from the first time point to the second time point, and a texture transformation probability matrix for the texture sub-region from the first time point to the second time point. This includes the following steps:

[0132] S41. For each time point in the time series graph, the graph attention network is used to perform spatial dimension message passing on the feature vector of the enhanced node along the first spatial adjacent edge and based on the weight of the first spatial adjacent edge, so as to obtain the spatially updated enhanced node feature vector; the graph attention network is used to perform spatial dimension message passing on the feature vector of the texture node along the second spatial adjacent edge and based on the weight of the second spatial adjacent edge, so as to obtain the spatially updated texture node feature vector.

[0133] Specifically, the spatial attention coefficient of the enhanced subspace graph is calculated as follows:

[0134]

[0135]

[0136] in, and Enhanced nodes and Node feature vectors; To share the linear transformation matrix; For learnable attention vectors, Indicates the transpose operation; This represents vector concatenation; LeakyReLU is a modified linear unit with leakage. This represents the weight of the adjacent edges in the first spatial region. is the learnable edge weight scaling factor, with a dimension of one; Indicates an enhanced node right Attention score; For nodes The set of spatial neighbors (including self-loops); To enhance nodes right The normalized attention coefficients are calculated. This step employs a multi-head attention mechanism, where each head calculates independently and then concatenates the results to obtain the spatially updated enhanced node feature vectors. The spatial update of the texture subspace graph is performed similarly.

[0137] S42. Using the graph attention network along the first time-transformation edge and based on the weight of the first time-transformation edge, the feature vector of the spatial update enhancement node at the first time point is passed to each enhancement sub-region node at the second time point to obtain the feature vector of the spatiotemporal fusion enhancement node; using the graph attention network along the second time-transformation edge and based on the weight of the second time-transformation edge, the feature vector of the spatial update texture node at the first time point is passed to each texture sub-region node at the second time point to obtain the feature vector of the spatiotemporal fusion texture node.

[0138] Specifically, cross-time attention uses the same attention coefficient calculation formula as S41, the neighbor set is all nodes at the second time point, and the edge weights use cross-time transformed edge weights.

[0139] S43. Input the spatiotemporal fusion enhancement node feature vectors into the gated loop unit in chronological order to obtain the enhancement node hidden state at each time step; input the spatiotemporal fusion texture node feature vectors into the gated loop unit in chronological order to obtain the texture node hidden state at each time step.

[0140] Specifically, the state update formula for the gated loop unit is: , , , .in, For time step The spatiotemporal fusion feature vector; For time step The hidden state, This is the hidden state of the previous time step; For time step Reset the door activation value; For time step Update the gate activation value; For time step Candidate hidden states; It is a sigmoid function; Represents element-wise product; , , This is the corresponding input weight matrix. , , This is the corresponding cyclic weight matrix. , , These are the corresponding bias vectors, all of which are learnable parameters. Initial hidden state It is obtained by passing the input features from the first time step through a fully connected layer and a hyperbolic tangent activation mapping.

[0141] S44. For each enhanced sub-region at the first time point, the hidden state of the enhanced node at the first time point is concatenated with the hidden state of the enhanced node in each enhanced sub-region at the second time point to obtain the first concatenation result; the first concatenation result is input into the multilayer perceptron to obtain the enhanced conversion tendency score; the enhanced conversion tendency score is normalized by softmax to obtain the enhanced conversion probability matrix.

[0142] Specifically, the multilayer perceptron is a two-layer fully connected network. The first layer is followed by a modified linear unit activation, and the second layer outputs a scalar. The expression for softmax normalization is: .in, To enhance sub-regions arrive The conversion tendency score This represents the corresponding predicted enhanced conversion probability; To increase the number of subregions.

[0143] S45. For each texture sub-region at the first time point, the hidden state of the texture node at the first time point is concatenated with the hidden state of the texture node in each texture sub-region at the second time point to obtain a second concatenation result; the second concatenation result is input into the multilayer perceptron to obtain the texture transformation tendency score; the texture transformation tendency score is normalized by softmax to obtain the texture transformation probability matrix.

[0144] Specifically, the calculation process on the texture side is completely symmetrical to that of S44, using independent network parameters.

[0145] Example 6:

[0146] The enhancement conversion rate is calculated based on the enhancement conversion probability matrix and the volume of the enhancement subregion; the texture conversion rate is calculated based on the texture conversion probability matrix and the volume of the texture subregion; and an early efficacy assessment result is generated based on the difference between the enhancement conversion rate and the texture conversion rate, including the following steps:

[0147] S51. Calculate the enhancement conversion rate based on the diagonal elements of the enhancement conversion probability matrix and the volume of the enhancement subregion; where the enhancement conversion rate characterizes the proportion of enhancement subregions undergoing type change; the formula for calculating the enhancement conversion rate is:

[0148]

[0149] in, This indicates an increased conversion rate. To increase the number of subregions, Indicates the first The volume of each enhanced subregion To enhance the first in the transformation probability matrix Line 1 The diagonal elements of the column represent enhanced subregions. The probability of remaining unchanged.

[0150] Specifically, in this formula, That is, the probability that the enhanced subregion transforms into other types; the numerator is the weighted sum of the volume of each enhanced subregion and its transformation probability, and the denominator is the sum of the volumes of all enhanced subregions; The larger the value, the more intense the redistribution of subregion types.

[0151] S52. Calculate the texture transformation rate based on the diagonal elements of the texture transformation probability matrix and the volume of the texture sub-region; where the texture transformation rate characterizes the proportion of texture sub-regions undergoing type changes; the formula for calculating the texture transformation rate is:

[0152]

[0153] in, Indicates the texture conversion rate. The number of texture subregions. Indicates the first The volume of each texture subregion; The th in the texture transformation probability matrix Line 1 The diagonal elements of the column represent texture subregions. The probability of remaining unchanged.

[0154] Specifically, in this formula, That is, the probability that the texture subregion transforms into other types; the numerator is the weighted sum of the volume of each texture subregion and its transformation probability, and the denominator is the sum of the volumes of all texture subregions; The larger the value, the more severe the redistribution of texture subregion types.

[0155] S53. Calculate the difference between the enhancement conversion rate and the texture conversion rate to obtain the enhancement-texture conversion difference index, and generate early efficacy assessment results based on the enhancement-texture conversion difference index.

[0156] Specifically, the enhancement-texture conversion difference index is defined as follows: ,in To enhance the conversion rate, This represents the texture transformation rate. This index is directly output as a result of continuous early efficacy assessment.

[0157] Furthermore, to locate the response origin region in the spatial dimension, a local enhancement-texture transformation difference index can be calculated for each voxel. For voxels... Let its enhanced subregion label at the first time point be... Texture sub-region label is ,but: .in, voxels Local enhancement-texture transformation difference index; To enhance sub-regions The probability of maintaining itself unchanged; For texture subregions The probability of maintaining itself unchanged. A positive LETDI value indicates that the stability of the texture subregion where the voxel is located is higher than that of the enhancement subregion, meaning that the enhancement pattern is more likely to change first than the texture pattern, and this voxel is a potential early response site. Three-dimensional pseudo-color rendering of the LETDI values ​​of all voxels generates an early therapeutic response heatmap. The highlighted areas in the heatmap represent the subregions where the vascular normalization effect first appears. This spatial heatmap serves as a visual supplement to the early therapeutic assessment results, providing clinicians with intuitive quantitative evidence to determine the origin and direction of the response and to formulate subsequent local intervention strategies.

[0158] The aforementioned method for early assessment of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging utilizes a pre-trained bi-branch decoupled encoder network. Under the constraint of minimizing the lower bound of mutual information, it maps multi-phase dynamic enhancement images and multi-sequence structural images into statistically independent enhancement embedding feature maps and texture embedding feature maps, thereby actively separating the entangled contrast agent dynamic features and tissue texture features in the feature space in traditional methods. Based on the decoupled representation, the enhancement embedding and texture embedding are sub-regionally divided, and a time-series graph that simultaneously carries spatial topology and temporal evolution information is constructed using each sub-region as a node and spatial adjacency and cross-temporal voxel transformation relationships as edges. Subsequently, a graph sequence encoder is used to jointly encode the spatial adjacency edges and cross-temporal transformation edges, independently outputting the enhancement sub-region transformation probability matrix and the texture sub-region transformation probability matrix, respectively capturing the enhancement pattern redistribution driven by vascular effects and the texture pattern redistribution driven by tissue necrosis. Finally, the difference between the volume-weighted enhancement conversion rate and the texture conversion rate is used as a quantitative indicator of early efficacy. This scheme uses a technical approach of "feature space decoupling - independent sub-region evolution modeling - conversion rate difference quantification" to enable the vascular normalization effect and tissue necrosis effect to be tracked and compared separately in their respective independent subspaces. This allows for sensitive detection of the early response to hepatocellular carcinoma conversion therapy, overcoming the technical defects of existing methods where the early weak therapeutic signal is diluted due to feature entanglement and the inability to distinguish the source and direction of response diffusion.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides a system for implementing the above-mentioned method for early evaluation of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the system for early evaluation of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging provided below can be found in the limitations of the method for early evaluation of the efficacy of hepatocellular carcinoma conversion therapy based on magnetic resonance imaging described above, and will not be repeated here.

[0161] In one exemplary embodiment, such as Figure 3 As shown, a magnetic resonance imaging-based early assessment system 30 for the efficacy of hepatocellular carcinoma conversion therapy is provided to implement the methods in the above-described embodiments. The system includes:

[0162] The multimodal image preprocessing module 31 is used to acquire liver magnetic resonance images of the patient at at least one time point before and after conversion therapy, preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region.

[0163] The feature decoupling encoding module 32 is used to input multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks into a pre-trained dual-branch decoupling encoder network to obtain statistically independent enhancement embedding feature maps and texture embedding feature maps. The dual-branch decoupling encoder network includes an enhancement encoding branch and a texture encoding branch. During training, the lower bound of the mutual information between enhancement embedding and texture embedding is minimized as a constraint to make the enhancement embedding feature map and texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features.

[0164] The sub-region segmentation and quantization module 33 is used to divide the tumor voxel into multiple enhanced sub-regions based on the enhanced embedding feature map and determine the volume of each enhanced sub-region; and to divide the tumor voxel into multiple texture sub-regions based on the texture embedding feature map and determine the volume of each texture sub-region.

[0165] The spatiotemporal graph construction module 34 is used to construct an enhanced subspace graph with multiple enhanced subregions as nodes and a texture subspace graph with multiple texture subregions as nodes. It establishes spatial adjacency edges within the first time point and cross-time transformation edges between the first time point and the second time point in the enhanced subspace graph and the texture subspace graph to form a time series graph.

[0166] The graph sequence evolution coding module 35 is used to input the time series graph into a pre-trained graph sequence encoder and output the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-time transformation relationship of the sub-region.

[0167] The efficacy assessment generation module 36 is used to calculate the enhancement conversion rate based on the enhancement conversion probability matrix and the volume of the enhancement subregion, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture subregion, and generate early efficacy assessment results based on the difference between the enhancement conversion rate and the texture conversion rate.

[0168] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0169] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0170] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0171] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for early evaluation of the efficacy of conversion therapy for hepatocellular carcinoma based on magnetic resonance imaging, characterized in that, The method includes: S1. Obtain liver magnetic resonance images of the patient at at least one time point before and after conversion therapy, and preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region. S2. Input the multi-phase dynamic enhancement image block and the multi-sequence structure image block into a pre-trained dual-branch decoupled encoder network to obtain statistically independent enhancement embedding feature maps and texture embedding feature maps; wherein, the dual-branch decoupled encoder network includes an enhancement coding branch and a texture coding branch, and during the training process, the lower bound of the mutual information between the enhancement embedding and the texture embedding is used as a constraint to make the enhancement embedding feature map and the texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features; S3. Based on the enhanced embedding feature map, the tumor voxel is divided into multiple enhanced subregions, and the volume of each enhanced subregion is determined; based on the texture embedding feature map, the tumor voxel is divided into multiple texture subregions, and the volume of each texture subregion is determined. S4. Construct an enhanced subspace graph with multiple enhanced subregions as nodes, construct a texture subspace graph with multiple texture subregions as nodes, establish spatial adjacency edges within the first time point and cross-time transformation edges between the first time point and the second time point in the enhanced subspace graph and the texture subspace graph, and form a time series graph. S5. Input the time series graph into a pre-trained graph sequence encoder, and output the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-time transformation relationship of the sub-region; S6. Calculate the enhancement conversion rate based on the enhancement conversion probability matrix and the volume of the enhancement sub-region, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture sub-region, and generate an early efficacy assessment result based on the difference between the enhancement conversion rate and the texture conversion rate.

2. The method according to claim 1, characterized in that, The training process of the dual-branch decoupled encoder network includes: S11. Input the multi-phase dynamic enhancement image block samples of the training samples into the enhancement coding branch, and input the multi-sequence structure image block samples of the training samples into the texture coding branch to obtain the training enhancement embedding feature map and the training texture embedding feature map respectively. S12. Extract training augmentation embedding vectors at multiple spatial locations from the training augmentation embedding feature map, and extract training texture embedding vectors at spatial locations corresponding to the training augmentation embedding vectors from the training texture embedding feature map. S13. Pair the training augmented embedding vectors with the training texture embedding vectors at the same spatial location to obtain positive sample pairs; pair the training augmented embedding vectors with the training texture embedding vectors at randomly shuffled spatial locations to obtain negative sample pairs; S14. Based on the positive sample pair and the negative sample pair, calculate the first mutual information lower bound estimation loss value using the pre-built discriminator network, and update the parameters of the discriminator network by maximizing the first mutual information lower bound estimation loss value; wherein, the input of the discriminator network is the concatenation result of the training augmented embedding vector and the training texture embedding vector, and the output is a mutual information estimation value used to estimate the mutual information lower bound between the training augmented embedding vector and the training texture embedding vector; S15. Based on the updated discriminator network, calculate the second mutual information lower bound estimation loss value for the positive sample pair and the negative sample pair, and update the parameters of the enhanced coding branch and the texture coding branch by minimizing the second mutual information lower bound estimation loss value; S16. Repeat S14 and S15 until the preset convergence condition is met to obtain the trained dual-branch decoupled encoder network; wherein, the first mutual information lower bound estimation loss value and the second mutual information lower bound estimation loss value are calculated according to the following formula: ； in, This represents the lower bound of mutual information to estimate the loss value. The number of spatial locations sampled in each batch. This refers to the discriminator network. Indicates the first positive sample pair Training augmented embedding vectors for each spatial location, Indicates the first positive sample pair Training texture embedding vectors at spatial locations, Indicates that the negative sample pair is related to the first... The training texture embedding vector corresponding to each spatial location. This represents an exponential function with the natural constant as its base. This represents the natural logarithm function.

3. The method according to claim 1, characterized in that, Based on the enhanced embedding feature map, the tumor voxel is divided into multiple enhanced subregions, and the volume of each enhanced subregion is determined. Based on the texture embedding feature map, the tumor voxel is divided into multiple texture sub-regions, and the volume of each texture sub-region is determined, including: S21. Input the multi-phase dynamic enhancement image block samples and multi-sequence structure image block samples of all historical patients in the training set before treatment into the dual-branch decoupled encoder network to obtain the training enhancement embedding feature map and the training texture embedding feature map. S22. Perform K-means clustering on the voxels in the training augmented embedding feature map to obtain multiple global augmented clustering centers; perform K-means clustering on the voxels in the training texture embedding feature map to obtain multiple global texture clustering centers; S23. For the current patient, extract the enhanced embedding vector of each voxel from the enhanced embedding feature map, calculate the first Euclidean distance between the enhanced embedding vector and each of the global enhanced cluster centers; assign the voxel to the global enhanced cluster center with the smallest first Euclidean distance to obtain the enhanced sub-region, and count the volume of each enhanced sub-region. S24. For the current patient, extract the texture embedding vector of each voxel from the texture embedding feature map, calculate the second Euclidean distance between the texture embedding vector and each of the global texture cluster centers; assign the voxel to the global texture cluster center with the smallest second Euclidean distance to obtain the texture sub-region, and count the volume of each texture sub-region.

4. The method according to claim 1, characterized in that, The step of establishing spatial adjacency edges within a first time point and cross-time transformation edges between the first and second time points in the enhanced subspace graph and the texture subspace graph to form a time series graph includes: S31. For each node in the enhanced subspace graph, calculate the enhanced embedding mean vector based on the enhanced embedding vector located in the corresponding enhanced subregion in the enhanced embedding feature graph; calculate the enhanced geometric feature vector based on the geometric properties of the enhanced subregion; concatenate the enhanced embedding mean vector with the enhanced geometric feature vector to obtain the enhanced node feature vector. S32. For each node in the texture subspace graph, calculate the texture embedding mean vector based on the texture embedding vector located in the corresponding texture subregion in the texture embedding feature graph; calculate the texture geometric feature vector based on the geometric properties of the texture subregion; concatenate the texture embedding mean vector with the texture geometric feature vector to obtain the texture node feature vector. S33. Based on the three-dimensional spatial position of each enhancement sub-region in the enhancement subspace graph, establish a first spatial adjacency edge between any two enhancement sub-regions that share at least one adjacent voxel; based on the three-dimensional spatial position of each texture sub-region in the texture subspace graph, establish a second spatial adjacency edge between any two texture sub-regions that share at least one adjacent voxel. S34. Based on the number and volume of adjacent voxel faces of the two enhanced subregions connected by each first spatial adjacent edge in the enhanced subspace graph, calculate the weight of the first spatial adjacent edge; based on the number and volume of adjacent voxel faces of the two texture subregions connected by each second spatial adjacent edge in the texture subspace graph, calculate the weight of the second spatial adjacent edge; wherein, the calculation formulas for the weights of the first spatial adjacent edge and the second spatial adjacent edge are as follows: ； ； in, Indicates a point in time Lower Enhanced Subregion With enhanced subregions The weight of the first spatial adjacent edge between them. Indicates enhanced subregion With enhanced subregions At the point of time The number of adjacent voxel faces, and Representing enhanced subregions and enhance sub-regions At the point of time Volume; Indicates a point in time Subtexture subregion With texture subregion The weights of the second spatial adjacent edges between them. Represents texture subregions With texture subregion At the point of time The number of adjacent voxel faces, and Representing texture subregions respectively and texture subregion At the point of time Volume; S35. Perform deformable registration on the multi-phase dynamic enhancement image block at the first time point and the multi-phase dynamic enhancement image block at the second time point to obtain the voxel correspondence between the tumor region at the first time point and the tumor region at the second time point. S36. Based on the voxel correspondence, count the number of first voxel transformations from the enhanced subregion at the first time point to the enhanced subregion at the second time point; use the number of first voxel transformations to form a first transformation counting matrix, perform row normalization on the first transformation counting matrix to obtain an enhanced experience transformation probability matrix; extract the enhanced experience transformation probability from the enhanced subregion at the first time point to the enhanced subregion at the second time point from the enhanced experience transformation probability matrix. S37. Based on the voxel correspondence, count the number of second voxel transformations from the texture sub-region at the first time point to the texture sub-region at the second time point; use the number of second voxel transformations to form a second transformation counting matrix, perform row normalization on the second transformation counting matrix to obtain a texture empirical transformation probability matrix; extract the texture empirical transformation probability from the texture sub-region at the first time point to the texture sub-region at the second time point from the texture empirical transformation probability matrix. S38. Establish a fully connected first time-transformation edge between the enhanced subregion at the first time point and the enhanced subregion at the second time point in the enhanced subspace graph, and use the corresponding enhanced experience transformation probability as the weight of the first time-transformation edge; establish a fully connected second time-transformation edge between the texture subregion at the first time point and the texture subregion at the second time point in the texture subspace graph, and use the corresponding texture experience transformation probability as the weight of the second time-transformation edge. S39. The enhanced subspace graph containing the first spatial adjacent edges and their weights, the first cross-time transformation edges and their weights, and the enhanced node feature vectors is used as the first time-series subspace graph corresponding to the enhanced subspace graph; the texture subspace graph containing the second spatial adjacent edges and their weights, the second cross-time transformation edges and their weights, and the texture node feature vectors is used as the second time-series subspace graph corresponding to the texture subspace graph; the time-series graph is composed of the first time-series subspace graph and the second time-series subspace graph.

5. The method according to claim 4, characterized in that, The step of inputting the time series graph into a pre-trained graph sequence encoder and outputting the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point, includes: S41. For each time point in the time series graph, the graph attention network is used to perform spatial dimension message passing on the enhanced node feature vector along the first spatial adjacent edge and based on the weight of the first spatial adjacent edge to obtain a spatially updated enhanced node feature vector; the graph attention network is used to perform spatial dimension message passing on the texture node feature vector along the second spatial adjacent edge and based on the weight of the second spatial adjacent edge to obtain a spatially updated texture node feature vector. S42. Using a graph attention network along the first cross-time transformation edge and based on the weight of the first cross-time transformation edge, the spatial update enhancement node feature vector at the first time point is passed to each enhancement sub-region node at the second time point to obtain the spatiotemporal fusion enhancement node feature vector; using a graph attention network along the second cross-time transformation edge and based on the weight of the second cross-time transformation edge, the spatial update texture node feature vector at the first time point is passed to each texture sub-region node at the second time point to obtain the spatiotemporal fusion texture node feature vector; S43. Input the spatiotemporal fusion enhancement node feature vector into the gated loop unit in chronological order to obtain the enhancement node hidden state at each time step; input the spatiotemporal fusion texture node feature vector into the gated loop unit in chronological order to obtain the texture node hidden state at each time step; S44. For each enhanced sub-region at the first time point, the hidden state of the enhanced node at the first time point is concatenated with the hidden state of the enhanced node in each enhanced sub-region at the second time point to obtain a first concatenation result; the first concatenation result is input into a multilayer perceptron to obtain an enhanced conversion tendency score; the enhanced conversion tendency score is normalized by softmax to obtain the enhanced conversion probability matrix. S45. For each texture sub-region at the first time point, the hidden state of the texture node at the first time point is concatenated with the hidden state of the texture node in each texture sub-region at the second time point to obtain a second concatenation result; the second concatenation result is input into a multilayer perceptron to obtain a texture transformation tendency score; the texture transformation tendency score is normalized by softmax to obtain the texture transformation probability matrix.

6. The method according to claim 1, characterized in that, The step of calculating the enhancement conversion rate based on the enhancement conversion probability matrix and the volume of the enhancement subregion, calculating the texture conversion rate based on the texture conversion probability matrix and the volume of the texture subregion, and generating an early efficacy assessment result based on the difference between the enhancement conversion rate and the texture conversion rate includes: S51. Calculate the enhancement conversion rate based on the diagonal elements of the enhancement conversion probability matrix and the volume of the enhancement subregion; wherein the enhancement conversion rate characterizes the proportion of enhancement subregions undergoing type change; the formula for calculating the enhancement conversion rate is: ； in, This indicates the enhanced conversion rate. To increase the number of subregions, Indicates the first The volume of each enhanced subregion The first in the enhanced transformation probability matrix Line number The diagonal elements of the column represent enhanced subregions. The probability of remaining unchanged; S52. Calculate the texture transformation rate based on the diagonal elements of the texture transformation probability matrix and the volume of the texture sub-region; wherein, the texture transformation rate characterizes the proportion of texture sub-regions undergoing type changes; the formula for calculating the texture transformation rate is: ； in, This indicates the texture conversion rate. The number of texture subregions. Indicates the first The volume of each texture subregion; The texture transformation probability matrix is ​​the first... Line number The diagonal elements of the column represent texture subregions. The probability of remaining unchanged; S53. Calculate the difference between the enhancement conversion rate and the texture conversion rate to obtain the enhancement-texture conversion difference index, and generate the early efficacy assessment result based on the enhancement-texture conversion difference index.

7. A magnetic resonance imaging-based early assessment system for the efficacy of conversion therapy for hepatocellular carcinoma, used to implement the method according to any one of claims 1 to 6, characterized in that, The system includes: The multimodal image preprocessing module is used to acquire liver magnetic resonance images of the patient at at least one time point before and after conversion therapy, and to preprocess the liver magnetic resonance images to obtain multi-phase dynamic enhancement image blocks and multi-sequence structure image blocks containing the tumor region. The feature decoupling encoding module is used to input the multi-phase dynamic enhancement image patch and the multi-sequence structure image patch into a pre-trained dual-branch decoupling encoder network to obtain statistically independent enhancement embedding feature maps and texture embedding feature maps. The dual-branch decoupling encoder network includes an enhancement encoding branch and a texture encoding branch. During training, minimizing the lower bound of the mutual information between the enhancement embedding and the texture embedding is used as a constraint to make the enhancement embedding feature map and the texture embedding feature map statistically independent, so as to separate the enhancement dynamic features and the organization texture features. The sub-region segmentation and quantization module is used to divide the tumor voxel into multiple enhanced sub-regions based on the enhanced embedding feature map and determine the volume of each enhanced sub-region; and to divide the tumor voxel into multiple texture sub-regions based on the texture embedding feature map and determine the volume of each texture sub-region. The spatiotemporal graph construction module is used to construct an enhanced subspace graph with multiple enhanced subregions as nodes, construct a texture subspace graph with multiple texture subregions as nodes, and establish spatial adjacency edges within a first time point and cross-time transformation edges between the first time point and the second time point in the enhanced subspace graph and the texture subspace graph to form a time series graph. The graph sequence evolution coding module is used to input the time series graph into a pre-trained graph sequence encoder and output the enhancement transformation probability matrix of the enhanced sub-region from the first time point to the second time point, and the texture transformation probability matrix of the texture sub-region from the first time point to the second time point; wherein, the graph sequence encoder is used to encode the spatial adjacency relationship and cross-temporal transformation relationship of the sub-region; The efficacy assessment generation module is used to calculate the enhancement conversion rate based on the enhancement conversion probability matrix and the volume of the enhancement sub-region, calculate the texture conversion rate based on the texture conversion probability matrix and the volume of the texture sub-region, and generate an early efficacy assessment result based on the difference between the enhancement conversion rate and the texture conversion rate.

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

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

Images