Intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images

Abstract

The application relates to a hepatocellular carcinoma postoperative recurrence intelligent prediction method based on multi-phase magnetic resonance images. The method comprises the following steps: acquiring multi-phase magnetic resonance images of an arterial phase, a portal phase, a delay phase and a hepatobiliary phase, and segmenting an arterial phase image to generate a tumor main body mask; generating a peritumoral micro-infiltration probability graph by adaptively morphologically dilating the tumor main body mask; generating a target displacement vector field and a multi-phase registration image by perceiving residual errors of the peritumoral micro-infiltration probability graph; generating a multi-phase registration residual confidence graph by fusing a local deformation anomaly of the target displacement vector field and a closed loop error; generating a composite feature vector by extracting a spatio-temporal decoupling feature of the multi-phase registration image based on the multi-phase registration residual confidence graph; generating a pure biological feature vector by orthogonally projecting and decoupling the composite feature vector; and generating a hepatocellular carcinoma postoperative recurrence risk score by classifying and mapping the pure biological feature vector. The method can inhibit registration residual interference, extract pure biological features and improve accuracy.

Description

Technical Field

[0001] This invention belongs to the field of postoperative prediction technology for hepatocellular carcinoma, and in particular relates to an intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging. Background Technology

[0002] With the development of medical imaging technology, multiphase magnetic resonance imaging (MRI) has become an important tool for assessing postoperative recurrence of hepatocellular carcinoma (HCC) due to its ability to reflect the hemodynamic distribution characteristics of contrast agents at different stages. Current methods for predicting postoperative recurrence of HCC mainly rely on extracting morphological and functional features of the tumor and its surrounding areas from multiphase MRI images for risk stratification.

[0003] In traditional techniques, when using multi-phase magnetic resonance imaging (MRI) images for postoperative recurrence prediction, the images of different phases are typically first spatially registered to eliminate anatomical positional shifts caused by respiratory motion. Then, the spatiotemporal features of the multi-phase images are directly extracted from the registered images. Finally, the extracted features are input into a model to generate a recurrence risk score. This approach implicitly assumes that the multi-phase images, after registration, have achieved perfect anatomical alignment at the voxel level.

[0004] However, traditional methods have significant drawbacks. The peritumoral microinvasive region of hepatocellular carcinoma is crucial in determining early postoperative recurrence, but this region exhibits extremely low signal contrast and blurred tissue boundaries on magnetic resonance imaging. Traditional registration methods are prone to producing unreliable voxel misalignment mappings in microinvasive regions lacking clear anchor points. When extracting features directly from the registered images, these grayscale differences caused by registration misalignment are mistakenly identified as genuine contrast agent kinetic changes, introducing numerous spurious artifacts. Furthermore, current techniques lack a mechanism to effectively separate the artifacts introduced by registration residuals from genuine biological features, resulting in severe interference from registration errors in the model and hindering accurate assessment of postoperative recurrence risk in hepatocellular carcinoma. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images, which can suppress interference from multi-phase registration residuals and extract pure biological features, in order to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides an intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging, including:

[0007] S1. Acquire multi-phase magnetic resonance images of the target object, and segment the tumor region in the arterial phase image to generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images.

[0008] S2. Adaptive morphological expansion of the tumor mask is performed to generate a probability map of peritumoral micro-infiltration.

[0009] S3. Based on the peritumoral microinfiltration probability map, residual sensing adaptive registration is performed on arterial phase images with portal venous phase images, delayed phase images and hepatobiliary phase images to generate target displacement vector field and multi-phase registration images.

[0010] S4. Perform local deformation anomaly and closed-loop error fusion quantization on the target displacement vector field to generate a multi-phase registration residual confidence map;

[0011] S5. Based on the confidence map of multi-phase registration residuals, perform residual perception spatiotemporal decoupling feature extraction on multi-phase registration images to generate composite feature vectors.

[0012] S6. Perform residual orthogonal projection decoupling on the composite feature vector to generate a pure biological feature vector;

[0013] S7. Classification mapping is performed based on pure biological feature vectors to generate a risk score for recurrence after hepatocellular carcinoma surgery.

[0014] In one embodiment, S3 includes:

[0015] S31. Calculate the local smoothing penalty weight based on the peritumoral microinfiltration probability map and generate an adaptive penalty weight map;

[0016] S32. Construct an adaptive smoothing regularization loss function based on the adaptive penalty weight graph. The formula for calculating the adaptive smoothing regularization loss function is as follows:

[0017] ;

[0018] in, To adaptively smooth the regularization loss value, Voxel positions in the adaptive penalty weight graph The weight value at that location, For the target displacement vector field at the voxel position Spatial gradient at that location To calculate the square norm;

[0019] S33. The spatial transformation network is jointly driven by the adaptive smoothing regularization loss function and the image similarity loss function, and nonlinear deformation is applied to the portal venous phase image, delayed phase image and hepatobiliary phase image to generate the initial displacement vector field.

[0020] S34. Based on the initial displacement vector field, perform spatial distortion transformation on the portal venous phase image, delayed phase image and hepatobiliary phase image to generate multi-phase registration image, and determine the initial displacement vector field as the target displacement vector field.

[0021] In one embodiment, S31 includes:

[0022] S311. Extract the micro-infiltration probability value of each voxel position in the peritumoral micro-infiltration probability map and generate a voxel micro-infiltration probability set.

[0023] S312. Based on the set of voxel micro-infiltration probabilities and the preset global basic weights, calculate the adaptive smoothing penalty weight for each voxel position and generate a set of voxel penalty weights.

[0024] S313. Generate an adaptive penalty weight map by combining the adaptive smoothing penalty weights of all voxel positions in the voxel penalty weight set.

[0025] S314. Based on the adaptive penalty weight graph, spatially weight the standard smoothing regularization term to generate the adaptive smoothing regularization loss function.

[0026] In one embodiment, S4 includes:

[0027] S41. Based on the target displacement vector field, calculate the absolute value deviation of the determinant of the spatial Jacobian matrix and generate a local deformation anomaly map.

[0028] S42. Based on the target displacement vector field and the corresponding reverse displacement vector field, calculate the identity transformation deviation norm and generate the closed-loop error norm spectrum.

[0029] S43. Multiply and fuse the local deformation anomaly map and the closed-loop error norm map to generate the original residual intensity map;

[0030] S44. Perform Gaussian low-pass filtering and normalization mapping on the original residual intensity map to generate a multi-phase registration residual confidence map.

[0031] In one embodiment, S42 includes:

[0032] S421. Extract the forward displacement field based on the target displacement vector field, and extract the reverse displacement vector field based on the corresponding reverse registration network to generate a two-way displacement field pair.

[0033] S422. Based on the bidirectional displacement field pair, calculate the composite displacement deviation vector and generate the closed-loop displacement deviation vector.

[0034] S423. Calculate the Euclidean norm of the closed-loop displacement deviation vector to generate the closed-loop error norm. The formula for calculating the closed-loop error norm is:

[0035] ;

[0036] in, voxel position The closed-loop error norm at that point, To calculate the L2 norm of a vector, Let x be the forward displacement of the target displacement vector field at the voxel position x. The reverse displacement vector field at the voxel position The reverse displacement, For spatial compounding operations, For the spatial coordinates of the voxel;

[0037] S424. Aggregate the closed-loop error norms of all voxel locations to generate a closed-loop error norm map.

[0038] In one embodiment, S5 includes:

[0039] S51. Perform phase-by-phase three-dimensional convolutional encoding on the multi-phase registration images to generate phase-by-phase convolutional feature maps;

[0040] S52. Based on the multi-phase registration residual confidence map, spatial attention masking is applied to the phase-by-phase convolutional feature map to generate a residual suppression feature map.

[0041] S53. Based on the multi-phase registration residual confidence map, the residual suppression feature map is penalized with temporal attention weights to generate a spatiotemporal aggregated feature map.

[0042] S54. Perform global average pooling on the spatiotemporal aggregated feature map to generate a composite feature vector.

[0043] In one embodiment, S6 includes:

[0044] S61. Based on the preset residual basis vector matrix, perform column space projection on the composite feature vector to generate residual feature components;

[0045] S62. Subtract the residual feature components from the composite feature vector to generate an orthogonalized intermediate feature vector;

[0046] S63. Minimize the inner product norm constraint on the orthogonalized intermediate eigenvectors and residual basis vector matrices to generate the orthogonality constraint loss function. The formula for calculating the orthogonality constraint loss function is as follows:

[0047] ;

[0048] in, The orthogonal constraint loss value. To calculate the matrix norm, This is the transpose of the orthogonalized intermediate eigenvectors. The residual basis vector matrix;

[0049] S64. Based on the orthogonality constraint loss function, the orthogonalized intermediate feature vector is iteratively updated to generate a pure biological feature vector.

[0050] Secondly, this application also provides an intelligent prediction device for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging, comprising:

[0051] The tumor segmentation module is used to acquire multi-phase magnetic resonance images of the target object, segment the tumor region in the arterial phase image, and generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images.

[0052] The probability map generation module is used to adaptively expand the tumor body mask to generate a probability map of peritumoral microinfiltration.

[0053] The residual registration module is used to perform residual-aware adaptive registration of arterial phase images with portal venous phase images, delayed phase images and hepatobiliary phase images based on the peritumoral microinfiltration probability map, and generate target displacement vector field and multi-phase registered images.

[0054] The residual quantization module is used to perform local deformation anomaly and closed-loop error fusion quantization on the target displacement vector field and generate multi-phase registration residual confidence map;

[0055] The spatiotemporal decoupling module is used to extract residual-aware spatiotemporal decoupling features from multi-phase registration images based on multi-phase registration residual confidence maps, and generate composite feature vectors.

[0056] The orthogonal decoupling module is used to perform residual orthogonal projection decoupling on composite feature vectors to generate pure biological feature vectors.

[0057] The risk scoring module is used to perform classification mapping based on pure biological feature vectors to generate a risk score for postoperative recurrence of hepatocellular carcinoma.

[0058] 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 the intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images as described in the first aspect.

[0059] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images as described in the first aspect.

[0060] The aforementioned intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging locates the tumor and adaptively expands to predict the peritumoral microinvasive region. Based on this, residual-aware adaptive registration is performed to avoid artifacts caused by forced alignment of the microinvasive region due to the lack of anchor points. Then, the local deformation anomalies and closed-loop errors of the registration displacement field are quantified to generate a residual confidence map, which accurately locates unreliable registration regions. Subsequently, during feature extraction, this residual confidence map is used for spatiotemporal decoupling to suppress residual region interference. Furthermore, orthogonal projection decoupling is used to completely remove the registration residual components from the composite features, thereby effectively avoiding the problem of false artifact features caused by registration misalignment being misjudged as real hemodynamic signals. This achieves deep decoupling between registration error and real biological information, significantly improving the accuracy and reliability of postoperative recurrence risk prediction for hepatocellular carcinoma. Attached Figure Description

[0061] 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.

[0062] Figure 1 A flowchart illustrating an intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging provided by the present invention.

[0063] Figure 2 This is a schematic diagram of the structure of an intelligent prediction device for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging provided by the present invention. Detailed Implementation

[0064] 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.

[0065] In one embodiment, such as Figure 1 As shown, a method for intelligent prediction of postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging is provided. This embodiment illustrates the application of this method to an intelligent prediction terminal. It is understood that this method can also be applied to an intelligent prediction server, and can also be applied to a system including an intelligent prediction terminal and an intelligent prediction server, and can be implemented through the interaction between the intelligent prediction terminal and the intelligent prediction server. In this embodiment, the method includes the following steps:

[0066] S1. Acquire multi-phase magnetic resonance images of the target object, and segment the tumor region in the arterial phase image to generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images.

[0067] Optionally, the intelligent prediction terminal can read the raw data of multi-phase magnetic resonance images of the target object from the medical image storage device through medical digital imaging and communication protocols. The multi-phase magnetic resonance images can include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images. The above four phase images reflect the tissue hemodynamic characteristics of the contrast agent in the arterial blood supply phase, portal venous blood supply phase, equilibrium distribution phase, and hepatocyte-specific uptake phase, respectively.

[0068] Specifically, the intelligent prediction terminal can preprocess the multi-phase magnetic resonance images it reads. Preprocessing operations may include removing background noise, resampling images with different scan slice thicknesses to a uniform voxel size, and normalizing the image grayscale values. The intelligent prediction terminal can use a U-shaped network to segment tumor regions in the preprocessed arterial phase images.

[0069] Specifically, the training process of the U-shaped network is as follows: Arterial phase images labeled with tumor region boundaries are collected as the training set. Preprocessed arterial phase images are used as network input, and binary masks of tumor regions manually labeled by radiologists are used as training labels. A weighted sum of Dessian similarity coefficient loss and binary cross-entropy loss is used as the loss function. An adaptive moment estimation optimizer iteratively updates the network parameters until the loss function converges to a stable range. After training, the U-shaped network outputs the probability value of each voxel in the arterial phase image belonging to the tumor region. The intelligent prediction terminal can binarize the probability values ​​to generate a tumor main mask. In the tumor main mask, the voxel value of the tumor region is 1, and the voxel value of the non-tumor region is 0.

[0070] S2. Adaptive morphological expansion of the tumor body mask is performed to generate a peritumoral micro-infiltration probability map.

[0071] Optionally, the intelligent prediction terminal can first extract the edge contour of the tumor body mask and calculate the local curvature value at each position on the edge of the tumor body mask. The intelligent prediction terminal can adaptively adjust the size of the structuring element for the morphological dilation operation based on the local curvature value. For convex or concave areas of the tumor edge with large local curvature, smaller structuring elements are used for dilation to avoid over-dilation that includes too much normal liver tissue; for smooth areas of the tumor edge with small local curvature, larger structuring elements are used for dilation to ensure coverage of any potential peritumoral micro-invasive range. The intelligent prediction terminal can use the adaptively adjusted structuring elements to perform multi-scale morphological dilation of the tumor body mask, obtaining a set of binary masks with different degrees of dilation. The intelligent prediction terminal can calculate the shortest Euclidean distance from each voxel position to the edge of the tumor body mask, mapping the shortest Euclidean distance to a micro-invasive probability value between 0 and 1. The closer a voxel position is to the edge of the tumor body, the higher its micro-invasive probability value; the farther a voxel position is from the edge of the tumor body, the lower its micro-invasive probability value. The intelligent prediction terminal can aggregate the microinvasive probability values ​​of all voxel locations to generate a peritumoral microinvasive probability map with the same size as the arterial phase image.

[0072] S3. Based on the peritumoral microinfiltration probability map, residual sensing adaptive registration is performed on arterial phase images with portal venous phase images, delayed phase images and hepatobiliary phase images to generate target displacement vector field and multi-phase registered images.

[0073] Optionally, the intelligent prediction terminal can use the arterial phase image as a fixed image and the portal venous phase image, delayed phase image, and hepatobiliary phase image as floating images, respectively, to perform residual-aware adaptive registration. Residual-aware adaptive registration introduces a peritumoral microinvasive probability map as a prior constraint into the traditional image registration framework. Addressing the characteristics of low signal contrast and blurred tissue boundaries in the peritumoral microinvasive region, it adaptively adjusts the smoothing regularization intensity during the registration process, allowing for greater local deformation in the peritumoral microinvasive region to achieve more accurate anatomical alignment.

[0074] Optionally, the intelligent prediction terminal can employ a deep learning-based spatial transformation network to implement the registration process. The input to the spatial transformation network is a fixed image, a floating image, and a peritumoral micro-infiltration probability map. The output is a target displacement vector field describing the spatial displacement of each voxel from the floating image to the fixed image. The intelligent prediction terminal can perform a voxel-by-voxel spatial distortion transformation on the floating image based on the target displacement vector field, mapping each voxel of the floating image to the corresponding coordinate position in the fixed image, generating a multi-phase registration image spatially aligned with the arterial phase image.

[0075] S4. Perform local deformation anomaly and closed-loop error fusion quantization on the target displacement vector field to generate a multi-phase registration residual confidence map.

[0076] Optionally, the intelligent prediction terminal can assess the registration reliability of each voxel position by quantifying the local consistency and bidirectional registration consistency of the target displacement vector field, and generate a multi-phase registration residual confidence map. Local deformation anomaly quantification identifies abnormal deformation regions that do not conform to the elastic deformation law of normal liver tissue by analyzing the spatial gradient change characteristics of the target displacement vector field; these abnormal deformation regions typically correspond to locations with large registration errors. Closed-loop error quantification assesses the self-consistency of the registration result by calculating the composite displacement deviation between forward and reverse registration. If the registration result is accurate, the composite of forward and reverse displacements should approximately equal the identity transformation. The intelligent prediction terminal can fuse the local deformation anomaly quantification results and the closed-loop error quantification results through voxel-by-voxel product to obtain the original residual intensity map. A higher value in the original residual intensity map indicates a larger registration error at the corresponding voxel position. The intelligent prediction terminal can perform Gaussian low-pass filtering on the original residual intensity map to remove noise interference, and then normalize the filtered residual intensity value to map it to the interval between 0 and 1 to generate a multi-phase registration residual confidence map. The higher the value of the multi-phase registration residual confidence map, the more reliable the registration result of the corresponding voxel position.

[0077] S5. Based on the residual confidence map of multi-phase registration, residual perception spatiotemporal decoupling feature extraction is performed on the multi-phase registration image to generate composite feature vector.

[0078] Optionally, the intelligent prediction terminal can perform phase-by-phase 3D convolutional encoding on multi-phase registered images to extract local morphological and texture features of each phase image, generating phase-by-phase convolutional feature maps. The intelligent prediction terminal can introduce multi-phase registration residual confidence maps as guiding information for feature extraction, weighting features in both spatial and temporal dimensions to achieve spatiotemporal decoupling feature extraction with residual awareness. In the spatial dimension, the intelligent prediction terminal can use the multi-phase registration residual confidence maps as spatial attention weights to perform voxel-by-voxel weighting on the phase-by-voxel convolutional feature maps, reducing the feature contribution of voxel positions with large registration errors and suppressing the propagation of artifact features. In the temporal dimension, the intelligent prediction terminal can calculate temporal attention weights based on the overall registration confidence of each phase image, globally weighting the feature maps of different phases to balance the differences in registration reliability between different phase images. The intelligent prediction terminal can then concatenate the multi-phase feature maps after spatial and temporal weighting along the channel dimension, and then perform global average pooling on the concatenated feature maps, converting the 3D feature maps into a 1D composite feature vector.

[0079] S6. Perform residual orthogonal projection decoupling on the composite feature vector to generate a pure biological feature vector.

[0080] Optionally, the intelligent prediction terminal can separate the true biological features from the artifact features introduced by registration errors in the composite feature vector based on the orthogonal projection principle in linear algebra. The intelligent prediction terminal can pre-learn a residual basis vector matrix representing the main distribution direction of registration artifact features using a large number of multi-phase magnetic resonance image samples containing varying degrees of registration errors. Each column of the residual basis vector matrix corresponds to an independent artifact feature basis vector. The intelligent prediction terminal can project the composite feature vector onto the artifact feature subspace spanned by the residual basis vector matrix to obtain the residual feature components contained in the composite feature vector. The intelligent prediction terminal can subtract the residual feature components from the composite feature vector to obtain a preliminary orthogonalized intermediate feature vector. To further improve the thoroughness of feature separation, the intelligent prediction terminal can introduce orthogonality constraints and iteratively optimize the orthogonalized intermediate feature vector to minimize the inner product norm of the orthogonalized intermediate feature vector and the residual basis vector matrix, ultimately obtaining a pure biological feature vector containing only true biological information.

[0081] S7. Classification mapping is performed based on pure biological feature vectors to generate a risk score for recurrence after hepatocellular carcinoma surgery.

[0082] Optionally, the intelligent prediction terminal can use a logistic regression model to achieve a classification mapping from pure biological feature vectors to postoperative recurrence risk scores for hepatocellular carcinoma.

[0083] Specifically, the training process of the logistic regression model is as follows: Cases of hepatocellular carcinoma patients with confirmed postoperative recurrence through clinical follow-up are collected as the training set. The multi-phase magnetic resonance imaging (MRI) images of each case are processed using the aforementioned steps to obtain the corresponding purified biological feature vector. This purified biological feature vector is used as the model input, and the binary follow-up label indicating whether the patient has experienced postoperative recurrence is used as the model output. The intelligent prediction terminal can use the maximum likelihood estimation method to solve for the parameters of the logistic regression model and optimize the model's classification accuracy. After training, the logistic regression model can output a probability value between 0 and 1. The intelligent prediction terminal can determine this probability value as a risk score for postoperative recurrence of hepatocellular carcinoma; a higher score indicates a higher risk of postoperative recurrence in the target individual.

[0084] The aforementioned intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging (MRI) images acquires multi-phase MRI images of the target object containing four key phases and generates a precise tumor mask. Adaptive morphological dilation is then used to obtain a probability map reflecting the likelihood of peritumoral microinvasiveness. Based on this probability map, residual-aware adaptive image registration is achieved. Quantitative registration residuals generate confidence maps to guide subsequent feature extraction. Finally, through two steps of residual-aware spatiotemporal decoupling feature extraction and residual orthogonal projection decoupling, true biological features and registration artifact features are thoroughly separated. A recurrence risk score is generated based on the pure features. This method effectively solves the problems of large registration errors and severe artifact interference in the peritumoral microinvasive region in traditional methods, significantly improving the accuracy and robustness of postoperative recurrence risk prediction for hepatocellular carcinoma and providing reliable technical support for the development of personalized postoperative follow-up and intervention plans in clinical practice.

[0085] In one embodiment, S3 may include:

[0086] S31. Based on the peritumoral microinfiltration probability map, calculate the local smoothing penalty weight and generate an adaptive penalty weight map.

[0087] Optionally, the intelligent prediction terminal can first extract the micro-infiltration probability value corresponding to each voxel position in the peritumoral micro-infiltration probability map. The micro-infiltration probability value reflects the likelihood of peritumoral micro-infiltration at the corresponding voxel position. The intelligent prediction terminal can load preset global base weights, which are pre-defined base smoothing penalty intensities applicable to normal liver tissue regions, used to constrain the smoothness of the displacement field in normal liver tissue regions. The intelligent prediction terminal can calculate the adaptive smoothing penalty weight for each voxel position based on the product relationship between the micro-infiltration probability value and the global base weight. The higher the micro-infiltration probability value, the lower the adaptive smoothing penalty weight corresponding to the voxel position, and vice versa. The intelligent prediction terminal can arrange the adaptive smoothing penalty weights of all voxel positions according to the spatial coordinates of the original image to generate an adaptive penalty weight map with the exact same size as the peritumoral micro-infiltration probability map.

[0088] S32. Construct an adaptive smoothing regularization loss function based on the adaptive penalty weight graph. The formula for calculating the adaptive smoothing regularization loss function is as follows:

[0089] ;

[0090] in, To adaptively smooth the regularization loss value, Voxel positions in the adaptive penalty weight graph The weight value at that location, For the target displacement vector field at the voxel position Spatial gradient at that location To calculate the square norm.

[0091] Optionally, the intelligent prediction terminal can construct a spatially variable adaptive smoothing regularization loss function based on an adaptive penalty weight map. This function can constrain the smoothness of the displacement field during registration, preventing unreasonable and drastic local deformations. The given formula for the adaptive smoothing regularization loss function differentiates the penalty for the smoothness of the displacement field in different regions by performing a weighted integral of the square of the displacement field gradient across the entire image domain. Normal liver tissue regions have higher weights and are subject to stronger smoothing constraints, while the peritumoral micro-infiltration regions have lower weights, allowing for greater local deformation freedom.

[0092] S33. The spatial transformation network is jointly driven by the adaptive smoothing regularization loss function and the image similarity loss function, and nonlinear deformation is applied to the portal venous phase image, the delayed phase image and the hepatobiliary phase image to generate the initial displacement vector field.

[0093] Optionally, the intelligent prediction terminal can weighted sum the adaptive smoothing regularization loss function and the image similarity loss function to obtain the total loss function used to drive the spatial transformation network. The image similarity loss function can employ mutual information loss, which effectively measures the statistical correlation between multi-phase magnetic resonance images and is suitable for multi-phase image registration scenarios where gray values ​​vary with contrast agent concentration. The intelligent prediction terminal can minimize the total loss function as the optimization objective and iteratively update the parameters of the spatial transformation network through a backpropagation algorithm, ensuring that the displacement field output by the spatial transformation network simultaneously satisfies the gray-level similarity requirements of floating and fixed images, as well as the adaptive smoothness requirements of the displacement field in different regions. When the total loss function value converges to a preset stable range, the spatial transformation network outputs the initial displacement vector field.

[0094] S34. Based on the initial displacement vector field, perform spatial distortion transformation on the portal venous phase image, delayed phase image and hepatobiliary phase image to generate multi-phase registration image, and determine the initial displacement vector field as the target displacement vector field.

[0095] Optionally, the intelligent prediction terminal can perform voxel-by-voxel spatial distortion transformations on the portal venous phase image, delayed phase image, and hepatobiliary phase image based on the initial displacement vector field. For each voxel position in the fixed image (i.e., the arterial phase image), the intelligent prediction terminal can obtain the corresponding coordinate position in the floating image based on the initial displacement vector field. Since this coordinate position is usually a non-integer coordinate, the intelligent prediction terminal can use trilinear interpolation to calculate the gray value of this non-integer coordinate position and assign the calculated gray value to the registration image of the corresponding voxel position in the fixed image. The intelligent prediction terminal can combine the separately generated portal venous phase registration image, delayed phase registration image, and hepatobiliary phase registration image into a multi-phase registration image, while determining the initial displacement vector field obtained through joint loss optimization as the target displacement vector field.

[0096] In the above embodiments, an adaptive penalty weight map is calculated based on the peritumoral microinfiltration probability map, a spatially variable adaptive smoothing regularization loss function is constructed, and an initial displacement vector field is generated by jointly driving a spatial transformation network in conjunction with an image similarity loss function. Then, a spatial distortion transformation is performed to obtain a multi-phase registration image and determine the target displacement vector field. This embodiment can dynamically adjust the smoothing constraint strength during the registration process, addressing the characteristics of low signal contrast and blurred boundaries in the peritumoral microinfiltration region. While ensuring the smoothness of registration in the normal liver tissue region, it allows for greater local deformation freedom in the peritumoral microinfiltration region, effectively solving the problem of voxel misalignment easily generated in the peritumoral microinfiltration region by traditional global smoothing registration, and significantly improving the registration accuracy of multi-phase magnetic resonance images in key recurrence-related regions.

[0097] In one embodiment, S31 may include:

[0098] S311. Extract the micro-infiltration probability value of each voxel position in the peritumoral micro-infiltration probability map to generate a voxel micro-infiltration probability set.

[0099] Optionally, the intelligent prediction terminal can traverse all voxel positions of the peritumoral micro-infiltration probability map row by row, column by column, and layer by layer according to the three-dimensional spatial coordinate order, and sequentially read the gray value corresponding to each voxel position. This gray value is the micro-infiltration probability value of the corresponding voxel position, and the value of the micro-infiltration probability value ranges from 0 to 1. The intelligent prediction terminal can bind and store each read micro-infiltration probability value with the corresponding three-dimensional spatial index to generate a voxel micro-infiltration probability set. The voxel micro-infiltration probability set can contain the micro-infiltration probability information of all voxel positions in the peritumoral micro-infiltration probability map, and can completely reflect the spatial distribution characteristics of the possibility of micro-infiltration in the peritumoral region.

[0100] S312. Based on the set of voxel micro-infiltration probabilities and the preset global basic weights, calculate the adaptive smoothing penalty weight for each voxel position and generate a set of voxel penalty weights.

[0101] Optionally, the intelligent prediction terminal can load preset global base weights. These preset global base weights are the baseline smoothing penalty intensity obtained statistically from a large number of registration experiments using multi-phase magnetic resonance images of normal liver tissue, used to provide a unified smoothing constraint benchmark for normal liver tissue regions. The intelligent prediction terminal can iterate through each element in the voxel micro-infiltration probability set, and calculate the adaptive smoothing penalty weight for the corresponding voxel position based on the micro-infiltration probability value corresponding to each element and the preset global base weights through a linear mapping relationship. The specific calculation formula is as follows:

[0102] ;

[0103] in, voxel position Adaptive smoothing penalty weights at the location, The preset global base weights, voxel position The micro-infiltration probability value at the location. The intelligent prediction terminal can bind and store all the calculated adaptive smoothing penalty weights with the corresponding three-dimensional spatial indexes to generate a voxel penalty weight set.

[0104] S313. Combine the adaptive smoothing penalty weights for all voxel positions in the voxel penalty weight set to generate an adaptive penalty weight map.

[0105] Optionally, the intelligent prediction terminal can create a three-dimensional empty array that perfectly matches the size, voxel spacing, and spatial coordinate system of the peritumoral microinfiltration probability map. This three-dimensional empty array serves as the initial state of the adaptive penalty weight map. The intelligent prediction terminal can iterate through each element in the voxel penalty weight set and, based on the three-dimensional spatial index of each element, assign the adaptive smoothing penalty weight value contained in that element to the corresponding coordinate position in the three-dimensional empty array. For voxel positions in the background region of the peritumoral microinfiltration probability map, the intelligent prediction terminal can assign preset global base weights to the corresponding coordinate positions, ensuring the integrity of the adaptive penalty weight map. After the assignment is complete, the intelligent prediction terminal can define this three-dimensional array as the adaptive penalty weight map, which can intuitively reflect the distribution of smoothing penalty intensity at different spatial locations.

[0106] S314. Based on the adaptive penalty weight graph, spatially weight the standard smoothing regularization term to generate the adaptive smoothing regularization loss function.

[0107] Optionally, the intelligent prediction terminal can load a standard smoothing regularization term. This term is a globally smoothing constraint widely used in traditional image registration methods. It constrains the overall smoothness of the displacement field by calculating the integral of the square of the spatial gradient over the entire image domain, preventing unreasonable and drastic deformation. The intelligent prediction terminal can multiply the integral kernel of the standard smoothing regularization term on a voxel-by-voxel basis with the adaptive penalty weight map, ensuring that the smoothing constraint strength at different spatial locations is proportional to the weight value at the corresponding location in the adaptive penalty weight map. The intelligent prediction terminal can then integrate the multiplied integral kernel over the entire image domain to generate an adaptive smoothing regularization loss function. This function enables differentiated smoothing constraints for different tissue regions.

[0108] In the above embodiments, a set of voxel micro-infiltration probabilities is generated by extracting voxel micro-infiltration probability values ​​from the peritumoral micro-infiltration probability map. An adaptive smoothing penalty weight set is generated by combining this with preset global base weights to calculate the adaptive smoothing penalty weight set for each voxel. These weights are then combined to generate a spatially aligned adaptive penalty weight map. Finally, the standard smoothing regularization term is spatially weighted to obtain the adaptive smoothing regularization loss function. This embodiment directly transforms the spatial distribution information of peritumoral micro-infiltration into the local smoothing constraint strength during the registration process, achieving a differentiated registration strategy for different tissue regions. This ensures both the smoothness of the displacement field and registration stability in normal liver tissue regions, while providing sufficient deformation degrees of freedom for peritumoral micro-infiltration regions with low signal contrast and blurred boundaries. It effectively avoids the voxel misalignment problem caused by traditional global smoothing registration in key recurrence-related regions, significantly improving the accuracy and reliability of multi-phase magnetic resonance image registration.

[0109] In one embodiment, S4 may include:

[0110] S41. Based on the target displacement vector field, calculate the absolute value deviation of the determinant of the spatial Jacobian matrix and generate a local deformation anomaly map.

[0111] Optionally, the intelligent prediction terminal can traverse all voxel positions of the target displacement vector field and calculate the spatial Jacobian matrix of the target displacement vector field for each voxel position. The target displacement vector field in three-dimensional space can contain displacement components in three directions. The Jacobian matrix is ​​composed of the partial derivatives of the three displacement components with respect to the three spatial coordinate axes, forming a third-order square matrix. The determinant value of the Jacobian matrix represents the volume change rate of the local region at that voxel position. The elastic deformation of normal liver tissue has the characteristic of continuous and smooth deformation, and the corresponding determinant value of the Jacobian matrix should be close to 1. The intelligent prediction terminal can calculate the difference between the absolute value of the determinant of the Jacobian matrix at each voxel position and 1 to obtain the local deformation anomaly degree at that voxel position. The intelligent prediction terminal can aggregate the local deformation anomaly degrees of all voxel positions to generate a local deformation anomaly degree map. The higher the value of the local deformation anomaly degree map, the less the local deformation at the corresponding voxel position conforms to the elastic deformation law of normal liver tissue, and the larger the registration error.

[0112] S42. Based on the target displacement vector field and the corresponding reverse displacement vector field, calculate the identity transformation deviation norm and generate a closed-loop error norm spectrum.

[0113] Optionally, the intelligent prediction terminal can use the arterial phase image as a floating image and the portal venous phase, delayed phase, and hepatobiliary phase images as fixed images, respectively. It then performs inverse registration using the same spatial transformation network and joint loss function as the forward registration process, obtaining the corresponding inverse displacement vector field. The intelligent prediction terminal can calculate the composite transformation result of the forward and inverse displacement vector fields. Ideally, an accurate registration result should satisfy the condition that the forward and inverse transformations are inverses of each other, and the composite transformation result is approximately equal to the identity transformation. The intelligent prediction terminal can calculate the deviation vector between the composite transformation result and the identity transformation, and then calculate the L2 norm of the deviation vector to obtain the closed-loop error norm for each voxel position. The intelligent prediction terminal can aggregate the closed-loop error norms for all voxel positions to generate a closed-loop error norm map. A higher value in the closed-loop error norm map indicates poorer registration self-consistency and a larger registration error for the corresponding voxel position.

[0114] S43. Multiply and fuse the local deformation anomaly map and the closed-loop error norm map to generate the original residual intensity map.

[0115] Optionally, the intelligent prediction terminal can perform voxel-by-voxel product fusion of the local deformation anomaly map and the closed-loop error norm map, multiplying the values ​​at corresponding voxel positions in the two maps to obtain the original residual intensity value at that voxel position. The local deformation anomaly assesses registration error from the perspective of the local geometric rationality of the displacement field, while the closed-loop error norm assesses registration error from the perspective of the bidirectional self-consistency of the registration result. These two evaluation indicators have independent and complementary characteristics. Product fusion can simultaneously amplify areas where both indicators show high errors, suppressing occasional high-value noise from a single indicator, and improving the accuracy and robustness of registration error detection. The intelligent prediction terminal can aggregate the original residual intensity values ​​at all voxel positions to generate an original residual intensity map; a higher value in the original residual intensity map indicates a larger overall registration error at the corresponding voxel position.

[0116] S44. Perform Gaussian low-pass filtering and normalization mapping on the original residual intensity map to generate a multi-phase registration residual confidence map.

[0117] Optionally, the intelligent prediction terminal can first perform Gaussian low-pass filtering on the original residual intensity map. Gaussian low-pass filtering uses a preset Gaussian convolution kernel to perform a three-dimensional convolution operation with the original residual intensity map, smoothing high-frequency noise components in the original residual intensity map while preserving the locally continuous registration residual distribution characteristics. The intelligent prediction terminal can then perform linear normalization mapping on the filtered residual intensity map, linearly mapping the filtered residual intensity values ​​to the interval between 0 and 1. Finally, the intelligent prediction terminal can invert the normalized residual intensity values ​​to obtain a multi-phase registration residual confidence map. Higher values ​​in the multi-phase registration residual confidence map indicate more reliable registration results at the corresponding voxel positions, while lower values ​​indicate larger registration errors at the corresponding voxel positions.

[0118] In the above embodiments, a local deformation anomaly map is generated by calculating the absolute value deviation of the determinant of the Jacobian matrix of the target displacement vector field, and a closed-loop error norm map is generated by the deviation norm of the identity transformation of the two-way displacement field. The original residual intensity map, comprehensively reflecting the registration error, is obtained through voxel-by-voxel product fusion. Then, a multi-phase registration residual confidence map is generated through Gaussian low-pass filtering for noise reduction and normalization mapping. This embodiment, through a two-dimensional fusion registration residual quantization method, can comprehensively and accurately characterize the spatial distribution of local errors generated during the registration process of multi-phase magnetic resonance images, effectively distinguishing between real tissue physiological deformation and abnormal deformation caused by registration errors. This provides reliable confidence guidance for the feature extraction process of residual perception, significantly reduces the interference of registration artifacts on the recurrence prediction model, and improves the accuracy and robustness of postoperative recurrence risk assessment for hepatocellular carcinoma.

[0119] In one embodiment, S42 may include:

[0120] S421. Extract the forward displacement field based on the target displacement vector field, and extract the reverse displacement vector field based on the corresponding reverse registration network to generate a two-way displacement field pair.

[0121] Optionally, the intelligent prediction terminal can directly extract the forward displacement field from the target displacement vector field. The forward displacement field describes the voxel-by-voxel spatial mapping relationship from the portal venous phase image, delayed phase image, and hepatobiliary phase image to the arterial phase image. The intelligent prediction terminal can use a spatial transformation network with the exact same structure as the forward registration process for reverse registration. During reverse registration, the intelligent prediction terminal can treat the floating image in the original forward registration as a fixed image and the fixed image in the original forward registration as a floating image, inputting them into the spatial transformation network for inference to obtain the corresponding reverse displacement vector field. The intelligent prediction terminal can bind the forward displacement field and reverse displacement vector field corresponding to each phase one by one to generate a bidirectional displacement field pair. The bidirectional displacement field pair can contain the registration displacement information in both the forward and reverse directions of the same phase image.

[0122] S422. Based on the bidirectional displacement field pair, calculate the composite displacement deviation vector and generate the closed-loop displacement deviation vector.

[0123] Optionally, the intelligent prediction terminal can traverse all voxel positions corresponding to the bidirectional displacement field pair and perform a spatial composite operation for each voxel position. The spatial composite operation is a continuous transformation process: first, applying a reverse displacement vector field to the original voxel coordinates to obtain intermediate coordinates; then, applying a forward displacement vector field to the intermediate coordinates to obtain the final composite coordinates. Since the intermediate coordinates output by the reverse displacement vector field are usually non-integer coordinates, the intelligent prediction terminal can use trilinear interpolation to calculate the forward displacement vector value corresponding to this non-integer coordinate position, ensuring the continuity and accuracy of the composite transformation. The intelligent prediction terminal can calculate the difference between the composite coordinates and the original voxel coordinates to obtain the closed-loop displacement deviation vector for that voxel position. The closed-loop displacement deviation vectors for all voxel positions together form the closed-loop displacement deviation vector set.

[0124] S423. Calculate the Euclidean norm of the closed-loop displacement deviation vector to generate the closed-loop error norm. The formula for calculating the closed-loop error norm is:

[0125] ;

[0126] in, voxel position The closed-loop error norm at that point, To calculate the L2 norm of a vector, Let x be the forward displacement of the target displacement vector field at the voxel position x. The reverse displacement vector field at the voxel position The reverse displacement, For spatial compounding operations, The coordinates of the voxel are in space.

[0127] Optionally, the intelligent prediction terminal can calculate the Euclidean norm for each closed-loop displacement deviation vector in the set of closed-loop displacement deviation vectors to generate the closed-loop error norm. The formula for calculating the closed-loop error norm is:

[0128] ;

[0129] in, Let x be the closed-loop error norm at voxel position x. To calculate the L2 norm of a vector, For the target displacement vector field at the voxel position Forward displacement, The reverse displacement vector field at the voxel position The reverse displacement, For spatial compounding operations, Let be the three-dimensional spatial coordinates of the voxel. The Euclidean norm is equal to the square root of the sum of the squares of the three spatial components of the closed-loop displacement deviation vector. It can comprehensively reflect the degree of displacement deviation in the three directions. The larger the value, the worse the self-consistency of the two-way registration result of the corresponding voxel position.

[0130] S424. Aggregate the closed-loop error norms of all voxel locations to generate a closed-loop error norm map.

[0131] Optionally, the intelligent prediction terminal can create a three-dimensional empty array that perfectly matches the size, voxel spacing, and spatial coordinate system of the arterial phase image, serving as the initial carrier for the closed-loop error norm map. The intelligent prediction terminal can iterate through all elements in the closed-loop displacement deviation vector set, and based on the three-dimensional spatial index of each element, assign the corresponding closed-loop error norm to the corresponding coordinate position in the three-dimensional empty array. For each phase in the portal venous phase, delayed phase, and hepatobiliary phase images, the intelligent prediction terminal can generate a corresponding single-phase closed-loop error norm map. All single-phase closed-loop error norm maps together form a complete closed-loop error norm map, which can intuitively reflect the differences in registration self-consistency at different spatial locations.

[0132] In the above embodiments, a bidirectional displacement field pair is formed by extracting the forward displacement field and generating a reverse displacement vector field using a spatially consistent spatial transformation network. The composite displacement deviation vector of the bidirectional displacement field is calculated to obtain the closed-loop displacement deviation vector. The comprehensive deviation degree in three-dimensional space is quantified using Euclidean norm to generate the closed-loop error norm. Finally, the closed-loop error norm map is generated by aggregating according to spatial coordinates. This embodiment can objectively evaluate the registration reliability of each voxel position from the perspective of the self-consistency of the registration results, effectively identify hidden registration misalignments caused by the blurring of the boundary of the peritumoral micro-infiltration area, make up for the limitations of single forward registration error detection, and provide a reliable basis for fusing local deformation anomalies to generate a more accurate multi-phase registration residual confidence map, thereby improving the accuracy and robustness of subsequent feature extraction and postoperative recurrence prediction of hepatocellular carcinoma.

[0133] In one embodiment, S5 may include:

[0134] S51. Perform phase-by-phase three-dimensional convolutional encoding on the multi-phase registration images to generate phase-by-phase convolutional feature maps.

[0135] Optionally, the intelligent prediction terminal can employ a three-dimensional convolutional neural network to extract features from multi-phase registered images phase by phase. The three-dimensional convolutional neural network can simultaneously capture the local morphological and textural features of multi-phase registered images in three spatial dimensions, making it suitable for feature encoding of three-dimensional medical images. The three-dimensional convolutional neural network consists of multiple stacked three-dimensional convolutional layers, modified linear unit activation layers, and three-dimensional pooling layers, transforming the grayscale information of the original image into high-level abstract features through layer-by-layer feature transformation.

[0136] Specifically, the training process of the 3D convolutional neural network is as follows: A large number of multi-phase registration images of hepatocellular carcinoma patients labeled with postoperative recurrence status are collected as the training set. Single-phase registration images are used as network input, and the corresponding patient's recurrence label is used as the supervision signal. The cross-entropy loss function is used as the training loss, and an adaptive moment estimation optimizer is used to iteratively update the network parameters until the loss function value converges to a stable range. After training, the 3D convolutional neural network encodes the arterial phase registration image, portal venous phase registration image, delayed phase registration image, and hepatobiliary phase registration image, respectively, generating four independent phase-wise convolutional feature maps. Each phase-wise convolutional feature map contains multi-level spatial feature information of the corresponding phase image.

[0137] S52. Based on the multi-phase registration residual confidence map, spatial attention masking is applied to the phase-by-phase convolution feature map to generate a residual suppression feature map.

[0138] Optionally, the intelligent prediction terminal can first adjust the size of the multi-phase registration residual confidence map to be completely consistent with the size of the phase-by-phase convolutional feature map. This size adjustment can employ trilinear interpolation to ensure a one-to-one spatial correspondence between the adjusted multi-phase registration residual confidence map and the phase-by-phase convolutional feature map. The intelligent prediction terminal can then use the adjusted multi-phase registration residual confidence map as a spatial attention weight matrix to perform voxel-by-voxel multiplication on the phase-by-voxel convolutional feature map, achieving spatial attention masking. Spatial attention masking dynamically adjusts the contribution of features at each voxel position based on the registration reliability. Voxel positions with higher values ​​in the registration residual confidence map have larger feature weights and higher feature contributions; conversely, voxel positions with lower values ​​in the registration residual confidence map have smaller feature weights and lower feature contributions. Through this method, the intelligent prediction terminal can effectively suppress the propagation of artifact features in areas with large registration errors, while preserving the true biological features of accurately registered areas, generating a residual suppression feature map.

[0139] S53. Based on the multi-phase registration residual confidence map, the residual suppression feature map is penalized with temporal attention weights to generate a spatiotemporal aggregated feature map.

[0140] Optionally, the intelligent prediction terminal can first calculate the global average confidence score of the multi-phase registration residual confidence map corresponding to each phase. The global average confidence score is the arithmetic mean of the registration residual confidence values ​​of all voxel positions in that phase, reflecting the overall registration quality of the phase image. The intelligent prediction terminal can calculate the corresponding temporal attention weight based on the global average confidence score of each phase. The temporal attention weight is proportional to the global average confidence score; the higher the overall registration quality, the larger the corresponding temporal attention weight, and vice versa. The intelligent prediction terminal can globally multiply the residual suppression feature map of each phase with the corresponding temporal attention weight to implement temporal attention weight penalty, globally suppressing the features of phases with poor overall registration quality and balancing the differences in registration reliability between different phase images. The intelligent prediction terminal can then stitch the four phase feature maps after temporal weighting along the channel dimension to generate a spatiotemporal aggregated feature map that simultaneously contains spatial and temporal dimensional information.

[0141] S54. Perform global average pooling on the spatiotemporal aggregated feature map to generate a composite feature vector.

[0142] Optionally, the intelligent prediction terminal can perform global average pooling on the spatiotemporal aggregated feature map. Global average pooling calculates the arithmetic mean of the positional feature values ​​of all voxels within each feature channel of the spatiotemporal aggregated feature map, converting each three-dimensional feature channel into a scalar value. Global average pooling can effectively reduce the feature dimensionality, reduce subsequent computation, and avoid the overfitting problem that is prone to occur in fully connected layers, while preserving the global semantic information of the feature map.

[0143] Optionally, the intelligent prediction terminal can arrange the scalar values ​​corresponding to all feature channels in channel order to generate a one-dimensional composite feature vector. The composite feature vector integrates feature information from multi-phase registration images at different spatial locations and time stages, and has undergone residual suppression processing in both spatial and temporal dimensions, providing a foundation for feature decoupling and classification mapping.

[0144] In the above embodiment, a three-dimensional convolutional neural network is used to encode multi-phase registration images in stages to extract multi-level spatial features. Based on the confidence map of multi-phase registration residuals, spatial attention masking is applied sequentially to suppress local artifact features, and temporal attention weights are used to penalize the registration quality differences between phases in the equilibrium period. Finally, global average pooling is used to generate a composite feature vector that integrates spatiotemporal information. This embodiment can simultaneously suppress artifact features introduced by registration errors from both spatial and temporal dimensions, maximizing the preservation of real biological features related to postoperative recurrence of hepatocellular carcinoma, and effectively improving the robustness and discriminative power of feature expression.

[0145] In one embodiment, S6 may include:

[0146] S61. Based on the preset residual basis vector matrix, the composite feature vector is projected into the column space to generate residual feature components.

[0147] Optionally, the intelligent prediction terminal can load a preset residual basis vector matrix, which can be pre-trained using principal component analysis. The pre-training process is as follows: collect a large number of multi-phase magnetic resonance image samples containing different degrees of registration errors, perform the aforementioned steps on each sample to generate a corresponding composite feature vector, and construct a sample set containing a large number of registration artifact features. Perform principal component analysis on the sample set, extract several principal component vectors with the largest variance contribution as residual basis vectors, and arrange all residual basis vectors in columns to form a residual basis vector matrix. The subspace spanned by the residual basis vector matrix is ​​the main distribution space of the registration artifact features. The intelligent prediction terminal can project the composite feature vectors into the column space of the residual basis vector matrix and calculate the projection components of the composite feature vectors in the artifact shadow space, i.e., the residual feature components. The specific calculation formula is as follows:

[0148] ;

[0149] in, For residual characteristic components, Let be the residual basis vector matrix. It is a composite feature vector. -1 is the matrix transpose operation, and -1 is the matrix inversion operation.

[0150] S62. Subtract the residual feature components from the composite feature vector to generate an orthogonalized intermediate feature vector.

[0151] Optionally, the intelligent prediction terminal can perform element-wise subtraction between the composite feature vector and the residual feature components to generate an orthogonalized intermediate feature vector. The orthogonalized intermediate feature vector is the component of the composite feature vector that is orthogonal to the column space of the residual basis vector matrix. According to the basic principle of orthogonal projection, the composite feature vector can be uniquely decomposed into two parts: the residual feature components located in the artifact space and the orthogonalized intermediate feature vector orthogonal to the artifact space. The residual feature components mainly contain artifact features introduced by registration errors, while the orthogonalized intermediate feature vector initially removes most of the artifact features, preserving the true biological features related to postoperative recurrence of hepatocellular carcinoma. Through a simple vector subtraction operation, the intelligent prediction terminal can quickly achieve the initial separation of artifact features and biological features.

[0152] S63. Minimize the inner product norm constraint on the orthogonalized intermediate eigenvectors and residual basis vector matrices to generate the orthogonality constraint loss function. The formula for calculating the orthogonality constraint loss function is as follows:

[0153] ;

[0154] in, The orthogonal constraint loss value. To calculate the matrix norm, This is the transpose of the orthogonalized intermediate eigenvectors. Let be the residual basis vector matrix.

[0155] Optionally, the intelligent prediction terminal can construct an orthogonality constraint loss function to quantify the degree of orthogonality between the orthogonalized intermediate feature vectors and the residual basis vector matrix, further optimizing the thoroughness of feature separation. The formula for calculating the orthogonality constraint loss function is:

[0156] ;

[0157] in, The orthogonal constraint loss value. To calculate the Frobenius norm of a matrix, the Frobenius norm is equal to the square root of the sum of the squares of all its elements. It provides a comprehensive measure of the overall size of the matrix. This is the transpose of the orthogonalized intermediate eigenvectors. Let be the residual basis vector matrix. When the orthogonalized intermediate eigenvectors are completely orthogonal to the residual basis vector matrix, all elements of their product matrix are 0, and the orthogonal constraint loss value is 0. The smaller the loss value, the more thorough the feature separation.

[0158] S64. Based on the orthogonality constraint loss function, the orthogonalized intermediate feature vector is iteratively updated to generate a pure biological feature vector.

[0159] Optionally, the intelligent prediction terminal can minimize the orthogonality constraint loss function as the optimization objective and iteratively update the orthogonalized intermediate feature vector using gradient descent. During each iteration, the intelligent prediction terminal can calculate the gradient of the orthogonality constraint loss function relative to the orthogonalized intermediate feature vector, and adjust the element values ​​of the orthogonalized intermediate feature vector according to the gradient direction and a preset learning rate, so that the orthogonality constraint loss value gradually decreases. The iteration process terminates when the orthogonality constraint loss value converges to a preset stable range, or when the number of iterations reaches a preset maximum number of iterations. After iteration, the orthogonality between the orthogonalized intermediate feature vector and the residual basis vector matrix reaches its optimal level, containing almost no registration artifact features. The intelligent prediction terminal can then determine this orthogonalized intermediate feature vector as a pure biological feature vector.

[0160] In the above embodiment, the composite feature vector is projected onto the artifact feature subspace using a preset residual basis vector matrix to obtain residual feature components. Vector subtraction is then used to initially separate the artifacts from the biological features. Finally, an orthogonality-constrained loss function is constructed to iteratively optimize the feature orthogonality, ultimately generating a pure biological feature vector. This embodiment can completely separate the artifact features introduced by registration errors from the true biological features in the composite feature vector, effectively eliminating the interference of registration residuals on subsequent classification predictions. It significantly improves the accuracy and generalization ability of the hepatocellular carcinoma postoperative recurrence risk prediction model, providing more reliable technical support for the development of personalized postoperative follow-up and intervention plans in clinical practice.

[0161] The aforementioned intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images acquires multi-phase magnetic resonance images including the arterial phase, portal venous phase, delayed phase, and hepatobiliary phase, and segments them to generate a tumor body mask. An adaptive morphological dilation is used to generate a probability map reflecting the spatial distribution of peritumoral microinvasiveness. Based on this, a spatially variable adaptive smoothing regularization loss function-driven spatial transformation network is constructed to achieve residual-aware adaptive registration for the peritumoral region, solving the problem of voxel misalignment easily generated in low-contrast microinvasive regions by traditional global smoothing registration. Subsequently, a multi-phase registration residual confidence map is generated by fusing local deformation anomalies and bidirectional closed-loop error quantization to accurately characterize the spatial distribution of registration errors. Then, residual-aware features are extracted based on the confidence map in both spatial and temporal dimensions to initially suppress artifact feature propagation. Finally, registration artifact features and real biological features are completely separated through residual orthogonal projection decoupling, and a postoperative recurrence risk score for hepatocellular carcinoma is generated based on the pure features.

[0162] This technical solution effectively solves the core problems of large registration errors in the peritumoral microinvasive area and the inability to effectively separate artifact features from real biological features in traditional methods for predicting recurrence after hepatocellular carcinoma surgery, which leads to low prediction accuracy. It significantly improves the accuracy and robustness of risk assessment for postoperative recurrence of hepatocellular carcinoma, and provides reliable technical support for the development of personalized postoperative follow-up and intervention plans in clinical practice.

[0163] 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.

[0164] Based on the same inventive concept, this application also provides an apparatus for implementing the above-mentioned intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent prediction device for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging provided below can be found in the limitations of the intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging described above, and will not be repeated here.

[0165] In one exemplary embodiment, such as Figure 2 As shown, a smart prediction device 10 for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging is provided, comprising:

[0166] The tumor segmentation module 11 can be used to acquire multi-phase magnetic resonance images of the target object, segment the tumor region in the arterial phase image, and generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images.

[0167] The probability map generation module 12 can be used to adaptively expand the tumor body mask to generate a probability map of peritumoral micro-infiltration.

[0168] The residual registration module 13 can be used to perform residual-aware adaptive registration of arterial phase images with portal venous phase images, delayed phase images and hepatobiliary phase images based on the peritumoral microinfiltration probability map, and generate target displacement vector field and multi-phase registered images.

[0169] The residual quantization module 14 can be used to perform local deformation anomaly and closed-loop error fusion quantization on the target displacement vector field to generate a multi-phase registration residual confidence map.

[0170] The spatiotemporal decoupling module 15 can be used to extract residual-aware spatiotemporal decoupling features from multi-phase registration images based on multi-phase registration residual confidence maps, and generate composite feature vectors.

[0171] The orthogonal decoupling module 16 can be used to perform residual orthogonal projection decoupling on composite feature vectors to generate pure biological feature vectors.

[0172] The risk scoring module 17 can be used to perform classification mapping based on pure biological feature vectors to generate a risk score for postoperative recurrence of hepatocellular carcinoma.

[0173] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images as described above.

[0174] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent prediction method for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance images as described above.

[0175] 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.

[0176] 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 intelligent prediction of postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging, characterized in that, The method includes: S1. Acquire multi-phase magnetic resonance images of the target object, and segment the tumor region in the arterial phase image to generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images. S2. Adaptive morphological expansion is performed on the tumor body mask to generate a peritumoral micro-invasion probability map; S3. Based on the peritumoral microinfiltration probability map, residual-aware adaptive registration is performed on the arterial phase image, the portal venous phase image, the delayed phase image, and the hepatobiliary phase image to generate a target displacement vector field and a multi-phase registration image; S4. Perform local deformation anomaly and closed-loop error fusion quantization on the target displacement vector field to generate a multi-phase registration residual confidence map; S5. Based on the multi-phase registration residual confidence map, perform residual perception spatiotemporal decoupling feature extraction on the multi-phase registration image to generate a composite feature vector; S6. Perform residual orthogonal projection decoupling on the composite feature vector to generate a pure biological feature vector; S7. Based on the pure biological feature vector, perform classification mapping to generate a risk score for recurrence after hepatocellular carcinoma surgery.

2. The method according to claim 1, characterized in that, The S3 includes: S31. Calculate the local smoothing penalty weight based on the peritumoral microinfiltration probability map and generate an adaptive penalty weight map. S32. Construct an adaptive smoothing regularization loss function based on the adaptive penalty weight graph. The calculation formula for the adaptive smoothing regularization loss function is as follows: ； in, To adaptively smooth the regularization loss value, Voxel positions in the adaptive penalty weight graph The weight value at that point, For the target displacement vector field at the voxel position Spatial gradient at that location To calculate the square norm; S33. The spatial transformation network is jointly driven by the adaptive smoothing regularization loss function and the image similarity loss function, and nonlinear deformation is applied to the portal venous phase image, the delayed phase image and the hepatobiliary phase image to generate an initial displacement vector field. S34. Based on the initial displacement vector field, perform spatial distortion transformation on the portal venous phase image, the delayed phase image, and the hepatobiliary phase image to generate the multi-phase registration image, and determine the initial displacement vector field as the target displacement vector field.

3. The method according to claim 2, characterized in that, S31 includes: S311. Extract the micro-infiltration probability value of each voxel position in the peritumoral micro-infiltration probability map to generate a voxel micro-infiltration probability set. S312. Based on the set of voxel micro-infiltration probabilities and the preset global basic weights, calculate the adaptive smoothing penalty weight for each voxel position and generate a set of voxel penalty weights. S313. Combine the adaptive smoothing penalty weights of all voxel positions in the voxel penalty weight set to generate an adaptive penalty weight map. S314. Based on the adaptive penalty weight graph, spatially weight the standard smoothing regularization term to generate the adaptive smoothing regularization loss function.

4. The method according to claim 1, characterized in that, The S4 includes: S41. Based on the target displacement vector field, calculate the absolute value deviation of the determinant of the spatial Jacobian matrix and generate a local deformation anomaly map. S42. Based on the target displacement vector field and the corresponding reverse displacement vector field, calculate the identity transformation deviation norm and generate a closed-loop error norm map. S43. Multiply and fuse the local deformation anomaly map and the closed-loop error norm map to generate the original residual intensity map; S44. Perform Gaussian low-pass filtering and normalization mapping on the original residual intensity map to generate the multi-phase registration residual confidence map.

5. The method according to claim 1, characterized in that, S42 includes: S421. Extract the forward displacement field based on the target displacement vector field, and extract the reverse displacement vector field based on the corresponding reverse registration network to generate a two-way displacement field pair. S422. Based on the bidirectional displacement field pair, calculate the composite displacement deviation vector and generate the closed-loop displacement deviation vector. S423. Calculate the Euclidean norm of the closed-loop displacement deviation vector to generate the closed-loop error norm. The formula for calculating the closed-loop error norm is as follows: ； in, voxel position The closed-loop error norm at that point, To calculate the L2 norm of a vector, Let x be the forward displacement of the target displacement vector field at the voxel position x. The reverse displacement vector field at the voxel position The reverse displacement, For spatial compounding operations, For the spatial coordinates of the voxel; S424. Aggregate the closed-loop error norms at all voxel locations to generate the closed-loop error norm map.

6. The method according to claim 1, characterized in that, The S5 includes: S51. Perform phase-by-phase three-dimensional convolutional encoding on the multi-phase registration image to generate phase-by-phase convolutional feature maps; S52. Based on the multi-phase registration residual confidence map, spatial attention masking is applied to the phase-by-phase convolution feature map to generate a residual suppression feature map; S53. Based on the multi-phase registration residual confidence map, apply temporal attention weight penalty to the residual suppression feature map to generate a spatiotemporal aggregation feature map; S54. Perform global average pooling on the spatiotemporal aggregated feature map to generate the composite feature vector.

7. The method according to claim 1, characterized in that, The S6 includes: S61. Based on the preset residual basis vector matrix, the composite feature vector is projected into the column space to generate residual feature components; S62. Subtract the residual feature component from the composite feature vector to generate an orthogonalized intermediate feature vector; S63. Minimize the inner product norm constraint on the orthogonalized intermediate eigenvector and the residual basis vector matrix to generate an orthogonality constraint loss function. The formula for calculating the orthogonality constraint loss function is as follows: ； in, The orthogonal constraint loss value. To calculate the matrix norm, This is the transpose of the orthogonalized intermediate eigenvectors. The residual basis vector matrix; S64. Based on the orthogonality constraint loss function, the orthogonalized intermediate feature vector is iteratively updated to generate the pure biological feature vector.

8. A smart predictive device for postoperative recurrence of hepatocellular carcinoma based on multi-phase magnetic resonance imaging, characterized in that, The device includes: The tumor segmentation module is used to acquire multi-phase magnetic resonance images of the target object, and to segment the tumor region in the arterial phase image to generate a tumor mask. The multi-phase magnetic resonance images include arterial phase images, portal venous phase images, delayed phase images, and hepatobiliary phase images. The probability map generation module is used to adaptively morphologically expand the tumor body mask to generate a probability map of peritumoral microinfiltration. The residual registration module is used to perform residual-aware adaptive registration of the arterial phase image with the portal venous phase image, the delayed phase image and the hepatobiliary phase image based on the peritumoral microinfiltration probability map, and generate a target displacement vector field and a multi-phase registration image; The residual quantization module is used to perform fusion quantization of local deformation anomalies and closed-loop errors on the target displacement vector field to generate a multi-phase registration residual confidence map. The spatiotemporal decoupling module is used to extract residual-aware spatiotemporal decoupling features from the multi-phase registration image based on the multi-phase registration residual confidence map, and generate a composite feature vector. An orthogonal decoupling module is used to perform residual orthogonal projection decoupling on the composite feature vector to generate a pure biological feature vector. The risk scoring module is used to perform classification mapping based on the pure biological feature vector to generate a risk score for postoperative recurrence of hepatocellular carcinoma.

9. 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 7.

10. 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 7.

Images