A method and system for radiology image data parsing
By registering and extracting features from PET and ultrasound images, and performing geometric correction and multi-scale feature fusion of CT images, the problems of multimodal image registration errors and vascular malformations have been solved, enabling highly sensitive detection and precise localization of malignant lesions.
Patent Information
- Application Number
- CN202510543847.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-28
- Publication Date
- 2025-11-11
- Estimated Expiration
- 2045-04-28
AI Technical Summary
In existing technologies, the accumulation of multimodal medical image registration errors, insufficient correction of vascular geometric distortions, and inadequate effectiveness of cross-scale feature fusion affect the accuracy of tumor diagnosis and vascular lesion analysis.
Physiological metabolic feature vectors are generated by registering and extracting features from PET and ultrasound images; polar coordinate sinusoidal distortion enhancement and scan bed removal are performed on CT images to generate geometrically corrected 3D medical images; multi-scale Gaussian pyramid decomposition is performed to extract cross-scale feature tensors; a dynamic hypergraph is constructed based on physiological metabolism and manifold feature vectors to generate cross-modal fusion feature vectors; and a diagnostic report is generated by combining the elastic wave equation.
It improves the sensitivity of malignant lesion detection, enables precise localization of structural-functional synergistic abnormalities, and provides a highly reliable basis for clinical decision-making.
Smart Images

Figure CN120070440B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for parsing radiological image data. Background Technology
[0002] In recent years, multimodal medical image fusion technology has played an important role in tumor diagnosis and vascular lesion analysis. Among them, affine transformation-based PET-CT-ultrasound multimodal registration, vascular structure enhancement, and cross-scale feature extraction have become research hotspots. Existing technologies mostly employ rigid registration or non-rigid registration algorithms based on mutual information. However, due to neglecting the initial parameter optimization of image metadata (such as scan bed position), registration errors accumulate, especially in preserving the spatial consistency of hemodynamic parameters (such as the drag index). Furthermore, geometric distortion correction of CT images often relies on isotropic filtering or morphological processing, making it difficult to effectively distinguish between scan bed artifacts and marginal vessels. Traditional Gaussian pyramid decomposition also tends to lose high-frequency details, limiting the cross-scale correlation of subsequent feature encoding.
[0003] The limitations of existing technologies are mainly reflected in two aspects: First, parameter optimization during image registration relies on global search, lacks constraints on the complexity of transformation parameters, and is prone to getting trapped in local optima, affecting the accurate mapping of metabolic activity standardized values; Second, vascular structure enhancement uses fixed threshold segmentation or linear interpolation, which cannot simulate the dynamic influence of periodic physiological movements on vascular deformation, leading to interruption of vascular continuity, which in turn affects the accuracy of subsequent manifold topology analysis and biomechanical modeling. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a method for parsing radiological image data to address the problems of accumulated multimodal image registration errors, insufficient correction of vascular geometric distortions, and inadequate effectiveness of cross-scale feature fusion.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] In a first aspect, the present invention provides a method for parsing radiological image data, comprising,
[0008] Register and extract features from PET and ultrasound images to generate physiological metabolic feature vectors;
[0009] Polar coordinate sinusoidal distortion enhancement and scan bed removal are performed on CT images to generate geometrically corrected three-dimensional medical images;
[0010] Multi-scale Gaussian pyramid decomposition is performed on geometrically corrected 3D medical images. Cross-scale feature tensors are extracted through coding state construction and entanglement operation, and the compressed composite feature tensor is output.
[0011] Based on the real part of the compressed composite features, the organ surface manifold is reconstructed, and the manifold feature vector is obtained by combining Gaussian curvature detection and blood vessel continuous cohomology analysis.
[0012] Based on physiological metabolic feature vectors and manifold feature vectors, a dynamic hypergraph is constructed to generate cross-modal fusion feature vectors;
[0013] Input the cross-modal fusion feature vector into the constraint generator, solve the elastic wave equation, and generate a diagnostic report.
[0014] As a preferred embodiment of the radiology image data analysis method described in this invention, the registration and feature extraction include the following steps:
[0015] Based on the patient's position and the scanning bed position, initialize the translation and scaling components of the affine transformation matrix;
[0016] The registration parameters were iteratively optimized using a quasi-Newton method to minimize the registration error between PET and ultrasound images;
[0017] The registered PET images are mapped to the CT space by inverse transformation, and the metabolic activity values of each voxel are calculated using the SUV normalization formula.
[0018] The tumor boundary was delineated in CT space and mapped to PET images. The maximum SUV value and the volume of voxels exceeding the threshold were extracted, and physiological metabolic feature vectors were generated by combining hemodynamic parameters.
[0019] As a preferred embodiment of the radiology image data analysis method described in this invention, the generation of geometrically corrected three-dimensional medical images includes the following steps.
[0020] The CT voxel data were converted to cylindrical coordinates and subjected to periodic sinusoidal perturbations.
[0021] Trilinear interpolation is used to inversely map the perturbed cylindrical coordinates to the Cartesian coordinate system to generate a distortion-enhanced image;
[0022] A three-dimensional binary mask is constructed based on the scanning bed position parameters. Bed artifacts are suppressed by pixel-by-pixel multiplication, and geometrically corrected three-dimensional medical images are output.
[0023] As a preferred embodiment of the radiology image data parsing method described in this invention, the extraction of cross-scale feature tensors includes the following steps:
[0024] A three-level Gaussian pyramid decomposition was performed on the geometrically corrected 3D medical images to construct spatial domain coding states and frequency domain enhanced coding states.
[0025] By fusing the spatial domain encoded state and the frequency domain enhanced encoded state through a parameterized entanglement gate, a composite state is generated, which is then compressed into a composite feature tensor using a complex orthogonal matrix projection.
[0026] As a preferred embodiment of the radiology image data analysis method described in this invention, the reconstruction of the organ surface manifold includes the following steps.
[0027] Based on the complex real part of the encoded state of the compressed composite feature tensor, the isosurface grid is extracted, the Riemannian metric tensor and discrete Gaussian curvature detection are calculated, and the one-dimensional continuous homology features of the vascular complex are extracted by combining the α-shape algorithm. The curvature anomaly and topological invariant are fused to generate the manifold feature vector.
[0028] As a preferred embodiment of the radiology image data parsing method described in this invention, the construction of the dynamic hypermap includes the following steps.
[0029] Physiological metabolic feature vectors and manifold feature vectors are concatenated to form hypergraph nodes, and hyperedge correlation matrices are dynamically generated based on Pearson correlation coefficients.
[0030] By aggregating cross-modal information through hypergraph convolution and backpropagating it, combined with degree matrix normalization, a fusion feature vector capturing the co-variation of structure and function is output.
[0031] As a preferred embodiment of the radiology image data analysis method described in this invention, the generation of the diagnostic report includes the following steps:
[0032] A finite element model is constructed based on fused feature vectors to solve the nodal displacement field, and the deformation is quantified and compared with a health threshold to mark the diseased area.
[0033] The probability of malignancy is predicted by Monte Carlo sampling of a random perturbation model, and a multimodal diagnostic report is generated by combining Gaussian curvature, continuous homology, and maximum SUV value.
[0034] Secondly, the present invention provides a system for analyzing radiological image data, comprising,
[0035] The image registration module registers and extracts features from PET metabolic images and ultrasound images to generate physiological metabolic feature vectors.
[0036] The geometric correction module performs polar coordinate sinusoidal distortion enhancement and scan bed removal on CT images to generate geometrically corrected three-dimensional medical images.
[0037] The encoding module performs multi-scale Gaussian pyramid decomposition on the geometrically corrected 3D medical image, extracts cross-scale feature tensors through encoding state construction and entanglement operation, and outputs compressed composite feature tensors.
[0038] The topology analysis module reconstructs the organ surface manifold based on the real part of the compressed composite features, and combines Gaussian curvature detection and blood vessel continuous homology analysis to obtain the manifold feature vector.
[0039] The hypergraph fusion module constructs a dynamic hypergraph based on physiological metabolic feature vectors and manifold feature vectors, generating cross-modal fusion feature vectors;
[0040] The elasticity diagnostic module takes the cross-modal fused feature vector as input to the constraint generator, solves the elastic wave equation, and generates a diagnostic report.
[0041] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the method for parsing radiological image data as described in the first aspect of the present invention.
[0042] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the radiology image data parsing method as described in the first aspect of the present invention.
[0043] The beneficial effects of this invention are as follows: In CT image geometric correction, the natural deformation of blood vessels affected by heartbeat is simulated by periodic sinusoidal perturbation and trilinear interpolation in cylindrical coordinates, while the scanning bed mask is used to accurately suppress artifacts; Multi-scale feature encoding solves the problem of high-frequency detail loss in traditional Gaussian pyramid decomposition by dynamically allocating weights through gradient amplitude and combining entanglement operations of spatial-frequency domain encoded states; In the biomechanical diagnosis stage, the elastic wave equation combined with the Monte Carlo random perturbation model quantifies deformation anomalies through finite element node displacement fields, and combines morphological-metabolic-topological multi-dimensional parameters to significantly improve the sensitivity of malignant lesion detection; Through dynamic hypergraph fusion of cross-modal features, the precise localization of structural-functional synergistic anomalies is achieved, providing a highly reliable basis for clinical decision-making. Attached Figure Description
[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1This is a flowchart of the radiology image data parsing method in Example 1.
[0046] Figure 2 This is a schematic diagram of image registration and feature extraction in Example 1.
[0047] Figure 3 This is a schematic diagram of feature extraction and manifold reconstruction in Example 1.
[0048] Figure 4 This is a schematic diagram of dynamic hypergraph construction and elasticity diagnosis in Example 1. Detailed Implementation
[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0052] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides a method for parsing radiological image data, including the following steps:
[0053] S1. Define an affine transformation matrix including translation components, rotation components, and scaling factor transformation parameters; solve for the optimal transformation parameters in the affine transformation matrix using an optimization algorithm to minimize the registration error and constrain the complexity of the parameters; set initial translation and scaling based on image metadata (such as patient position and scanning bed position); use the quasi-Newton method to iteratively update the optimal transformation parameters until the registration error change rate reaches the maximum number of iterations to obtain the optimized affine transformation matrix; map the PET image and ultrasound image to CT space through the inverse transformation of the affine transformation matrix for alignment to obtain the registered PET metabolic image and ultrasound image.
[0054] To further illustrate, when ultrasound images are aligned to CT space, spatial consistency between blood flow velocity and resistance index is preserved.
[0055] For the registered PET images, the standardized value of metabolic activity is calculated for each voxel, expressed as:
[0056] ;
[0057] in, Represents voxels, Voxel representation Standardized intake values, dimensionless. This indicates that the PET image after registration is in voxel Radioactivity count at the location. Indicates the injection dosage. Indicates patient weight.
[0058] The tumor boundary is delineated in CT space and mapped to the registered PET image. All voxels within the boundary are traversed, and the maximum metabolic activity standardized value is recorded. A critical threshold for the metabolic activity standardized value is set, and the volume of voxels exceeding the critical threshold is selected. The voxel volume is calibrated according to the CT resolution to obtain the calibrated voxel volume. The physiological metabolic feature vector of PET and ultrasound images is formed by combining the maximum metabolic activity standardized value, voxel volume, blood flow velocity and resistance index.
[0059] S2. Store the CT images as DICOM format CT voxel data; for each slice in the CT voxel data, convert from Cartesian coordinates to cylindrical coordinates with the center of the 3D image as the origin, while keeping the inter-slice coordinates unchanged. The cylindrical coordinate system includes radial distance, polar angle, and axial height, as expressed below:
[0060] ;
[0061] in, This represents the radial distance, which is the distance from the voxel to the center of the image. Represents the three-dimensional voxel coordinates in the original Cartesian coordinate system. This represents the coordinates of the image center, used to translate the origin of the Cartesian coordinate system to the image center. The polar angle (in radians) represents the azimuth angle of a voxel relative to its center point, ranging from [0, 2π). The height axis corresponds directly to the Z-axis (layer number or slice number) in the Cartesian coordinate system.
[0062] To further explain, radial distance refers to the distance from the center in the XY plane, polar angle refers to the rotation angle in the XY plane, and axial height refers to the layer position along the Z-axis, representing the height axis, which directly corresponds to the Z-axis of the Cartesian coordinate system (layer number or slice number).
[0063] By jointly applying periodic sinusoidal perturbations in the radial dimension and layer direction, the continuity of blood vessels and organs is enhanced, as expressed by:
[0064] ;
[0065] in, Indicates the radial distance after the disturbance. This indicates the amplitude of the disturbance and controls the intensity of vascular deformation. It represents the angular frequency, determines the number of periods of a sine wave, and simulates periodic physiological movements (such as heartbeat). This represents the radial threshold, used to limit the disturbance to act only on... The voxels protect the central region (such as large blood vessels) from being affected, representing a rectangular function, defined as:
[0066] ;
[0067] when At that time, the central area is shielded, and disturbance is applied only to the peripheral blood vessels.
[0068] The perturbed cylindrical coordinates are mapped back to Cartesian coordinates, and a distortion-enhanced 3D image is generated through trilinear interpolation, as shown in the following expression:
[0069] The expression for the inverse mapping to the Cartesian coordinate system is:
[0070] ;
[0071] in, This represents the perturbed Cartesian coordinates. Represents the polar angle corresponding to the original Cartesian coordinates. The trigonometric function values.
[0072] The trilinear interpolation formula is:
[0073] ;
[0074] in, This indicates the distortion-enhanced 3D image in the original Cartesian coordinates. voxel values at the location, , , This means rounding down the perturbed Cartesian coordinates to obtain the nearest neighbor integer grid point. This represents a 3D neighborhood index, traversing the 8 neighboring voxels around the perturbed Cartesian coordinates. Represents the weighting factor, distortion coordinates and adjacent voxels in Distance decay weights in the direction (linear interpolation).
[0075] It should be noted that the purpose of inverse mapping to the Cartesian coordinate system is to restore the perturbed cylindrical coordinates to Cartesian coordinates and generate the distorted spatial position. Trilinear interpolation solves the problem that the distorted coordinates may be non-integer. By smoothly distributing the contribution values of the surrounding 8 voxels through trilinear interpolation, jagged artifacts are avoided and the continuity of the blood vessel edge is maintained.
[0076] For each slice, the number of slice layers covered in the Z-axis direction is determined based on the physical position of the scanning bed; within each Z-axis slice, the projected radius of the scanning bed is calculated from the reconstructed field of view diameter and the bed position, expressed as:
[0077] ;
[0078] in, This indicates the coverage radius of the scanning bed in the XY plane. Indicates the diameter of the reconstructed field of view. This indicates the position parameters of the scanning bed in the Z-axis direction.
[0079] Based on the projection radius of the scanning bed, a three-dimensional binary mask is defined, marking the scanning bed area as 0 and the anatomical area as 1, with the expression as follows:
[0080] ;
[0081] in, Represents a three-dimensional binary mask. Indicates the starting position of the slice. This indicates the end position of the slice.
[0082] Each mask layer is multiplied pixel-by-pixel with the distortion-enhanced 3D image to suppress scanning bed artifacts, resulting in a geometrically corrected 3D image.
[0083] To further explain, the voxel value in the scanning bed region (Mask=0) is set to zero to eliminate metal artifacts or interference from the bed signal, while the anatomical region (Mask=1) retains the distorted and enhanced vascular signal.
[0084] S3. Perform three levels of downsampling on the geometrically corrected image. The first level of sampling is to perform 2×2 neighborhood average pooling on the geometrically corrected image, downsampling to... ,in and The height and width of the geometrically corrected image are defined; the second-level downsampling involves repeatedly performing average pooling and downsampling on the first-layer image to... The third-level downsampling involves performing average pooling and downsampling on the second-layer image to the target resolution. .
[0085] To further explain, the Gaussian pyramid extracts global structural information from the image step by step through multi-scale decomposition with three downsampling steps, thus suppressing high-frequency details.
[0086] The gradient of the image after multi-scale decomposition is calculated using the Sobel operator, expressed as follows:
[0087] ;
[0088] in, This represents the image after three downsampling steps. and Indicates coordinates and gradient components on, and Indicates the image in coordinates and Sobel convolution kernels on top.
[0089] Based on the image gradient, calculate the gradient magnitude (L1 norm) of each image, expressed as:
[0090] ;
[0091] in, This represents the L1 norm of the gradient.
[0092] The weight coefficients are generated based on the gradient magnitude, and the expression is as follows:
[0093] ;
[0094] in, Represents the coordinates of the image center The weighting coefficients.
[0095] Based on the gradient magnitude, calculate the global normalization factor to ensure that the total energy of the coded state is 1. The expression is:
[0096] ;
[0097] in, This represents the normalization factor.
[0098] Combining the weights with the position ground state generates the spatial domain encoded state, expressed as:
[0099] ;
[0100] in, Represents the spatial domain coding state. This represents the two-dimensional position ground state.
[0101] The multi-scale decomposed image is mapped to the frequency domain using a Fast Fourier Transform (FFT). A Gaussian low-pass filter is designed to suppress high-frequency noise and preserve low-frequency texture in the frequency domain. The Gaussian filter is then multiplied element-wise with the frequency domain signal to obtain a filtered complex matrix in the frequency domain, enhancing low-frequency features. The amplitude spectrum of the filtered frequency domain signal is calculated and normalized, as shown in the expression:
[0102] ;
[0103] in, Represents the normalized complex matrix in the frequency domain. Represents frequency coordinates. This represents the filtered frequency domain complex matrix. The frequency domain complex matrix representing the original frequency domain signal. This represents a very small constant, avoiding the error of dividing by zero when the amplitude spectrum is zero.
[0104] By using inverse Fourier transform, the normalized frequency domain complex matrix is mapped back to the spatial domain to obtain the reconstructed spatial domain signal, which is the frequency domain enhanced coded state.
[0105] The spatial domain encoded state and the frequency domain enhanced encoded state are combined into a high-dimensional composite state through tensor product, where tensor product represents the independent superposition of spatial and frequency domain features. A parameterized entanglement gate is constructed to force the spatial coordinates to be associated with the frequency domain pattern, thereby achieving feature fusion and forming a composite state. An entanglement gate is applied to the composite state to generate an entangled state. A learnable complex orthogonal matrix that satisfies orthogonal constraints is defined, and the frequency domain pattern vector at each spatial location is linearly projected to obtain the compressed composite feature tensor.
[0106] S4. The complex real part of the encoded state of the input compressed composite feature tensor is used to extract the isosurface grid using the improved MarchingCubes algorithm, defined as follows:
[0107] ;
[0108] in, The triangulated isosurface mesh represents the differential manifold model of the blood vessel. Represents the vertices of a 3D mesh. Represents the vertex index. Represents a three-dimensional real vector space. Indicates entangled state at vertex The real part of the value reflects the characteristic strength. This represents the global mean of the real parts. The standard deviation of the real part of the value. This represents a dynamic threshold that filters out regions with low feature intensity (such as outliers or noise).
[0109] Define local parameterized coordinates at the grid vertices and compute the Riemannian metric tensor, expressed as:
[0110] ;
[0111] in, Represents the Riemannian metric tensor, which describes the local geometric properties of a surface. and Represents locally parameterized coordinates. Represents the vertex index.
[0112] Based on the discrete Gaussian curvature formula, the curvature of the vertex is calculated, and the expression is:
[0113] ;
[0114] in, The Gaussian curvature of the vertex is represented by... Represents the set of vertices in the 1-ring neighborhood of a vertex. This represents the interior angle of the neighborhood triangle at its vertex. This represents the Voronoi area of a vertex (the weighted sum of the areas of neighboring triangles). This represents a neighborhood index.
[0115] To further explain, Gaussian curvature reflects the degree of local bending:
[0116] >0 indicates a protrusion (such as a blood vessel bifurcation point); <0 indicates a depression (such as a narrowing of a blood vessel); =0 indicates a flat region.
[0117] Abnormal curvature vertices are selected from Gaussian curvature as candidate lesion regions; the α-shape algorithm is applied to construct vascular complexes and extract topological invariants; one-dimensional continuous homology features (ring structure) of the complexes are extracted; the candidate lesion regions, the extracted topological invariants, and the one-dimensional continuous homology features of the complexes are fused to generate manifold feature vectors of vascular geometry and topology.
[0118] S5. Concatenate the manifold feature vector and the physiological metabolic feature vector along the feature dimension to generate hypergraph node features; select the center node based on feature importance, traverse all nodes, use the Pearson correlation coefficient to calculate the correlation between the current node and the center node, set a correlation threshold based on experience, and add nodes with correlation higher than the correlation threshold to the corresponding hyperedge to form a dynamic association group matrix.
[0119] To further explain, a hyperedge represents a group of vascular regions that exhibit synergistic changes in both structural and functional modes. For example, abnormal curvature and increased metabolic activity in a segment of a blood vessel may belong to the same hyperedge.
[0120] By combining the initial features of hypergraph nodes and the association matrix of hyperedges, a degree matrix of nodes and hyperedges is constructed for normalized feature propagation. The degree matrices of nodes and hyperedges are transformed by the weight matrix in the hypergraph convolution formula to learn the nonlinear combination of cross-modal features and enhance feature representation. Hyperedge information aggregation is performed on the transformed degree matrices of nodes and hyperedges, aggregating the cross-modal information received by each node through the hyperedge, and converting the cross-modal information into node-hyperedge relationships. This enables the back propagation of information from hyperedges to nodes, progressively increasing the feature dimension, generating cross-modal fusion features, and capturing global cross-modal associations.
[0121] S6. Extract vascular elastic parameters and ultrasound hemodynamic data from physiological metabolic characteristics. Based on the biomechanical properties and clinical data of healthy blood vessels, set elastic parameters with healthy reference values and lesion correction values, as well as tissue density and vibration frequency. Set fixed constraints for proximal vascular displacement and apply dynamic loads of periodic pressure to the distal end (branch ends). Discretize the triangular mesh into a finite element mesh using a tetrahedral element generation algorithm to generate nodes and elements. Interpolate the displacement field within each element using linear shape functions to construct local stiffness and mass matrices. Merge the local stiffness and mass matrices of all elements to form a global linear system. Solve the linear system using a sparse direct solver (such as LU decomposition) to obtain the nodal displacement field. Calculate the displacement amplitude for each node in the nodal displacement field and quantify the total deformation. Statistically calculate the mean and standard deviation of displacement amplitudes within the healthy region. Define an abnormal threshold and compare the calculated displacement amplitude with the abnormal threshold. When the calculated displacement amplitude is greater than the abnormal threshold, it is marked as an abnormal lesion area.
[0122] A pre-trained random perturbation model is loaded and the Dropout mechanism is activated. During the testing phase, the random neuron dropping strategy during training is retained. By randomly masking part of the neuron weights, the posterior distribution of the Bayesian neural network is simulated, transforming the deterministic model into a probabilistic model. The posterior distribution of the parameters is approximated by Monte Carlo sampling to obtain the trained random perturbation model. Cross-modal fusion features are used as input, and through multiple independent forward propagations, 20% of the neuron weights are randomly dropped each time (such as randomly masking part of the weight matrix in a fully connected layer). The model output after each perturbation is the predicted malignancy probability, reflecting the prediction fluctuation caused by parameter randomness.
[0123] Based on abnormal lesion regions, Gaussian curvature, continuous homology features, and maximum standardized uptake values are extracted to screen out lesion regions that simultaneously satisfy morphological abnormalities, topological abnormalities, and metabolic activity abnormalities, forming a lesion set; based on the malignancy probability prediction results, malignancy and benignity are determined to generate diagnostic results; if multiple lesions exist, the lesion with the maximum standardized uptake value is used as the primary diagnostic basis.
[0124] This embodiment also provides a system for analyzing radiological image data, including:
[0125] The image registration module registers and extracts features from PET metabolic images and ultrasound images to generate physiological metabolic feature vectors.
[0126] The geometric correction module performs polar coordinate sinusoidal distortion enhancement and scan bed removal on CT images to generate geometrically corrected three-dimensional medical images.
[0127] The encoding module performs multi-scale Gaussian pyramid decomposition on the geometrically corrected 3D medical image, extracts cross-scale feature tensors through encoding state construction and entanglement operation, and outputs compressed composite feature tensors.
[0128] The topology analysis module reconstructs the organ surface manifold based on the real part of the compressed composite features, and combines Gaussian curvature detection and blood vessel continuous homology analysis to obtain the manifold feature vector.
[0129] The hypergraph fusion module constructs a dynamic hypergraph based on physiological metabolic feature vectors and manifold feature vectors, generating cross-modal fusion feature vectors;
[0130] The elasticity diagnostic module takes the cross-modal fused feature vector as input to the constraint generator, solves the elastic wave equation, and generates a diagnostic report.
[0131] This embodiment also provides a computer device suitable for use in radiology image data analysis methods, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the radiology image data analysis method proposed in the above embodiment.
[0132] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0133] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the radiology image data parsing method as described in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0134] In summary, this invention simulates the natural deformation of blood vessels affected by heartbeats through periodic sinusoidal perturbations and trilinear interpolation in cylindrical coordinates during CT image geometric correction, while precisely suppressing artifacts using a scanning bed mask. Multi-scale feature encoding solves the problem of high-frequency detail loss in traditional Gaussian pyramid decomposition by dynamically allocating weights based on gradient amplitude and combining entanglement operations in the space-frequency domain. In the biomechanical diagnostic stage, the elastic wave equation combined with a Monte Carlo random perturbation model quantifies deformation anomalies through finite element node displacement fields, and integrates morphological, metabolic, and topological multi-dimensional parameters, significantly improving the sensitivity of malignant lesion detection. Furthermore, dynamic hypergraph fusion of cross-modal features enables precise localization of structural-functional co-abnormalities, providing a highly reliable basis for clinical decision-making.
[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for analyzing radiological image data, characterized in that: include, Register and extract features from PET and ultrasound images to generate physiological metabolic feature vectors; Performing polar coordinate sinusoidal distortion enhancement and scan bed removal on CT images to generate geometrically corrected 3D medical images includes the following steps. The CT voxel data were converted to cylindrical coordinates and subjected to periodic sinusoidal perturbations. Trilinear interpolation is used to inversely map the perturbed cylindrical coordinates to the Cartesian coordinate system to generate a distortion-enhanced image; A three-dimensional binary mask is constructed based on the scanning bed position parameters. Bed artifacts are suppressed by pixel-by-pixel multiplication, and geometrically corrected three-dimensional medical images are output. Multi-scale Gaussian pyramid decomposition is performed on geometrically corrected 3D medical images. Cross-scale feature tensors are extracted through coding state construction and entanglement operation, and the compressed composite feature tensor is output. Based on the real part of the compressed composite features, the organ surface manifold is reconstructed, and the manifold feature vector is obtained by combining Gaussian curvature detection and blood vessel continuous cohomology analysis. Based on physiological metabolic feature vectors and manifold feature vectors, a dynamic hypergraph is constructed to generate cross-modal fusion feature vectors; Input the cross-modal fusion feature vector into the constraint generator, solve the elastic wave equation, and generate a diagnostic report.
2. The method for analyzing radiological image data as described in claim 1, characterized in that: The registration and feature extraction Includes the following steps, Based on the patient's position and the scanning bed position, initialize the translation and scaling components of the affine transformation matrix; The registration parameters were iteratively optimized using a quasi-Newton method to minimize the registration error between PET and ultrasound images; The registered PET images are mapped to the CT space by inverse transformation, and the metabolic activity values of each voxel are calculated using the SUV normalization formula. The tumor boundary was delineated in CT space and mapped to PET images. The maximum SUV value and the volume of voxels exceeding the threshold were extracted, and physiological metabolic feature vectors were generated by combining hemodynamic parameters.
3. The method for analyzing radiological image data as described in claim 1, characterized in that: The extraction of cross-scale feature tensors Includes the following steps, A three-level Gaussian pyramid decomposition was performed on the geometrically corrected 3D medical images to construct spatial domain coding states and frequency domain enhanced coding states. By fusing the spatial domain encoded state and the frequency domain enhanced encoded state through a parameterized entanglement gate, a composite state is generated, which is then compressed into a composite feature tensor using a complex orthogonal matrix projection.
4. The method for analyzing radiological image data as described in claim 1, characterized in that: The reconstructed organ surface manifold includes the following steps: Based on the complex real part of the encoded state of the compressed composite feature tensor, the isosurface grid is extracted, the Riemannian metric tensor and discrete Gaussian curvature detection are calculated, and the one-dimensional continuous homology features of the vascular complex are extracted by combining the α-shape algorithm. The curvature anomaly and topological invariant are fused to generate the manifold feature vector.
5. The method for analyzing radiological image data as described in claim 1, characterized in that: The construction of the dynamic hypergraph includes the following steps. Physiological metabolic feature vectors and manifold feature vectors are concatenated to form hypergraph nodes, and hyperedge correlation matrices are dynamically generated based on Pearson correlation coefficients. By aggregating cross-modal information through hypergraph convolution and backpropagating it, combined with degree matrix normalization, a fusion feature vector capturing the co-variation of structure and function is output.
6. The method for analyzing radiological image data as described in claim 1, characterized in that: The generation of the diagnostic report includes the following steps. A finite element model is constructed based on fused feature vectors to solve the nodal displacement field, and the deformation is quantified and compared with a health threshold to mark the diseased area. The probability of malignancy is predicted by Monte Carlo sampling of a random perturbation model, and a multimodal diagnostic report is generated by combining Gaussian curvature, continuous homology, and maximum SUV value.
7. A system for analyzing radiological image data, based on the method for analyzing radiological image data according to any one of claims 1 to 6, characterized in that: include, The image registration module registers and extracts features from PET metabolic images and ultrasound images to generate physiological metabolic feature vectors. The geometric correction module performs polar coordinate sinusoidal distortion enhancement and scan bed removal on CT images to generate geometrically corrected three-dimensional medical images. The encoding module performs multi-scale Gaussian pyramid decomposition on the geometrically corrected 3D medical image, extracts cross-scale feature tensors through encoding state construction and entanglement operation, and outputs compressed composite feature tensors. The topology analysis module reconstructs the organ surface manifold based on the real part of the compressed composite features, and combines Gaussian curvature detection and blood vessel continuous homology analysis to obtain the manifold feature vector. The hypergraph fusion module constructs a dynamic hypergraph based on physiological metabolic feature vectors and manifold feature vectors, generating cross-modal fusion feature vectors. The elasticity diagnostic module takes the cross-modal fused feature vector as input to the constraint generator, solves the elastic wave equation, and generates a diagnostic report.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the radiology image data parsing method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the radiology image data parsing method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Brain network cross-modal generation method based on diffusion model
CN118942648A
Medical decision-oriented multi-modal data dynamic fusion and labeling method and system
CN119377894A