A deep learning-based method for extracting lesion structures from cardiac magnetic resonance images

Abstract

This invention relates to the fields of computer vision and medical image processing technology, and discloses a method for extracting lesion structures from cardiac magnetic resonance imaging (MRI) images based on deep learning. The method first retrieves dynamic cardiac cine sequences and delayed gadolinium-enhanced sequences from the same anatomical location, extracts the centroid trajectory of the left ventricular blood pool, and performs cross-sequence axial alignment and non-rigid deformation correction. Subsequently, a local orthogonal coordinate system based on the myocardial anatomical topology is constructed, and pixel-level motion vector fields with physical constraints are calculated in the dynamic cardiac cine sequences. By projecting the motion vectors onto the local orthogonal coordinate system and performing vector clipping according to the centripetal contraction physical constraint, an effective motion vector field is generated. Finally, dynamic features and image grayscale features are fused, and a logic gating operator is used to extract the lesion structure mask.

Description

Technical Field

[0001] This invention relates to the field of computer vision and medical image processing technology, specifically to a method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning. Background Technology

[0002] Cardiac magnetic resonance imaging (MRI) is a crucial basis for clinical assessment of myocardial function and histopathological status. Dynamic cardiac cine sequences and delayed gadolinium-enhanced sequences are used to observe cardiac contractile dynamics and myocardial fibrosis, respectively. During actual examinations, differences in the patient's respiratory status and the timing of imaging often lead to significant shifts or non-rigid deformations in the anatomical and spatial positions of these two types of imaging sequences. Existing alignment techniques often struggle to achieve precise pixel-level matching of multimodal data, resulting in errors in subsequent fusion analysis.

[0003] Furthermore, current deep learning-based methods for extracting lesion structures primarily rely on the spatial grayscale distribution characteristics of images, lacking effective modeling of the physiological motion patterns of the heart. Because the heart undergoes complex concentric contractions, radial thickening, and circumferential rotations during the cardiac cycle, single data-driven models are prone to producing identification results that do not conform to biomechanical logic when processing image regions with low signal-to-noise ratios or interference from blood pool turbulence. Such algorithmic models, lacking physical prior constraints, cannot accurately distinguish between the overall physiological displacement of the ventricles and localized pathological motor dysfunction, thus affecting the boundary accuracy of lesion mask extraction and the consistency with clinical diagnosis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging (MRI) images. This method solves the problems arising from variations in the patient's respiratory state and limitations in the imaging principles of MRI equipment during cardiac MRI image processing. These limitations lead to generalized slice shifts and non-rigid deformations between dynamic cardiac cine sequences and delayed gadolinium enhancement sequences, making direct matching of multimodal data difficult. Furthermore, traditional data-driven extraction schemes lack constraints on the physiological motion logic of the heart, easily producing identification results that do not conform to biomechanical characteristics when dealing with blood pool turbulence artifacts and low signal-to-noise ratio regions, resulting in insufficient boundary accuracy and anatomical consistency in lesion structure extraction.

[0005] The first aspect of this invention provides a method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning, comprising the following steps: This method involves retrieving dynamic cardiac cine sequence and delayed gadolinium-enhanced sequence image data from the same anatomical location.

[0006] After acquiring the data, the centroid trajectory of the left ventricular blood pool in different sequences is extracted, and the slice overlap relationship is calculated. Cross-sequence axial alignment is performed to obtain aligned image data. In this process, the optimal slice offset is determined by calculating the left ventricular blood pool centroid coordinate sequence for each slice and using the minimum mean square error. Using the centroid coordinate sequence as a control point set, a transformation function is constructed using a thin plate spline interpolation algorithm to achieve pixel-level non-rigid deformation correction within the slice.

[0007] Subsequently, using the aligned delayed gadolinium-enhanced image sequence, a local orthogonal coordinate system based on the myocardial anatomical topology is constructed in reference space. A binary myocardial mask is generated by extracting the boundaries of the endocardium and endocardium, and the Laplace equation is solved based on this mask to generate the anatomical distance field potential energy distribution. The gradient of the potential energy distribution is calculated and normalized to obtain the geometric radial basis vectors of each pixel. Then, the geometric circumferential basis vectors are obtained through rotation transformation, providing an anatomically meaningful reference framework for myocardial motion analysis.

[0008] Simultaneously, using the aligned dynamic cardiac film sequence, pixel-level motion vector fields for the entire cardiac cycle are calculated. Pixel displacements are captured using a multi-scale depth optical flow model, and physical constraints are introduced into the loss function. The divergence constraint correction of the displacement vector field is performed using the prior knowledge of myocardial incompressibility to ensure the physical continuity of the motion field.

[0009] After obtaining the motion vector field, it is projected into the local orthogonal coordinate system, and vector clipping is performed according to the centripetal contraction physical constraint to generate an effective motion vector field. The motion vector is decomposed into radial thickening components and circumferential shear displacement components through dot product operation. A dynamic consistency measure is constructed based on the motion coordination criterion of healthy myocardium during contraction to identify the attenuation degree of the circumferential shear displacement component, and normal physiological displacement interference is eliminated using a preset deviation index threshold to identify candidate lesion regions with abnormal motion.

[0010] Finally, the dynamic characteristics of the effective motion vector field and the image grayscale features are fused to extract the lesion structure mask. The temporal energy integral within the lesion candidate region is calculated to generate a local motion disability score, measuring the consistency between the motion abnormality boundary and the myocardial anatomical edge in the normal direction, and constructing a structural correlation weight. Combining the image grayscale distribution characteristics, an adaptive logic gating function and high / low dual thresholds are used to execute hysteresis judgment logic, locking the damaged core area and restoring the boundary functional transition zone. Furthermore, the extracted lesion region is subjected to a logical AND operation using the aforementioned myocardial binary mask to confine the lesion region within the myocardial boundary, and a morphological opening operation is performed to output a three-dimensional lesion mask.

[0011] A second aspect of the present invention provides a deep learning-based system for extracting lesion structures from cardiac magnetic resonance imaging, comprising: The image acquisition module is used to acquire dynamic cardiac cine sequence and delayed gadolinium enhancement sequence image data of the same anatomical site. The axial alignment module is used to extract the centroid trajectory of the left ventricular blood pool in different sequences and calculate the layer overlap relationship, perform cross-sequence axial alignment and non-rigid deformation correction, and output aligned image data; The topology modeling module is used to construct a local orthogonal coordinate system for the myocardium based on the aligned image data, and to generate geometric radial basis vectors and geometric circumferential basis vectors. The motion extraction module is used to calculate the pixel-level motion vector field with physical constraints in the aligned dynamic cardiac movie sequence; The vector clipping module is used to project motion vectors onto a local orthogonal coordinate system and perform physical suppression processing according to the centripetal contraction physical constraint to output an effective motion vector field. The results extraction module is used to integrate the dynamic features of the effective motion vector field with the image grayscale features to identify and extract the lesion structure mask.

[0012] This invention provides a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging. It offers the following advantages: 1. This invention achieves precise spatial mapping between dynamic cardiac cine sequences and delayed gadolinium-enhanced sequences by extracting the centroid trajectory of the left ventricular blood pool and performing cross-sequence axial alignment and non-rigid deformation correction. This technique eliminates the multimodal data mismatch problem caused by the subject's respiratory motion or scanning plane offset, providing a unified anatomical reference basis for the subsequent extraction of lesion structures.

[0013] 2. This invention achieves effective separation of physiological displacement interference and pathological motion abnormalities by constructing a local orthogonal coordinate system based on the anatomical topology of the myocardium and performing vector clipping based on centripetal contraction constraints. By decomposing the motion vector into radial and circumferential components and performing physical amplitude suppression, motion artifacts caused by overall cardiac translation and rotation are eliminated, improving the system's specificity in identifying myocardial motion dysfunction regions.

[0014] 3. This invention utilizes a logic gating operator to integrate the temporal energy integral of the effective motion vector field with image grayscale features, and combines high and low dual thresholds to execute hysteresis judgment logic. This multi-dimensional feature fusion mechanism compensates for the lack of robustness of single image features in identifying transmural damaged areas, ensuring the accuracy of lesion core area localization while maintaining the connectivity of the output mask boundary and the integrity of the anatomical structure. Attached Figure Description

[0015] Figure 1 This is a structural diagram of the extraction system according to an embodiment of the present invention; Figure 2This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the construction of the myocardial geometric reference frame according to an embodiment of the present invention, where: a is the generated geometric radial basis vector. Distribution diagram, with arrows pointing from the endocardium to the epicardium, representing the normal direction of the myocardial wall; b is the generated geometric circumferential basis vector. Distribution diagram, with arrows arranged counterclockwise along the tangent of the myocardial wall; Figure 4 The exercise disability index of this invention Spatial distribution extraction map. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See attached document Figure 1 , Figure 1 This is a structural diagram of a deep learning-based cardiac magnetic resonance imaging lesion structure extraction system according to an embodiment of the present invention. The present invention provides a deep learning-based cardiac magnetic resonance imaging lesion structure extraction system, comprising: an image acquisition module, an axial alignment module, a topology modeling module, a motion extraction module, a vector clipping module, and a result extraction module.

[0018] Image acquisition module: Used to retrieve the subject's raw image data. The data includes dynamic cardiac cine sequences and delayed gadolinium-enhanced sequences of the same anatomical location. The image acquisition module standardizes the acquired sequence data and stores it in a cache for later retrieval.

[0019] Axial alignment module: Connects to the image acquisition module. Extracts the centroid trajectory of the left ventricular blood pool from the two sequences, determines the overlap relationship of the slices by calculating the minimum mean square error of the centroid sequence, performs axial alignment, and achieves pixel-level alignment using a non-rigid transformation operator.

[0020] Topology modeling module: Receives aligned image data. Constructs a local orthogonal coordinate system based on myocardial anatomy within the reference space of the delayed gadolinium enhancement sequence, and calculates the geometric radial and circumferential basis vectors of each pixel within the myocardium.

[0021] Motion extraction module: Used for pixel-level motion tracking of the entire cardiac cycle sequence in a dynamic cardiac movie sequence. It calls the depth optical flow operator to calculate displacement deviations and generate a motion vector field.

[0022] Vector Clipping Module: Connects the topology modeling module and the motion extraction module. It projects the motion vector field onto a local orthogonal coordinate system, performs amplitude suppression based on the physiological logic of the heart's centripetal contraction, removes noise, and outputs the effective motion vector field.

[0023] Result extraction module: Integrates grayscale features and dynamic features to generate a motion disability index map, uses logic gating operators to identify and extract lesion structures, and outputs the final three-dimensional lesion mask.

[0024] See attached document Figure 2 , Figure 2 This is a flowchart of a method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning, according to an embodiment of the present invention. The specific steps are as follows: S11, retrieve the subject's dynamic cardiac cine sequence and delayed gadolinium enhancement sequence image data; S12, extract the centroid trajectory of the left ventricular blood pool in different sequences and calculate the layer overlap relationship, and perform cross-sequence axial alignment; S13, Identify the boundaries of the myocardium and endocardium and construct a local orthogonal coordinate system based on anatomical topology; S14, using the depth optical flow algorithm to calculate the pixel-level motion vector field of the entire cardiac cycle in a dynamic cardiac movie sequence; S15, project the motion vector onto the local orthogonal coordinate system, perform vector physical clipping according to the centripetal contraction constraint, and generate an effective motion vector field; S16, calculate the time-domain energy integral of the effective motion vector field to generate a motion disability index map; S17. Establish a logic gating operator, integrate gray-scale distribution features, motor disability indicators and gradient collinearity features, and extract the cardiac lesion structure mask.

[0025] See attached document Figure 1 , Figure 1 This is a structural diagram of a deep learning-based cardiac magnetic resonance imaging lesion structure extraction system according to an embodiment of the present invention. In this embodiment, the image data acquisition and preprocessing step S11 in the deep learning-based cardiac magnetic resonance imaging lesion structure extraction method specifically includes the following execution process: S111, Acquire raw image data. The system retrieves the subject's cardiac MRI image sequences, including dynamic cardiac cine sequences with a time dimension (typically 20 to 50 frames) and static delayed gadolinium-enhanced sequences. The system ensures that both sets of sequences cover the same left ventricular target area on an anatomical level by identifying slice location information in the sequences, thereby eliminating spatial truncation errors.

[0026] S112, Spatial resolution normalization. The system identifies the voxel size information of the two sequences and uses a third-order spline interpolation algorithm to spatially resample the images, unifying the voxel specifications to a preset standard resolution. As a preferred approach, this embodiment sets the pixel spacing after resampling to between 1.0mm and 1.5mm to achieve a balance between clinical computational accuracy and processing efficiency.

[0027] S113, Image Intensity Normalization. To eliminate the influence of different scanning devices or parameters on image contrast, and to improve the stability of subsequent deep learning models for feature extraction, the system performs grayscale mapping on the resampled image. Considering that there may be a very small number of artifacts with extremely high grayscale values ​​in magnetic resonance images, this embodiment does not directly use the maximum value of the entire image. Instead, it calculates the cumulative distribution function of image pixel intensity, takes the 1st percentile as the low threshold, and takes the 99th percentile as the high threshold, thereby eliminating the interference of extreme noise points. The calculation logic of grayscale mapping is as follows: ; In the above formula, Represents pixels exist The original grayscale value at that moment; This represents the low grayscale threshold determined by the cumulative distribution function in the image sequence; This represents the high grayscale threshold determined based on the cumulative distribution function; The preset smoothing factor has a value range of 10. -6 Up to 10 -3 Used to prevent in and An anomaly occurs when the values ​​are equal, with the denominator being zero. For delayed gadolinium-enhanced sequences, the corresponding time variable in the formula... fixed. The normalized grayscale value is constrained to be around 0 to 1.

[0028] S114, Image Region Cropping and Localization. The system utilizes a lightweight convolutional neural network to automatically identify the center location of the left ventricular blood pool. Using this center point as a reference, the system performs a uniform window crop of 128×128 or 256×256 pixels to ensure that the model input fully includes the myocardium and cardiac chamber regions.

[0029] In this embodiment, the axial consistency calibration and spatial registration step S12 in a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging specifically includes the following execution process: S121, Left ventricular blood pool centroid sequence extraction. The system segments the left ventricular blood pool region in each slice of the dynamic cardiac cine sequence in the terminal diastolic frame and the delayed gadolinium enhancement sequence. For automatic segmentation of the blood pool region, those skilled in the art can use threshold-based segmentation methods or predefined convolutional neural network operators; the specific implementation methods are well-known in the art and will not be elaborated here. After obtaining the blood pool mask for each slice, the system calculates the geometric centroid coordinates of the blood pool region within that slice. Based on the physical principle of image centroid, for the first... Layer slice, its centroid coordinates The calculation logic is as follows: ; In the above formula, Indicates the first Axial position index of the layer slice; This indicates the total number of pixels within the blood pool mask of the slice; Indicates the first blood pool region The planar coordinates of each pixel; The preset infinitesimal constant has a value of 10. -8， Used when the segmentation algorithm fails to identify a valid blood pool region, i.e. When the value is 0, it prevents division by zero errors in the calculation formula. The system generates a dynamic cardiac cinematic sequence centroid trajectory vector representing the trajectory of the left ventricular anatomical centerline by traversing all slices. With delayed gadolinium-enhanced sequence centroid trajectory vector .

[0030] S122, Cross-Sequence Slice Migration Compensation. Considering the randomness of respiratory depth and lung filling status during different scan sequences, which leads to slice drift in the anatomical height of the heart within the body, the Mi system determines the correspondence between slices by matching the aforementioned centroid trajectory vector. The system searches for the optimal slice misalignment by calculating the minimum mean square error within a sliding window. The calculation logic is as follows: ; In the above formula, Indicates the centroid of a dynamic cardiac cinematic sequence; Indicates the centroid of the delayed gadolinium enhancement sequence; The offset variable for the traversed levels; The preset search space is usually set to a range of positive and negative 5 level index values, which is determined based on the maximum respiratory displacement deviation in clinical experience; This is the set of layers where the two sequences overlap at the current offset.

[0031] During the optimization process, if If the number of elements in the set is less than 3, the system determines that the geometric features of the offset position are insufficient to support robust alignment, and thus skips the calculation of the offset. By solving this optimization problem, the system establishes a logical index mapping between the two sequences along the axis, ensuring that the two sets of images have a definite correspondence in the layer dimension, thereby completing the coarse alignment of the slice level at the macroscopic level.

[0032] S123, In-plane non-rigid deformation correction. The system uses a thin-plate spline interpolation algorithm to construct the non-rigid transformation function. Regularization constraints are introduced into the equation solving process to improve numerical stability. This step, through nonlinear coordinate distortion, not only achieves overlap at the contour level but also ensures that each pixel within the myocardium acquires its corresponding kinematic properties under the anatomical reference of the delayed gadolinium enhancement sequence.

[0033] This embodiment uses a thin-plate spline interpolation algorithm to construct a non-rigid transformation function. This algorithm, based on the principle of minimizing bending energy, establishes the projection relationship from the source coordinate system to the target reference coordinate system using the extracted control point set. To prevent matrix singularity issues caused by overly concentrated control point distribution or redundant points when solving the interpolation weight matrix, the system introduces regularization constraints in the equations: ; In the above formula, The kernel matrix is ​​constructed based on the Euclidean distance between control points and is used to characterize the contribution weight of each control point to the spatial distortion. It is the identity matrix; This is the regularization smoothing parameter, and its value range is set to 10. -6 Up to 10 -4， Introducing this term can significantly increase the condition number of the matrix, ensuring numerical stability when solving large-scale linear equation systems. Let be the interpolation weight matrix to be solved; is the target coordinate vector of the corresponding control point in the delayed gadolinium enhancement sequence.

[0034] S124, Resampling Alignment Verification. The system utilizes the solved transformation function. Each frame of the dynamic cardiac cine sequence is resampled pixel by pixel. This process effectively solves the common problem of spatial inaccuracy in cardiac image analysis caused by misalignment of respiratory movements and cardiac cycles, thereby suppressing the risk of misdiagnosis of lesions caused by anatomical misalignment at the source.

[0035] In this embodiment, the myocardial geometric reference frame construction step S13 in a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging specifically includes the following execution process: S131, Myocardial Boundary Recognition and Mask Extraction. The system identifies the anatomical boundaries of the left ventricular myocardium on spatially registered delayed gadolinium-enhanced images. This process extracts the endocardial and epicardial contours using a segmentation algorithm, thereby defining the complete myocardial tissue region. In this embodiment, the segmentation algorithm preferably employs a U-Net symmetric network structure based on residual connections. Multi-scale features are extracted during the encoding stage, and skip connections are used during the decoding stage to recover boundary details lost due to downsampling. During training, the network uses a weighted combination of Dice similarity coefficient and cross-entropy as the loss function to address the sample imbalance problem caused by the extremely small proportion of myocardial tissue in the image. The system generates a binary mask of the myocardium based on the identified contours, where pixels belonging to the myocardial tissue are marked as 1, and the remaining background areas are marked as 0.

[0036] S132, Anatomical Distance Field Construction. Based on the anatomical feature of uneven cardiac wall thickness, to establish a depth coordinate system that can be normalized across samples, the system constructs Euclidean distance transformation maps for the endocardial and epicardial boundaries respectively. Let the pixel point... The minimum distance to the endocardium is The minimum distance to the epicardium is The system generates a potential field reflecting the relative depth of a pixel within the ventricular wall by solving the Laplace equation. This process is achieved by setting Dirichlet boundary conditions, where the potential energy at the endocardium is 0 and the potential energy at the epicardium is 1. In actual discrete computation, to improve computational efficiency, as a preferred approach, this potential field can be approximated using the following normalization formula: ; In the above formula, Represents pixels within a myocardial mask; This represents the Euclidean distance from the pixel to the inner membrane. This represents the Euclidean distance from the pixel to the outer film. This is a preset stability constant, and its value range is usually set to 10. -9 Up to 10 -7 The constant is introduced to account for the fact that when calculating boundary pixels, the denominator may approach zero due to floating-point rounding errors; this term ensures the robustness of numerical calculations. Potential field The value smoothly transitions from 0 at the endocardium to 1 at the epicardium, and the physical meaning of its gradient points to the anatomical radial direction of the myocardial wall.

[0037] S133, Geometric radial basis vector generation. The system is based on the potential field. The gradient distribution is used to determine the radial reference direction at each pixel. Considering that the centripetal contraction of the heart always physically occurs along the wall normal, the system calculates the spatial gradient of the potential field at the current pixel location and performs normalization to obtain the geometric radial basis vector. The calculation logic is as follows: ; In the above formula, Indicates the potential field at a point Spatial gradient operator at; The second norm of a vector; This is the regularization coefficient, with a value range of 10. -6 Up to 10 -4 The rationale for setting this coefficient is that in regions where the myocardial wall thickness is extremely small, gradient calculations may produce numerical singularities. Nonlinear suppression of gradient magnitude can be achieved, ensuring that the directional properties of the basis vectors dominate. This radial basis vector... It accurately describes the physiological normal direction of the pixel point perpendicular to the myocardial wall in terms of anatomical structure.

[0038] S134, Generation of Geometric Circumferential Basis Vectors. Based on the established radial basis vectors, the system constructs local tangent directions through orthogonal transformations. Based on the geometric prior of the left ventricle's approximately circular topology in the short-axis section, the contraction vectors of its myocardial fibers are mainly distributed in the tangent plane perpendicular to the radial direction. The system uses a rotation matrix to perform a 90-degree rotation on the radial basis vectors, thereby obtaining the geometric circumferential basis vectors. : ; In the above formula, This is a preset orthogonal rotation matrix; and These represent the horizontal and vertical components of the radial basis vectors in the Cartesian coordinate system of the image, respectively. Based on the left-handed coordinate system criterion used in this embodiment, this rotation matrix ensures that the circumferential basis vectors are arranged counterclockwise along the myocardial wall. Through this step, the system establishes a locally orthogonal reference frame that adaptively changes with the anatomical morphology at each pixel within the myocardium. .

[0039] S135, Reference frame consistency check. Considering the jagged noise that may exist at the image segmentation edges, these disturbances can cause non-physical oscillations in the gradient direction locally. The system performs spatial smoothing filtering on the generated basis vector field. The system uses a Gaussian kernel function to perform weighted convolution on the radial and circumferential fields respectively. In this embodiment, the standard deviation of the Gaussian kernel is... The standard deviation is set to 1.0 to 2.5 pixels. This range is chosen because: if the standard deviation is too small, vector abrupt changes caused by segmentation spurs cannot be eliminated; if the standard deviation is too large, geometric details at sharp bends in the myocardium (such as the attachment point of the papillary muscle) will be blurred. This step ensures the continuity and robustness of the spatial distribution of the local orthogonal reference frame.

[0040] In this embodiment, the spatiotemporal motion field extraction step S14 in a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging specifically includes the following execution process: S141, Input data serialization. To guide model focusing, the system utilizes the non-rigid transformation function obtained in step S123. The binary mask of myocardium generated in the reference space of the delayed gadolinium enhancement (LGE) sequence is resampled into the time coordinate system of the dynamic cardiac cinema sequence, and element-wise multiplication is performed on the input image to achieve spatial masking across sequences, preserving only the myocardium and its adjacent regions of interest.

[0041] S142, Construction of the Deep Optical Flow Estimation Model. As a preferred approach, the system employs a convolutional neural network with multi-scale feature extraction and local correlation measurement capabilities to capture pixel-level non-rigid motion vectors. Structurally, this network includes a feature encoder, a correlation bottleneck layer, and a motion field decoder. The feature encoder uses a Siamese network topology, utilizing convolutional kernels with shared weights. and High-dimensional semantic features are extracted in parallel to generate a 5-layer feature pyramid from coarse to fine.

[0042] Specifically, each coding unit contains two convolutional kernels with a size of [missing information]. The convolutional layers achieve spatial resolution downsampling through stride convolutions with a stride of 2. The correlation bottleneck layer constructs a four-dimensional cost body representing the probability distribution of inter-pixel displacements by performing dot product operations within local radius regions of the feature map. The motion field decoder, based on this cost body, achieves progressive resolution recovery through cascaded deconvolutional layers and cross-layer skip connections, and predicts the displacement residual of the current pixel at each scale. For the implementation of the underlying operators of this deep learning model, those skilled in the art can use a predefined residual block structure, the specific implementation of which is a well-known technique in the field and will not be elaborated here.

[0043] S143, Spatiotemporal motion vector field generation. The system uses the constructed model to perform inference operations on adjacent frames, outputting a description of the myocardial pixel points from... Time's up Vector field of displacement change at time intervals Based on the principles of non-rigid continuum mechanics, this vector field is composed of horizontal and vertical components, and its expression is as follows: ; In the above formula, Represents the pixel coordinates within the myocardial region; For time step index; It represents the displacement magnitude of a pixel along the horizontal axis of the image; This represents the displacement amplitude along the vertical axis. Considering the potential for missing local features in cardiac MRI images due to blood turbulence or magnetic field inhomogeneity, the system introduces a third stability operator in the denominator when normalizing or quantifying the displacement field. Its preferred value is 10. -8 This is to ensure the stability of numerical calculations when the displacement vector approaches zero.

[0044] S144, Model Training under Physical Constraints. To ensure that the extracted motion field conforms to the physical logic of the incompressibility of the cardiac muscle and physiological contraction, this embodiment employs a multi-criteria joint loss function during the model training phase. This loss function... The calculation logic is as follows: ; In the above formula, This is the photometric consistency term, used to measure the average absolute error of pixels between the distorted frame after predicted displacement and the target frame. Its purpose is to ensure the accuracy of motion vectors at the visual feature level by minimizing grayscale differences between images. This is a spatial smoothing term used to penalize abrupt changes in the first-order gradient of the displacement field. Its technical purpose is to constrain the rate of change of motion vectors between adjacent pixels, ensuring the continuity of myocardial motion in spatial distribution and preventing the generation of non-physical, isolated high-frequency noise. Its form of expression is For physical constraints, i.e., displacement field The squared divergence. According to biomechanical principles, myocardial tissue is approximately incompressible during the cardiac cycle, and its motion field should satisfy the constraint that the divergence is close to zero. By introducing this term, random expansion or contraction artifacts generated in low-contrast regions can be significantly suppressed.

[0045] and These are preset spatial smoothing weights and physical constraint weights, respectively. As a preferred option, The value is set to a range of 0.05 to 0.2. If the value is too low, the displacement field is easily affected by local noise and will produce spikes; if the value is too high, it will erase the fine motion gradient at the boundary of the lesion.

[0046] The value range is set to 0.1 to 0.3. The reason for setting this range is that the physical constraint strength of the myocardium needs to be slightly higher than the spatial smoothing strength to ensure that in areas with weak magnetic resonance imaging signals (such as the middle of the myocardium), the model can still output reasonable physiological displacement based on the prior knowledge of tissue incompressibility.

[0047] By strengthening physical constraints By adjusting the weights, the system can effectively correct random motion noise that does not conform to physiological laws. This training strategy, which combines data-driven (photometric term) and physical prior (divergence constraint term), ensures the anatomical rationality of the motion trajectory and provides a robust data foundation for the subsequent accurate identification of pathological movement disorders.

[0048] S145, Spatiotemporal Trajectory Integration. Based on the acquired instantaneous displacement field, the system performs temporal integration to reconstruct the complete motion trajectory of the myocardial pixel throughout the entire cardiac cycle. For any pixel at time k... The cumulative displacement relative to the initial end-diastolic frame is obtained by bilinear interpolation and accumulation of instantaneous vectors along the motion path. To compensate for global positional drift caused by minor body movements of the subject during scanning, the system uses the myocardial binary mask generated in step S13 to spatially mask the displacement field and calculates the spatiotemporal mean of all pixel displacement vectors belonging only to the myocardial mask region within the complete cycle. Subsequently, the system subtracts the mean component calculated for the myocardial region from the cumulative trajectory at each moment, thereby effectively eliminating the interference of background tissue movement outside the pericardium (such as the lungs, chest wall, etc.) on the calculation of myocardial motion trajectory, thus achieving accurate global motion compensation.

[0049] Based on the aforementioned spatiotemporal motion feature extraction logic combining deep learning and physical constraints, the system successfully transforms dynamic image sequences into a refined motion vector distribution. This process not only achieves pixel-level displacement accuracy but also, through the introduction of physical terms, solves the feature tracking failure problem that traditional algorithms are prone to in low-contrast myocardial regions. This highly preserved spatiotemporal motion field provides a data benchmark for the subsequent step S15, which uses the anatomical coordinate system for basis vector projection, ensuring that the system can identify and separate abnormal dynamic features related to cardiac lesions.

[0050] In this embodiment, the vector clipping process under physical constraints in a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging specifically includes the following execution process: S151, Decoupling of Motion Vector Field Spatial Resampling and Projection. Considering that the deep optical flow network may have performed downsampling operations, the system first uses a bilinear interpolation algorithm to resample the motion vector field to the same spatial resolution as the myocardial geometric mask. Subsequently, based on the physical decomposition logic of the non-rigid deformation of the heart, the system projects the motion vector onto the anatomical radial and circumferential basis vectors through dot product operations: ; ; In the above formula, Represents the pixel coordinates within the myocardial tissue; This serves as the phase index for the cardiac cycle; This represents the amplitude of radial motion after decoupling. A positive value physically represents the centripetal thickening of the myocardial wall during contraction, while a negative value represents the eccentric thinning during diastole. The amplitude of circumferential motion characterizes the shear displacement of myocardial fibers along the wall. This process realizes the transformation of abstract pixel displacement into an anatomical dynamic index with physiological significance.

[0051] because It is built on the registered anatomical reference, and the projection process realizes the transformation of abstract pixel displacement into physiological radial thickening / circumferential shortening components.

[0052] S152, Construction of a Dynamic Consistency Metric. Based on the physical synergy of radial thickening and circumferential shortening followed by healthy myocardium during systole, a dynamic consistency evaluation function was systematically constructed to quantitatively characterize the degree to which local motion deviates from physiological patterns. This function calculates the motion deviation index of each pixel by comprehensively considering the radial-circumferential motion ratio and the synchronicity of motion phase. ; In the above formula, The preset physiological coordination coefficient is set to a value ranging from 0.6 to 1.2. This coefficient is determined based on the statistical distribution ratio of radial and circumferential strain in a healthy left ventricle under systolic conditions. The preset stability value is 10. -7 When local ischemia or fibrosis of the myocardium leads to impaired contractility, radial displacement... It will decay significantly, making Approaching the maximum value of 1, the saliency labeling of the anomalous dynamic region is achieved.

[0053] S153, an adaptive vector clipping system based on physical constraints. The system utilizes a deviation index to perform nonlinear filtering on the overall motion vector, aiming to eliminate background motion that conforms to physiological expectations from the overall pulsation, thereby accurately locating candidate lesion regions. The clipped feature motion vectors... Perform the following logical judgment: ; In the above formula, The threshold value for global physical logic determination is set between 0.45 and 0.65. This pruning mechanism based on physical causality solves the technical difficulty of traditional difference methods in distinguishing between pathological abnormal movements and physiological displacements.

[0054] S154, Spatiotemporal integration and mask generation of lesion structures. The system performs spatiotemporal accumulation processing on the clipped residual vector to calculate the sum of abnormal motion energy of each pixel within a complete cardiac cycle. Based on this, the system uses a Markov random field model to spatially smooth the energy distribution map, optimizes the edge consistency of the lesion region by minimizing the energy functional, and generates a preliminary binary lesion structure mask accordingly.

[0055] In this embodiment, step S16 of the method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning specifically includes the following execution process: S161, Calculation of local motion disability score. The system receives the binary lesion structure mask generated in step S15 and the physically clipped feature motion vector field. Based on the physical causal relationship of the difference in kinetic energy work done by diseased tissue and healthy tissue within a unit cardiac cycle, the motion disuse index of each pixel in the spatial dimension is calculated. This index aims to quantitatively describe the degree of functional decline in damaged myocardium by comparing the kinetic energy of normal tissue with that of candidate lesion areas. Its calculation logic is as follows: ; In the above formula, Represents the pixel coordinates within the lesion mask area; The total number of sampling phases for a complete cardiac cycle; The original displacement vector within the healthy reference region; This represents the Euclidean norm. As a preferred method, the reference area Defined as the set of healthy myocardial voxels in the current image slice that are not identified as lesions and are more than 5 pixels away from the lesion boundary. This represents the total number of pixels within the set. The preset numerical stability operator has a value of 10. -8 Through this normalization process, the system converts absolute displacement energy into a ratio of relative functional loss.

[0056] S162, Correlation Analysis of Structural and Dynamic Gradients. The system uses the Sobel operator to calculate the energy field gradient vector. With image grayscale gradient vector And construct gradient direction aligned weights : ; In the above formula, This represents the dot product operation of vectors. The physical meaning of this weight lies in measuring the degree of consistency between the diffusion boundary of motion abnormalities and the anatomical edge of the myocardium in the normal direction. If... A value close to 1 indicates that the loss of motor function is strictly limited to the anatomical boundaries of the myocardial wall. This characteristic is a key criterion for distinguishing between pathological parenchymal lesions and random motion noise.

[0057] S163, Multi-dimensional Weighted Determination of Lesion Grade. Based on the aforementioned motor disability index and gradient alignment weights, the system performs a final risk classification of the lesion area using multi-criteria linear weighted logic. Total Score The expression is as follows: ; In the above formula, and These are preset feature weight coefficients. In this embodiment, the feature weight coefficients are... The value range is set to 0.6 to 0.8. The value range is set to 0.2 to 0.4, and it satisfies... The constraints. The rationale for choosing this allocation ratio is the actual amount of motor function decline. This is the core criterion for determining tissue dysfunction, while the gradient term serves as an auxiliary correction term for structural consistency, used to eliminate false positives from non-anatomical regions. The system sets three threshold levels to... Mapped to different pathological states: a score of 0.1 to 0.3 indicates mild functional impairment, 0.3 to 0.7 indicates moderate ischemic risk, and a score greater than 0.7 indicates severe disability or necrosis.

[0058] S164, Spatiotemporal Consistency Verification and Artifact Removal. To ensure the dynamic robustness of quantitative indicators in complex sequences, the system projects the extracted lesion mask back to the original time axis for closed-loop logic verification. The system calculates the stability of the disability score of the lesion region at end-systole and end-diastole. If the variance of the score of a certain lesion cluster exceeds a preset fluctuation threshold throughout the entire time phase, for example, 0.5, the region is determined to be affected by transient metal artifacts or magnetic field inhomogeneities, resulting in non-pathological fluctuations, and is reset to zero. This verification logic based on temporal consistency ensures that the quantitative results accurately reflect the continuous dynamic performance of myocardial tissue under physiological load.

[0059] In this embodiment, the multi-dimensional logic-gated lesion structure extraction step S17 in a deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging specifically includes the following execution process: S171, Multi-source feature tensor fusion. The system summarizes the time-domain average of the dynamic consistency deviation index calculated in step S15. The local kinetic incapacity index obtained in step S16 and gradient alignment weights Based on the complementary but numerically different nature of physical features in lesion representation, the system first performs a nonlinear mapping of the feature space to ensure comparability of each feature during the decision-making process. The time-domain average deviation index is calculated as follows: ; In the above formula, For myocardial pixel coordinates; This represents the total number of phases in the cardiac cycle. This represents the instantaneous deviation index. The system uses the min-max normalization algorithm to map the features of the above three dimensions to... The standard area, this preprocessing process aims to eliminate the decision bias caused by different physical dimensions, and provide a unified input tensor for subsequent logic gating.

[0060] S172, Construction of Adaptive Logic Gating Function. The system introduces a nonlinear logic gating operator to perform asymmetric nonlinear activation on the fused features. This gating function is designed as a continuously differentiable function with soft-margin characteristics to simulate the probabilistic capture of occult lesions, and its mapping logic is as follows: ; In the above formula, Represents pixels The overall confidence score of the lesion structure; It is the standard Sigmoid activation function; , and This is the preset gated gain coefficient. The physical reason for choosing this ratio is to prioritize the dominance of dynamic anomaly characteristics in the determination, while utilizing structural gradient information for spatial calibration. This is the gating bias term, with a value set between 0.55 and 0.75. It is used to control the activation sensitivity of the gating operator and prevent false detections triggered by low-energy noise.

[0061] S173, a multidimensional logic-gated decision based on hysteresis thresholds. The system is based on calculated confidence scores. Binarization gating is performed. To address the structural fragmentation problem that easily arises when extracting diffuse lesions using a single global threshold, this embodiment employs a dual-threshold hysteresis decision logic with spatial coherence constraints: ; In the above formula, The core threshold for high reliability ranges from 0.75 to 0.85. The low reliability edge threshold has a value range of 0.45 to 0.55. This represents the neighborhood connectivity determination of a pixel. The technical significance of this dual-threshold mechanism lies in using a high threshold to accurately locate the severely damaged myocardial core area, and using this as a seed point, using a low threshold to restore the slightly damaged functional transition zone at the boundary under connectivity constraints.

[0062] S174, Structural modification under anatomical topological constraints. Since myocardial tissue has a definite wall thickness constraint physiologically, the system utilizes the myocardial geometric mask obtained in step S13. The extracted preliminary lesion structure mask is subjected to anatomical consistency verification, and the final lesion structure mask is output. The system uses the myocardial geometric mask obtained in step S13. Preliminary mask for extraction Anatomical consistency verification is performed. The system executes a logical AND operation to ensure that the extracted lesion area is completely confined within the myocardial anatomical boundaries, i.e. For potentially non-physically isolated regions, the system further invokes the morphological opening operator for smoothing. The radius of the morphologically processed structural elements is preferably set to be less than 1 / 3 of the average myocardial wall thickness to prevent the loss of small lesions due to over-smoothing, thereby outputting the final accurate lesion structure mask.

[0063] In this example, a 58-year-old male subject was selected, clinically diagnosed with old myocardial infarction with ventricular aneurysm formation.

[0064] Input data: Dynamic cardiac cinematic sequence (CINE): 25 frames / cardiac cycle, layer thickness 8mm, pixel pitch 1.4mm.

[0065] Delayed gadolinium enhancement sequence (LGE): static 10-slice section used to observe fibrotic scars.

[0066] Observation of the technology implementation process: S12 phase: The system detected two levels of respiratory drift shift in the axial direction between the CINE and LGE sequences. Spatial compensation was achieved through centroid trajectory alignment.

[0067] S13 stage: In the left ventricular lateral wall region, the system constructs a local orthogonal basis vector field.

[0068] S15 stage: In the lateral wall segment of this patient, the original motion vector projection shows radial displacement. Approaching 0, while the healthy area mm. Deviation index It quickly climbed to 0.82 in that region, triggering the physical clipping mechanism.

[0069] S17 stage: Through dual threshold hysteresis determination, the system identifies the transmural infarction area located from the anterior wall to the lateral wall, and the final generated mask eliminates the artifact noise caused by the turbulence of the blood pool in the left ventricle.

[0070] To verify the effectiveness of the present invention, this embodiment constructs a test set containing 120 clinical samples (50 healthy samples and 70 samples of various myocardial lesions) and compares it with traditional algorithms (threshold-based segmentation algorithm, standard U-Net segmentation, and optical flow method without physical constraints).

[0071] 1. Definition of Evaluation Indicators The following indicators were used to evaluate the accuracy of lesion structure extraction: Dice similarity coefficient (DSC): measures the degree of overlap between the extracted mask and the clinical gold standard (manually annotated by experts).

[0072] Hausdorff distance (HD): Measures the maximum deviation between the extracted boundary and the true boundary (unit: mm).

[0073] Motion field consistency error (MCE): measures whether the extracted motion vectors conform to physiological incompressibility.

[0074] 2. Experimental data comparison table.

[0075] Experimental results show that, according to the above table and appendix Figure 3 and appendix Figure 4 It can be seen that the proposed solution significantly outperforms traditional methods in terms of Dice coefficient. This is especially true in regions with low signal-to-noise ratios, due to the introduction of physical constraint terms. By combining anatomical radial / circumferential projection, the system can effectively distinguish between "pseudo-displacement caused by the overall swing of the heart and dynamic incompetence caused by myocardial damage," thereby reducing the misdiagnosis rate from 15.2% to 3.1%.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning, characterized in that, Includes the following steps: Retrieve dynamic cardiac cine sequence and delayed gadolinium enhancement sequence image data of the same anatomical site; Extract the centroid trajectory of the left ventricular blood pool from different sequences and calculate the slice overlap relationship. Perform cross-sequence axial alignment to obtain aligned image data. In the reference space of the aligned delayed gadolinium enhancement sequence, a local orthogonal coordinate system based on the myocardial anatomical topology is constructed; Using the aligned dynamic cardiac film sequence, the pixel-level motion vector field of the entire cardiac cycle is calculated; The motion vector field is projected onto the local orthogonal coordinate system, and vector clipping is performed according to the centripetal contraction physical constraint to generate an effective motion vector field; By fusing the dynamic characteristics of the effective motion vector field with the image grayscale characteristics, a lesion structure mask is extracted.

2. The method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning according to claim 1, characterized in that, The process of performing cross-sequence axial alignment includes: Calculate the left ventricular blood pool centroid coordinate sequence for each layer and use the minimum mean square error to search for the optimal layer offset. Using the left ventricular blood pool centroid coordinate sequence as the control point set, a transformation function is constructed using a thin-plate spline interpolation algorithm to perform non-rigid deformation correction within the plane.

3. The method for extracting cardiac lesion structures based on deep learning from magnetic resonance imaging according to claim 1, characterized in that, The construction of a local orthogonal coordinate system based on the myocardial anatomical topology includes: Extract the boundaries between the endocardium and endocardium and generate a binary mask of the myocardium; Based on the aforementioned binary mask of the myocardium, the Laplace equation is solved to generate the potential energy distribution of the anatomical distance field. Calculate and normalize the gradient of the potential energy distribution of the anatomical distance field to obtain the geometric radial basis vector of each pixel; Perform a rotation transformation on the geometric radial basis vector to obtain the geometric circumferential basis vector of each pixel and construct a local orthogonal coordinate system.

4. The method for extracting cardiac lesion structures based on deep learning from cardiac magnetic resonance imaging according to claim 1, characterized in that, The calculation of the pixel-level motion vector field of the entire cardiac cycle includes: Construct a multi-scale depth optical flow model to capture pixel-level displacement; A physical constraint term is introduced into the loss function of the multi-scale depth optical flow model, and the displacement vector field is corrected by divergence constraint using the prior of myocardial incompressibility.

5. The method for extracting cardiac lesion structures based on deep learning from magnetic resonance imaging according to claim 1, characterized in that, The vector clipping based on the centripetal contraction physical constraint includes: The motion vector field is decomposed into a radial thickening component and a circumferential shear displacement component using dot product operations. A dynamic consistency measure is constructed based on the motion coordination criteria of healthy myocardium during systole to identify the degree of attenuation of the circumferential shear displacement component; By using a preset deviation threshold, normal physiological displacement is eliminated, and candidate lesion regions with abnormal movement are identified.

6. The method for extracting cardiac lesion structures based on deep learning from magnetic resonance imaging according to claim 5, characterized in that, The fusion of the dynamic characteristics of the effective motion vector field and the image grayscale characteristics includes: Calculate the time-domain energy integral of the effective motion vector field within the candidate lesion region to generate a local motion disability score; The consistency between the boundary of motion abnormalities and the anatomical edge of the myocardium in the normal direction is measured, and structural correlation weights are constructed. The confidence score of the lesion is calculated by combining the gray-scale distribution characteristics of the image.

7. The method for extracting cardiac lesion structures based on deep learning from magnetic resonance imaging according to claim 6, characterized in that, The extracted lesion structure mask includes: An adaptive logic gating function is constructed using the lesion confidence score; A high and low threshold dual-threshold execution lag judgment logic is adopted. The high threshold is used to lock the damaged core area, and this is used as a seed point to restore the boundary functional transition zone under connectivity constraints.

8. The method for extracting cardiac lesion structures based on deep learning from magnetic resonance imaging according to claim 7, characterized in that, After extracting the lesion structure mask, the method further includes: performing a logical AND operation on the extracted lesion region using the myocardial binary mask to confine the lesion region within the myocardial boundary; Perform a morphological opening operation on the result of the logical AND operation to smooth the edges and output a three-dimensional lesion mask.

9. The method for extracting lesion structures from cardiac magnetic resonance imaging based on deep learning according to claim 1, characterized in that, Image preprocessing is also included before performing cross-sequence axial alignment: The voxel specifications of image data are unified using a third-order spline interpolation algorithm; The grayscale distribution of the image is statistically analyzed, and the 1st and 99th percentiles are used as thresholds to perform normalization mapping through the cumulative distribution function.

10. A deep learning-based system for extracting lesion structures from cardiac magnetic resonance imaging, implemented by the deep learning-based method for extracting lesion structures from cardiac magnetic resonance imaging as described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire dynamic cardiac cine sequence and delayed gadolinium enhancement sequence image data of the same anatomical site. The axial alignment module is used to perform centroid trajectory alignment and non-rigid deformation correction between sequences, and outputs aligned image data; The topology modeling module is used to construct a local orthogonal coordinate system for the myocardium based on the aligned image data and generate geometric radial basis vectors and circumferential basis vectors; The motion extraction module is used to calculate the pixel-level motion vector field with physical constraints in the aligned dynamic cardiac movie sequence; The vector clipping module is used to project motion vectors onto a local orthogonal coordinate system and perform physical suppression, outputting an effective motion vector field. The results extraction module is used to fuse dynamic features and grayscale features to identify and extract lesion structure masks.

Images