A method and system for identifying magnetocardiographic abnormal regions based on image segmentation

Abstract

This invention discloses a method and system for identifying magnetic field abnormalities (MPA) regions based on image segmentation. The method generates a signal-to-noise ratio (SNR) weighted MPA signal dataset; forms an adaptive position-encoded feature set and fuses it with the SNR-weighted MPA signal dataset to generate a fused MPA feature tensor; obtains a FlowSDF symbolic distance field; obtains the boundary of the abnormal region with sub-pixel precision; generates a topologically accurate abnormal region mask; obtains multiple sets of FlowSDF symbolic distance fields and calculates their spatial variance; applies dynamic morphological filtering to the abnormal region mask based on the confidence evaluation results to obtain the final abnormal region mask; and generates a result for identifying small lesions in MPA regions. This invention performs reverse evolutionary refinement within a narrow-band refinement region, thereby maintaining the geometric continuity of the boundary at the scale of small lesions.

Description

Technical Field

[0001] This invention relates to the field of magnetic field technology, and in particular to a method and system for identifying abnormal regions of magnetic field based on image segmentation. Background Technology

[0002] Magnetocardiography (MCG) is a non-invasive biomedical engineering technique that uses the weak magnetic field generated by the heart's electrical activity for detection. Its physical basis stems from the Biot-Savart law; the magnetocardiogram signal is essentially a continuously distributed physical field, exhibiting a smooth and gradual change in magnetic flux density within the chest space. Compared to electrocardiography (ECG), MCG theoretically has higher spatial resolution potential and shows promise for applications in early myocardial ischemia, focal conduction abnormalities, and the localization of small electrophysiological lesions.

[0003] In existing technologies, the identification and segmentation of abnormal areas of the heart's magnetic field usually draws on semantic segmentation methods in the field of computer vision. The heart's magnetic field signal is usually converted into a regular grid image through interpolation, and then the image is classified pixel by pixel. The technical approach discretizes the continuous heart's magnetic field into a finite number of pixel units, resulting in a stepped or sawtooth shape in the boundary representation. When the signal amplitude of the abnormal area of ​​the heart's magnetic field is low or is affected by environmental noise, the pixel-level classification model is prone to misjudgment at the boundary of the abnormal area, resulting in boundary breaks, isolated noise points or multiple pseudo-connected regions, thereby destroying the topological integrity of the abnormal area. Especially at the scale of small lesions, the discretization error has a more significant impact on the geometric shape judgment.

[0004] Furthermore, existing magnetocardiogram (MCC) detection devices typically employ multi-channel superconducting quantum interference devices (SQUs) or atomic magnetometer arrays for measurement. Due to the high cost of sensors, the number of channels in clinical devices is usually limited, resulting in sparse sensor distribution within the chest space and significant physical spacing between adjacent sensors. Under hardware conditions, the number of sampling points for MCC signals in space is limited, leading to restricted spatial resolution. Existing techniques typically use linear interpolation or spline interpolation methods to spatially reconstruct sparse measurement points, converting them into high-resolution images for subsequent processing. However, pre-interpolation methods only perform numerical smoothing on the original discrete data and cannot recover the true magnetic field variation structure within the blind zones between sensors. When small lesions are located in the gaps between multiple sensors, missed detections or localization errors are likely, with localization errors potentially significantly exceeding the actual size of the lesion. Simultaneously, the lack of a quantification mechanism for the spatial uncertainty of segmentation results and an adaptive post-processing strategy based on confidence levels leads to insufficient reliability of abnormal region masks. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for identifying magnetic field abnormalities based on image segmentation. This invention performs reverse evolution refinement within a narrow-band refinement region, thereby maintaining the geometric continuity of the boundary at the scale of small lesions.

[0006] A method for identifying magnetic field abnormality regions based on image segmentation according to an embodiment of the present invention includes:

[0007] The raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device is collected and preprocessed to obtain the preprocessed magnetic field signal. The signal-to-noise ratio weighting coefficient is calculated for each sensor channel. The signal-to-noise ratio weighting coefficient is multiplied by the corresponding preprocessed magnetic field signal to generate a signal-to-noise ratio weighted magnetic field signal dataset.

[0008] The sensor spatial coordinate information is mapped to a high-dimensional feature space. The mapping result is multiplied by the signal-to-noise ratio weighting coefficient to form an adaptive position coding feature set. This set is then fused with the signal-to-noise ratio weighted magnetocardiogram signal dataset to generate a fused magnetocardiogram feature tensor.

[0009] An improved FlowSDF multi-resolution bidirectional velocity field prediction model is constructed based on the FlowSDF architecture. The fused magnetocardiogram feature tensor is used as a conditional input to obtain the FlowSDF symbolic distance field.

[0010] The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region, and the corresponding gradient and curvature information is calculated. Based on the gradient and curvature information, the position of the zero isosurface is updated iteratively with an adaptive step size to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel precision anomaly region.

[0011] Perform topological consistency constraints and connected component integrity checks on the boundaries of sub-pixel precision anomalous regions to generate a topologically accurate anomalous region mask.

[0012] In the inference stage, Monte Carlo random inactivation inference is implemented on the improved FlowSDF multi-resolution bidirectional velocity field prediction model to obtain multiple sets of FlowSDF symbol distance fields and calculate the spatial variance. Based on the spatial variance, confidence evaluation results are generated. Dynamic morphological filtering is performed on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask.

[0013] Based on the final abnormal region mask and the FlowSDF symbolic distance field, a set of quantitative indicators for the abnormal region is calculated to form the identification results of small lesions in the magnetic field abnormal region.

[0014] Optionally, the acquisition and preprocessing of the raw multi-channel magnetic resonance imaging (MRIE) signal data output by the multi-channel MRIE detection device includes:

[0015] Acquire raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device;

[0016] Time correction is performed on the original multi-channel magnetocardiogram (MCC) signal data to obtain the time-aligned original multi-channel MCC signal at the correction time.

[0017] Denoising filtering was performed on the time-aligned original multi-channel magnetocardiogram (MCC) signal to obtain the filtered MCC signal.

[0018] Baseline drift correction is performed on the filtered magnetocardiogram signal to obtain the baseline drift-corrected magnetocardiogram signal;

[0019] Time window segmentation is performed on the baseline drift-corrected magnetocardiogram signal to obtain time window signal segments;

[0020] Amplitude normalization is performed on the time window signal segments to obtain the preprocessed magnetocardiogram signal;

[0021] Calculate the signal-to-noise ratio weighting coefficient of the k-th sensor channel based on the short-time energy density;

[0022] The signal-to-noise ratio (SNR) weighting coefficient is multiplied by the corresponding preprocessed magnetocardiogram (MCC) signal to obtain the SNR-weighted MMC signal. The spatial coordinates of each sensor are associated with and aggregated with the corresponding SNR-weighted MMC signal to form an SNR-weighted MMC signal dataset.

[0023] Optionally, mapping the sensor spatial coordinate information to a high-dimensional feature space includes:

[0024] Obtain the spatial coordinates of each sensor that correspond one-to-one with the signal-to-noise ratio-weighted magnetocardiogram signal dataset;

[0025] Map the spatial coordinates of the k-th sensor to a high-dimensional feature space to obtain the high-dimensional coordinate feature vector of the k-th sensor;

[0026] Multiply the signal-to-noise ratio weighting coefficient by the corresponding high-dimensional coordinate feature vector to obtain the adaptive position coding feature vector of the k-th sensor;

[0027] The weighted magnetic field signal characteristic scalar is obtained by integrating the signal-to-noise ratio weighted magnetocardiogram signal of the k-th sensor channel in the m-th time window over the time interval corresponding to the time window and then dividing it by the length of the time window.

[0028] The adaptive position coding feature vector of the k-th sensor is fused with the weighted magnetocardiogram signal feature scalar of the sensor in the m-th time window to obtain the fused magnetocardiogram feature vector.

[0029] The fused magnetocardiogram feature vectors of all sensor channels across all time windows are correlated and aggregated to form a fused magnetocardiogram feature tensor.

[0030] Optionally, the construction of the improved FlowSDF multi-resolution bidirectional velocity field prediction model based on the FlowSDF architecture includes:

[0031] A sensor influence weighting field is constructed based on the total number of sensor channels, the signal-to-noise ratio weighting coefficient, and the sensor spatial coordinates.

[0032] Set a low-resolution grid scale and construct a low-resolution initial symbolic distance field on the low-resolution grid;

[0033] Based on the FlowSDF architecture, a low-resolution positive velocity field prediction function is constructed, and a physical divergence constraint is applied to the low-resolution positive velocity field prediction function. The physical divergence constraint is a weighted constraint on the divergence consistency of the velocity field through the sensor influence weight field.

[0034] In the low-resolution positive velocity field prediction function, the fused magnetic cardiometa-feature tensor is used as the conditional input to predict the low-resolution positive velocity field under different continuous time variables. By continuously evolving the low-resolution initial symbolic distance field along the low-resolution positive velocity field, the low-resolution initial symbolic distance field gradually converges into a low-resolution coarse-resolution symbolic distance field that conforms to the spatial distribution characteristics of small lesions in the magnetic cardiometa-abnormal region.

[0035] The results of the low-resolution coarse-resolution symbolic distance field at the continuous time variable t=1 are improved to a high-resolution grid through spatial interpolation to form a high-resolution initial symbolic distance field.

[0036] Based on the high-resolution initial symbol distance field, a narrow-band refinement region is constructed to define the boundary refinement range of tiny lesions in the magnetic anomaly region with a diameter smaller than the sensor spacing.

[0037] Based on the FlowSDF architecture, a high-resolution inverse velocity field prediction function is constructed in the narrowband refinement region, and physical divergence constraints are applied to the high-resolution inverse velocity field prediction function.

[0038] As the high-resolution inverse velocity field prediction function evolves from 1 to 0 in continuous time, the high-resolution initial symbolic distance field is numerically updated at each discrete time step. When the continuous time variable evolves to the termination time, a FlowSDF symbolic distance field with uniform resolution and continuous differentiability defined on the entire high-resolution grid is obtained.

[0039] Optionally, the step of extracting the zero isosurface from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region and calculating the corresponding gradient and curvature information includes:

[0040] The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region;

[0041] The spatial gradient information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the spatial gradient vector;

[0042] The curvature information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the curvature;

[0043] Within the neighborhood of the initial boundary of the magnetic anomaly region, a sub-pixel virtual sampling grid is constructed;

[0044] Based on curvature, the adaptive step size for sub-pixel space reconstruction is calculated;

[0045] Based on the adaptive step size, the boundary points of the initial boundary of the magnetic anomaly region are iteratively updated within the sub-pixel virtual sampling grid;

[0046] When the number of iterations reaches the set number of iterations or the absolute value of the sign distance at the boundary point position in the nth iteration is less than the preset convergence threshold, the iteration is terminated, and all updated boundary point positions are taken as the boundaries of the sub-pixel precision abnormal region.

[0047] Optionally, performing topological consistency constraints and connected component integrity checks on the boundaries of sub-pixel precision anomaly regions includes:

[0048] In the magnetocardiogram reconstruction space, the boundary of the sub-pixel precision anomalous region is mapped to a discrete grid used to generate the anomalous region mask, and a discrete boundary indicator function is obtained. Based on the discrete boundary indicator function, the closed region enclosed by the boundary of the sub-pixel precision anomalous region is filled to obtain the initial anomalous region mask.

[0049] Perform connected component decomposition based on the initial abnormal region mask to obtain a set of connected components, and mark each connected component as a connected component unit;

[0050] For each connected component, calculate the number of connected component voxels and the set of connected component boundary points;

[0051] Based on the number of connected voxels, a connected component integrity check is performed on the connected component set. Connected components with a number of connected voxels greater than or equal to the minimum connected component voxel threshold are retained as candidate lesion connected components. Topological consistency constraints are then applied to obtain the constrained candidate lesion connected components.

[0052] If the minimum boundary distance between any two candidate lesion connected units is less than or equal to the topological break merging threshold, then a connection path is constructed between the boundary point pairs corresponding to the minimum boundary distance, and the spatial position covered by the connection path is assigned a value of 1 in the initial abnormal region mask to generate the bridged abnormal region mask.

[0053] Perform a union update on the bridged abnormal region mask by applying the constrained candidate lesion connected components to obtain a topologically accurate abnormal region mask.

[0054] Optionally, the implementation of Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference phase includes:

[0055] In each Monte Carlo random inactivation inference, a random inactivation mask is applied to the network unit of the improved FlowSDF multi-resolution bidirectional velocity field prediction model to form a Monte Carlo random inactivation inference model, and the FlowSDF symbolic distance field is output by the Monte Carlo random inactivation inference model.

[0056] The mean symbol distance field of the FlowSDF symbol distance field is calculated at each spatial location in the magnetocardiogram reconstruction space, and the spatial variance field of the symbol distance is calculated through the mean symbol distance field.

[0057] Confidence assessment results are generated based on the symbolic distance space variance field.

[0058] Based on the confidence assessment results, dynamic morphological filtering is performed on the topology-fidelity outlier mask to obtain the outlier mask after dynamic morphological filtering, which is denoted as the final outlier mask.

[0059] Optionally, the calculation of the anomaly region quantization index set based on the final anomaly region mask and the FlowSDF symbolic distance field includes:

[0060] Calculate the spatial area of ​​the abnormal region based on the final abnormal region mask;

[0061] Calculate the geometric center of the abnormal region based on the final abnormal region mask;

[0062] Based on the FlowSDF symbolic distance field and the boundary of the sub-pixel precision anomaly region, the curvature value is calculated and a set of boundary curvature distributions is formed.

[0063] Based on the FlowSDF symbol distance field, the peak value of the magnetic field gradient inside the anomaly region is calculated;

[0064] The spatial area, geometric center location, boundary curvature distribution set, and magnetic field gradient peak of the abnormal region are combined to form a quantitative index set of the abnormal region, which is then output together with the final abnormal region mask to form the identification result of small lesions in the cardiac magnetic abnormal region.

[0065] Optionally, a method for identifying magnetic field abnormalities based on image segmentation includes:

[0066] The data acquisition and processing module acquires the raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device and performs preprocessing to obtain preprocessed magnetic field signals, and generates a signal-to-noise ratio weighted magnetic field signal dataset.

[0067] The feature fusion module fuses the adaptive position-coded feature set with the signal-to-noise ratio-weighted magnetocardiogram (MCG) signal dataset to generate a fused MCG feature tensor.

[0068] An improved FlowSDF module was developed, and an improved FlowSDF multi-resolution bidirectional velocity field prediction model was constructed based on the FlowSDF architecture to obtain the FlowSDF symbol range field.

[0069] The sub-pixel space reconstruction module calculates the corresponding gradient and curvature information, and uses an adaptive step size to iteratively update the zero isosurface position based on the gradient and curvature information to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel accuracy anomaly region.

[0070] The verification module performs topological consistency constraints and connectivity integrity checks on the boundaries of sub-pixel precision abnormal regions, and generates a topologically accurate abnormal region mask.

[0071] The filtering module performs Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference stage, obtains multiple sets of FlowSDF symbol distance fields and calculates the spatial variance, generates confidence evaluation results based on the spatial variance, and performs dynamic morphological filtering on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask.

[0072] The identification results module calculates a set of quantitative indicators for abnormal regions based on the final abnormal region mask and the FlowSDF symbolic distance field, forming the identification results of small lesions in the magnetic field abnormal region.

[0073] The beneficial effects of this invention are:

[0074] (1) Based on the improved FlowSDF continuous geometric field generation mechanism, this invention realizes the topological fidelity reconstruction of the boundary of the magnetic anomaly region, eliminates the boundary jaggedness and topological breakage caused by pixel discretization error, transforms the traditional pixel-level classification and segmentation paradigm into the generation paradigm of continuous symbolic distance field, and constructs a FlowSDF multi-resolution bidirectional velocity field prediction model that includes sensor influence weight field and physical divergence constraint, so that the geometric shape of the magnetic anomaly region gradually converges into a symbolic distance field that satisfies physical consistency during continuous time evolution. The symbolic distance field evolution is driven by the transport equation to ensure that the zero isosurface is naturally a continuous closed surface, and reverse evolution refinement is performed in the narrow band refinement region, thereby maintaining the geometric continuity of the boundary at the scale of small lesions.

[0075] (2) This invention achieves super-resolution localization capability for small lesions at the sub-sensor scale through the collaborative design of FlowSDF and sub-pixel spatial reconstruction interpolation algorithm. A symbol distance field with uniform resolution and continuous differentiability is generated by FlowSDF, and a virtual sampling grid of sub-pixels is constructed in the neighborhood of the zero isosurface. The boundary is iteratively updated based on the curvature adaptive step size, so that the boundary position converges to the sub-pixel accuracy position that satisfies the symbol distance of zero. The subsequent sub-pixel reconstruction is performed in the continuous geometric field, avoiding the smoothing artifacts introduced by the previous interpolation. For small lesions with a diameter smaller than the sensor spacing, high-frequency geometric information can be recovered in the narrow band refinement region.

[0076] (3) In the inference stage, the present invention performs multiple Monte Carlo random inactivation inferences on the improved FlowSDF multi-resolution bidirectional velocity field prediction model to obtain multiple sets of symbol distance fields and calculate the spatial variance field. The symbol distance spatial variance field is converted into a confidence evaluation result, and the radius of the morphological structural element is dynamically adjusted according to the confidence level. This enables the low confidence region to perform stronger smoothing and hole repair operations, while the high confidence region retains the original boundary details, thus realizing the adaptive structural optimization of the abnormal region mask. Attached Figure Description

[0077] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0078] Figure 1 This is a flowchart of a method for identifying magnetic anomaly regions based on image segmentation proposed in this invention;

[0079] Figure 2 This is a block diagram of the improved FlowSDF multi-resolution bidirectional velocity field prediction model in the image segmentation-based magnetic anomaly region identification method proposed in this invention. Detailed Implementation

[0080] Example 1: Reference Figures 1-2 A method for identifying magnetic field abnormality regions based on image segmentation, comprising:

[0081] The raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device is collected and preprocessed to obtain the preprocessed magnetic field signal. The signal-to-noise ratio weighting coefficient is calculated for each sensor channel. The signal-to-noise ratio weighting coefficient is multiplied by the corresponding preprocessed magnetic field signal to generate a signal-to-noise ratio weighted magnetic field signal dataset.

[0082] In this embodiment, the raw multi-channel magnetic resonance imaging (MRCI) signal data output by the multi-channel MRCI detection device is acquired and preprocessed, including:

[0083] Acquire raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device;

[0084] In Example 1, the total number of sensor channels of the magnetocardiogram detection device is set to K, and k represents the kth sensor channel. The original multi-channel magnetocardiogram signal of the kth sensor channel at the sampling time is acquired. The original multi-channel magnetocardiogram signal represents the measured value of the magnetocardiogram magnetic field strength at the sampling time.

[0085] Time correction is performed on the original multi-channel magnetocardiogram (MCC) signal data to obtain the time-aligned original multi-channel MCC signal at the correction time.

[0086] Denoising filtering was performed on the time-aligned original multi-channel magnetocardiogram (MCC) signal to obtain the filtered MCC signal.

[0087] Baseline drift correction is performed on the filtered magnetocardiogram signal to obtain the baseline drift-corrected magnetocardiogram signal;

[0088] Time window segmentation is performed on the baseline drift-corrected magnetocardiogram signal to obtain time window signal segments;

[0089] In Example 1, the time window index is set to m, and the value of m ranges from 1 to M. The baseline drift-corrected magnetocardiogram signal is divided into M time window signal segments according to the time window length value, so as to obtain the magnetocardiogram signal segment of the k-th sensor channel in the m-th time window.

[0090] Amplitude normalization is performed on the time window signal segments to obtain the preprocessed magnetocardiogram signal;

[0091] Calculate the signal-to-noise ratio weighting coefficient of the k-th sensor channel based on the short-time energy density;

[0092] In Example 1, the preprocessed magnetocardiogram signal of the k-th sensor channel in each time window is squared, integrated and averaged in the corresponding time window, and then averaged over all time windows to obtain the short-time energy density.

[0093] ;

[0094] in, This represents the short-time energy density of the k-th sensor channel over all time windows, used to measure the average energy level of the preprocessed magnetocardiogram signal of the k-th sensor channel per unit time. This represents the sensor channel index, where k ranges from 1 to K, and K represents the total number of sensor channels in the magnetocardiogram (MCC) detection device. Indicates the total number of time windows. Indicates the time window index. The value range is 1 to , used to identify the A time window, Indicates the length of the time window. Indicates the calibration time. This represents the set of time intervals corresponding to the m-th time window. This indicates that the k-th sensor channel is within the m-th time window, and the calibration time is... Preprocessing of magnetocardiogram signals at that time This indicates that the preprocessed magnetocardiogram signal of the k-th sensor channel within the m-th time window is within the calibration time. The instantaneous energy density component at that location.

[0095] The signal-to-noise ratio (SNR) weighting coefficient of the k-th sensor channel is obtained by dividing the short-time energy density value by the sum of the short-time energy density value and the noise reference energy density value. The SNR weighting coefficient is used to represent the reliability of the preprocessed magnetocardiogram signal of the k-th sensor channel, and the noise reference energy density value represents the energy measure of the noise level of the k-th sensor channel.

[0096] The noise reference energy density value is not arbitrarily set, but is calculated by collecting ambient background magnetic field data for a period of time in an unloaded state without a subject in the multi-channel magnetocardiogram detection device, and using the average short-time energy density as the system's baseline noise level. Furthermore, the sensor spatial coordinates are mapped to a high-dimensional feature space using a Fourier feature mapping strategy. This involves projecting the low-dimensional three-dimensional coordinates onto a high-dimensional manifold using a set of preset frequency sine and cosine functions, forming a high-dimensional coordinate feature vector. This improves the FlowSDF multi-resolution bidirectional velocity field prediction model's ability to capture high-frequency geometric details (sharp boundaries and subtle protrusions) in small lesions in magnetocardiogram abnormalities, overcoming the spectral bias problem of conventional linear coordinate input.

[0097] The signal-to-noise ratio (SNR) weighting coefficient is multiplied by the corresponding preprocessed magnetocardiogram (MCC) signal to obtain the SNR-weighted MMC signal. The spatial coordinates of each sensor are associated with and aggregated with the corresponding SNR-weighted MMC signal to form an SNR-weighted MMC signal dataset.

[0098] The sensor spatial coordinate information is mapped to a high-dimensional feature space. The mapping result is multiplied by the signal-to-noise ratio weighting coefficient to form an adaptive position coding feature set. This set is then fused with the signal-to-noise ratio weighted magnetocardiogram signal dataset to generate a fused magnetocardiogram feature tensor.

[0099] In this embodiment, mapping sensor spatial coordinate information to a high-dimensional feature space includes:

[0100] Obtain the spatial coordinates of each sensor that correspond one-to-one with the signal-to-noise ratio-weighted magnetocardiogram signal dataset;

[0101] Map the spatial coordinates of the k-th sensor to a high-dimensional feature space to obtain the high-dimensional coordinate feature vector of the k-th sensor;

[0102] In Example 1, the dimension of the high-dimensional feature space is set to D, and a coordinate mapping function is constructed. By inputting the spatial coordinates of the k-th sensor into the coordinate mapping function, the high-dimensional coordinate feature vector of the k-th sensor is obtained. The spatial coordinates of the k-th sensor represent the three-dimensional spatial position of the k-th sensor in the device coordinate system, and the high-dimensional coordinate feature vector represents the coordinate feature representation of the spatial coordinates of the k-th sensor in the high-dimensional feature space.

[0103] Multiply the signal-to-noise ratio weighting coefficient by the corresponding high-dimensional coordinate feature vector to obtain the adaptive position coding feature vector of the k-th sensor;

[0104] The weighted magnetic field signal characteristic scalar is obtained by integrating the signal-to-noise ratio weighted magnetocardiogram signal of the k-th sensor channel in the m-th time window over the time interval corresponding to the time window and then dividing it by the length of the time window.

[0105] The adaptive position coding feature vector of the k-th sensor is fused with the weighted magnetocardiogram signal feature scalar of the sensor in the m-th time window to obtain the fused magnetocardiogram feature vector.

[0106] In Example 1, the adaptive position coding feature vector is used as a vector of length D, and the weighted magnetocardiogram signal feature scalar is extended into a one-dimensional scalar feature component and appended to the end of the adaptive position coding feature vector to form a fused magnetocardiogram feature vector of length D plus 1. The fused magnetocardiogram feature vector represents the joint feature representation after combining the weighted magnetocardiogram signal feature scalar of the sensor in the m-th time window as a new feature dimension while keeping the spatial position feature dimension of the k-th sensor unchanged.

[0107] The fused magnetocardiogram feature vectors of all sensor channels across all time windows are correlated and aggregated to form a fused magnetocardiogram feature tensor.

[0108] An improved FlowSDF multi-resolution bidirectional velocity field prediction model is constructed based on the FlowSDF architecture. The fused magnetocardiogram feature tensor is used as a conditional input to obtain the FlowSDF symbolic distance field.

[0109] In this embodiment, an improved FlowSDF multi-resolution bidirectional velocity field prediction model is constructed based on the FlowSDF architecture, including:

[0110] A sensor influence weighting field is constructed based on the total number of sensor channels, the signal-to-noise ratio weighting coefficient, and the sensor spatial coordinates.

[0111] In Example 1, the sensor influence weight field is used to impose a higher constraint strength on the spatial region that is more reliable and closer to the sensor within the magnetocardiogram reconstruction space.

[0112] ;

[0113] Wherein, ρ(x) represents the value of the sensor influence weight field at spatial location x, used to measure the comprehensive constraint strength of this spatial location after being subjected to the weighted influence of each sensor in the magnetic resonance imaging space, and x represents the spatial coordinates in the magnetic resonance imaging space, used to carry the three-dimensional spatial position variable of the geometric morphology of small lesions in the magnetic resonance imaging abnormality area. This represents the signal-to-noise ratio weighting coefficient for the k-th sensor channel. This represents the three-dimensional spatial coordinates of the k-th sensor in the device coordinate system. The parameter represents the spatial diffusion scale of the sensor's influence on the weight field, used to control the attenuation range of the sensor's influence in the magnetic resonance imaging space. exp(·) represents the natural exponential function, and K is the total number of sensor channels.

[0114] Set a low-resolution grid scale and construct a low-resolution initial symbolic distance field on the low-resolution grid;

[0115] In Example 1, a low-resolution grid scale is set, and a Gaussian mixture probability distribution is constructed based on the short-time energy density of the sensor. A symbolic distance transformation is performed on the Gaussian mixture probability distribution to generate a prior low-resolution initial symbolic distance field that satisfies the constraint that the gradient magnitude is 1. The prior low-resolution initial symbolic distance field provides a geometric initialization state containing coarse spatial location information for manifold evolution, thereby accelerating evolution convergence.

[0116] Based on the FlowSDF architecture, a low-resolution positive velocity field prediction function is constructed, and a physical divergence constraint is applied to the low-resolution positive velocity field prediction function. The physical divergence constraint is a weighted constraint on the divergence consistency of the velocity field through the sensor influence weight field.

[0117] In Example 1, physical divergence constraints ensure that the divergence changes in the velocity field do not produce abnormal fluctuations near the boundary of the magnetic field abnormality region, thus guaranteeing the natural transition of the shape of the small lesions in the magnetic field abnormality region. Physical divergence constraints combine the weighted influence of the sensor with the evolution process of the velocity field to generate a low-resolution weighted divergence residual. By controlling the magnitude of the low-resolution weighted divergence residual, the evolution of the symbolic distance field is optimized to avoid errors or boundary shapes that do not conform to physical laws.

[0118] ;

[0119] in, This represents the weighted velocity field resulting from the scalar field and the vector field multiplied together. This represents the low-resolution weighted divergence residual. This represents the low-resolution positive velocity field prediction function.

[0120] In the low-resolution positive velocity field prediction function, the fused magnetic cardiometa-feature tensor is used as the conditional input to predict the low-resolution positive velocity field under different continuous time variables. By continuously evolving the low-resolution initial symbolic distance field along the low-resolution positive velocity field, the low-resolution initial symbolic distance field gradually converges into a low-resolution coarse-resolution symbolic distance field that conforms to the spatial distribution characteristics of small lesions in the magnetic cardiometa-abnormal region.

[0121] ;

[0122] in, Represents the gradient operator, This represents the low-resolution initial symbol range field.

[0123] The results of the low-resolution coarse-resolution symbolic distance field at the continuous time variable t=1 are improved to a high-resolution grid through spatial interpolation to form a high-resolution initial symbolic distance field.

[0124] In Example 1, the spatial interpolation method calculates the weighted distance value of the neighboring low-resolution grid nodes for each high-resolution grid node based on the corresponding spatial position relationship of the low-resolution grid nodes in the high-resolution grid, generates a high-resolution interpolated symbolic distance field, performs symbolic distance renormalization processing on the high-resolution interpolated symbolic distance field, so that the symbolic distance gradient magnitude at each high-resolution grid node is close to 1, restores the geometric consistency of the symbolic distance field, and obtains the high-resolution initial symbolic distance field.

[0125] Based on the high-resolution initial symbol distance field, a narrow-band refinement region is constructed to define the boundary refinement range of tiny lesions in the magnetic anomaly region with a diameter smaller than the sensor spacing.

[0126] In Example 1, the absolute value of the symbol distance at each spatial location is calculated based on the high-resolution initial symbol distance field. Spatial locations with a symbol distance absolute value less than a preset narrowband threshold are identified as narrowband candidate regions. The sensor spacing is the minimum Euclidean distance between any two adjacent sensor spatial coordinates. The preset narrowband threshold is set to a proportion less than the sensor spacing to ensure that the narrowband candidate regions cover the neighborhood of small lesions in the magnetic anomaly region with a diameter smaller than the sensor spacing. All narrowband candidate regions are spatially connected and isolated regions that do not meet the minimum area constraint are removed to obtain the narrowband refined region.

[0127] Based on the FlowSDF architecture, a high-resolution inverse velocity field prediction function is constructed in the narrowband refinement region, and physical divergence constraints are applied to the high-resolution inverse velocity field prediction function.

[0128] In Example 1, the high-resolution inverse velocity field prediction function generates an inverse velocity field based on the spatial feature information of the small lesions in the magnetic field abnormality region. The high-resolution initial symbolic distance field is refined through the inverse evolution equation of the symbolic distance field, and the boundary details of the high-resolution initial symbolic distance field are optimized. The refinement process is aimed at refining the boundary of the small lesions in the magnetic field abnormality region, ensuring that the identified boundary has not only geometric accuracy but also physical consistency.

[0129] As the high-resolution inverse velocity field prediction function evolves from 1 to 0 in continuous time, the high-resolution initial symbolic distance field is numerically updated at each discrete time step. When the continuous time variable evolves to the termination time, a FlowSDF symbolic distance field with uniform resolution and continuous differentiability defined on the entire high-resolution grid is obtained.

[0130] In Example 1, the process of the high-resolution reverse velocity field prediction function evolving from 1 to 0 in continuous time is defined as the inverse numerical integration process along the characteristic line of the transport equation. The manifold evolution update based on the high-order numerical integration operator is performed on the high-resolution initial symbolic distance field. The evolution direction of the inverse numerical integration process is along the negative gradient flow direction of the high-resolution reverse velocity field, which is used to minimize the geometric variational reconstruction error of the boundary of the magnetic anomaly region and recover the high-frequency geometric topological features.

[0131] At each time step of the inverse numerical integration process, the spatial position of the zero isosurface is monitored in real time. If the Euclidean distance between the zero isosurface and the boundary of the narrowband refinement region is detected to be less than a safety threshold (in Example 1, it is...) If the current zero isosurface is used as the center, the narrowband re-initialization mechanism is triggered, and the narrowband refinement region is reconstructed to prevent the boundary of small lesions in the magnetic anomaly region from overflowing the computational domain during the refinement process.

[0132] Simultaneously, a magnetic field forward projection verification operator based on the Biot-Savart law is constructed to convert the FlowSDF symbolic distance field output by S41 into a current density distribution model. The theoretical magnetic field intensity vector of the current density distribution model at each sensor coordinate is calculated in the forward direction, and the magnetic field residual norm between the theoretical magnetic field intensity vector and the weighted original multi-channel magnetocardiogram signal dataset is calculated.

[0133] To ensure that the generated geometric structure conforms to the laws of electromagnetic physics, this embodiment introduces an optimization step during the physical consistency test in the inference stage, rather than a remedial measure triggered only when the error exceeds the limit. Specifically, the FlowSDF symbolic distance field is converted into an equivalent current density distribution through differentiable operations. Using a physical forward projection operator based on the Biot-Savart law, the theoretical magnetic field strength generated by the equivalent current density at each sensor spatial coordinate is calculated. The magnetic field residual norm between the theoretical magnetic field strength and the signal-to-noise ratio weighted magnetic-cardiogram signal dataset is constructed as a physical loss function. The gradient descent method is used to fine-tune and optimize the FlowSDF symbolic distance field, forcing the output results to meet the constraints of Maxwell's equations, ensuring that the identification results of small lesions in the magnetic-cardiogram abnormal region have strict physical authenticity.

[0134] The basic architecture of the improved FlowSDF multi-resolution bidirectional velocity field prediction model in Example 1 mainly includes a feature input layer, a low-resolution forward evolution layer, and a high-resolution reverse refinement layer.

[0135] Feature Input Layer: Configured to receive fused magnetic-cardiogram feature tensors as conditional input variables for the entire dynamic system.

[0136] To effectively map the discretely distributed fused magnetocardiogram (MCC) feature tensor to the continuous MCC reconstruction space, the improved FlowSDF multi-resolution bidirectional velocity field prediction model introduces a feature query mechanism based on spatial attention. Specifically, for any continuous spatial coordinate point in a low-resolution grid or narrow-band refinement region, the improved FlowSDF multi-resolution bidirectional velocity field prediction model calculates the relative distance between the spatial coordinate point and all sensor spatial coordinates, and uses the relative distance as the query weight to perform weighted aggregation of the feature vectors in the fused MCC feature tensor. In this way, discrete sensor-level features are transformed into local continuous field feature vectors at the spatial coordinate points. The local feature vectors are then fed into the neural network units of the FlowSDF architecture to drive the numerical prediction of the low-resolution forward velocity field prediction function and the high-resolution reverse velocity field prediction function in the continuous space.

[0137] Low-resolution forward evolution layer: Defined on a low-resolution grid, it contains a low-resolution forward velocity field prediction function. The low-resolution forward evolution layer introduces physical divergence constraints through the influence of the sensor on the weight field, driving the low-resolution initial symbolic distance field to evolve forward over time.

[0138] High-resolution inverse refinement layer: Defined within a narrow refinement region of the high-resolution mesh, it contains a high-resolution inverse velocity field prediction function. The high-resolution inverse refinement layer updates the symbolic distance field after resolution enhancement by performing inverse numerical integration, and outputs a continuously differentiable FlowSDF symbolic distance field.

[0139] Network Units and Random Deactivation: The network units that constitute the velocity field prediction function are configured with a random deactivation mask interface, which supports Monte Carlo random deactivation inference during the inference phase to generate confidence assessment results.

[0140] This implementation overcomes the spatial resolution bottleneck of traditional magnetocardiography (MCG) limited by sparse sensor arrays by constructing a multi-resolution FlowSDF architecture under physical constraints. By solving the inverse problem of magnetocardiography in continuous manifold space, the generated magnetic field shape strictly follows Maxwell's equations, effectively eliminating non-physical artifacts generated by pure data-driven models. The introduced sensor influence weight field and narrowband refinement strategy enable the algorithm to accurately reconstruct the geometric topology of tiny lesions at the sub-sensor scale, achieving sub-pixel-level precise localization of tiny ischemic lesions with diameters smaller than the physical spacing between sensors. This improves the clinical diagnostic sensitivity and robustness of magnetocardiography in the early screening of tiny lesions.

[0141] The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region, and the corresponding gradient and curvature information is calculated. Based on the gradient and curvature information, the position of the zero isosurface is updated iteratively with an adaptive step size to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel precision anomaly region.

[0142] In this embodiment, a zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region, and the corresponding gradient and curvature information are calculated, including:

[0143] The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region;

[0144] In Example 1, the initial boundary of the magnetic field abnormality region is the set of spatial points x that make the sign distance equal to zero. The initial boundary of the magnetic field abnormality region is used to represent the initial geometric contour of the small lesions in the magnetic field abnormality region.

[0145] The spatial gradient information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the spatial gradient vector;

[0146] In Example 1, the spatial gradient vector is obtained by taking the partial derivatives of the symbolic distance field at spatial position x along the three spatial coordinate directions and combining them into a vector. The spatial gradient vector is used to describe the direction of normal change of the boundary of the magnetic anomaly region at that spatial position.

[0147] The curvature information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the curvature;

[0148] In Example 1, after normalizing the spatial gradient vector, spatial divergence operation is performed on the normalized gradient vector to obtain curvature. Curvature is used to describe the degree of local curvature of the boundary of the small lesion in the magnetic anomaly region at spatial location x.

[0149] Within the neighborhood of the initial boundary of the magnetic anomaly region, a sub-pixel virtual sampling grid is constructed;

[0150] In Example 1, a sub-pixel sampling interval is set, which is smaller than the scale of the high-resolution grid. This allows the sub-pixel virtual sampling grid to achieve denser spatial sampling within the high-resolution grid, enabling more precise spatial localization of the boundaries of small lesions in the magnetic field abnormality region.

[0151] Based on curvature, the adaptive step size for sub-pixel space reconstruction is calculated.

[0152] In Example 1, the absolute value of curvature at spatial location x is obtained. The absolute value of curvature is multiplied by the curvature modulation coefficient and added to obtain the step size scaling denominator. The reference step size parameter is divided by the step size scaling denominator to obtain the adaptive step size at spatial location x. The adaptive step size is used to automatically reduce the boundary update distance in the high curvature region of the boundary of small lesions in the magnetic anomaly region, and maintain a large boundary update distance in the low curvature region, so as to avoid oscillation error caused by crossing the zero isosurface during the sub-pixel spatial reconstruction process.

[0153] Based on the adaptive step size, the boundary points of the initial boundary of the magnetic anomaly region are iteratively updated within the sub-pixel virtual sampling grid;

[0154] In Example 1, the number of iterations is set to be... Let n denote the nth iteration, where n ranges from 0 to... The boundary point position of the nth iteration is defined as... The boundary point position in the nth iteration By the position of the boundary point in the nth iteration The signed distance value at a given point is divided by the sum of the magnitude of the spatial gradient vector and the gradient stability factor to obtain the normalized distance correction. This normalized distance correction is then multiplied by the adaptive step size, and the result is applied to the boundary point position in the nth iteration along the unit direction of the spatial gradient vector. Perform reverse correction to obtain the boundary point positions in the (n+1)th iteration. .

[0155] ;

[0156] in, Indicates the first The location of the boundary point in the next iteration. Indicates the first The location of the boundary point in the next iteration. Indicates the first The signed distance value at the boundary point location in the next iteration. Indicates the first The spatial gradient vector at the boundary point of the next iteration. Represents the gradient stabilization factor. is the adaptive step size, and is a dimensionless constant.

[0157] When the number of iterations reaches the set number of iterations or the absolute value of the sign distance at the boundary point position in the nth iteration is less than the preset convergence threshold, the iteration is terminated, and all updated boundary point positions are taken as the boundaries of the sub-pixel precision abnormal region.

[0158] The boundary of the subpixel precision anomaly region is a set of spatial points composed of the positions of the boundary points after iterative updates within the subpixel virtual sampling grid. The boundary of the subpixel precision anomaly region is used to represent the subpixel precision geometric contour of tiny lesions in the magnetic anomaly region.

[0159] Perform topological consistency constraints and connected component integrity checks on the boundaries of sub-pixel precision anomalous regions to generate a topologically accurate anomalous region mask.

[0160] In this embodiment, topological consistency constraints and connected component integrity checks are performed on the boundaries of sub-pixel precision abnormality regions, including:

[0161] In the magnetocardiogram reconstruction space, the boundary of the sub-pixel precision anomalous region is mapped to a discrete grid used to generate the anomalous region mask, and a discrete boundary indicator function is obtained. Based on the discrete boundary indicator function, the closed region enclosed by the boundary of the sub-pixel precision anomalous region is filled to obtain the initial anomalous region mask.

[0162] In Example 1, the discrete boundary indicator function is obtained by determining whether the spatial position in the discrete grid belongs to the set of spatial points formed by the boundary of the sub-pixel precision abnormal region, and is used to identify the spatial position belonging to the boundary of the sub-pixel precision abnormal region in the discrete grid.

[0163] If a point belongs to the set of spatial points formed by the boundaries of the sub-pixel precision anomaly region, the discrete boundary indicator function is set to 1; if it does not belong to the set of spatial points formed by the boundaries of the sub-pixel precision anomaly region, the discrete boundary indicator function is set to 0. Based on the discrete boundary indicator function, the closed region enclosed by the boundaries of the sub-pixel precision anomaly region is filled to obtain the initial anomaly region mask. The initial anomaly region mask is used to represent the internal region of the small lesions in the magnetic field anomaly region in the discrete grid. The initial anomaly region mask is obtained by filling the closed boundary enclosed by the discrete boundary indicator function. The spatial position inside the closed boundary is assigned a value of 1, and the spatial position outside the closed boundary is assigned a value of 0.

[0164] Perform connected component decomposition based on the initial abnormal region mask to obtain a set of connected components, and mark each connected component as a connected component unit;

[0165] The connected component set is used to represent the set of anomalous regions that are connected to each other in a discrete grid. Connected component decomposition is obtained by traversing and clustering the spatial locations with a value of 1 in the initial anomalous region mask according to their spatial adjacency.

[0166] For each connected component, calculate the number of connected component voxels and the set of connected component boundary points;

[0167] The number of connected voxels represents the number of spatial locations with a value of 1 in a connected unit. The set of connected boundary points represents the set of boundary spatial locations between the connected unit and its neighboring units with a value of 0. The number of connected voxels is obtained by counting the number of all spatial locations with a value of 1 in a connected unit.

[0168] For each spatial location with a value of 1 in a connected component, the adjacent spatial locations under the preset neighborhood structure are detected. If there is at least one adjacent spatial location with a value of 0 in the initial abnormal region mask, the spatial location is marked as a connected component boundary point. All spatial locations marked as connected component boundary points constitute the connected component boundary point set. The preset neighborhood structure is the spatial adjacency relationship defined in the discrete grid of magnetic resonance imaging reconstruction. The spatial adjacency relationship is used to determine whether two spatial locations are directly adjacent in space.

[0169] Based on the number of connected voxels, a connected component integrity check is performed on the connected component set. Connected components with a number of connected voxels greater than or equal to the minimum connected component voxel threshold are retained as candidate lesion connected components. Topological consistency constraints are then applied to obtain the constrained candidate lesion connected components.

[0170] In Example 1, the minimum connected voxel threshold is used to suppress pseudo-isolated abnormal regions caused by magnetic field noise or sparse sensor sampling errors. Connected units with a number of connected voxels less than the minimum connected voxel threshold are removed as a set of isolated noise connected regions.

[0171] By traversing all boundary point pairs between the boundary point set of the first candidate lesion connected domain unit and the boundary point set of the second candidate lesion connected domain unit, the Euclidean distance between each pair of boundary points is calculated and the minimum value is taken to obtain the minimum boundary distance. The minimum boundary distance is used to measure the topological break gap between the two candidate lesion connected domain units.

[0172] ;

[0173] in, Represents the L2 norm operation. and The spatial coordinates of the reconstructed magnetic field are given. Minimum boundary distance This is the first connected component of the candidate lesion. This is the second candidate lesion connected domain unit.

[0174] If the minimum boundary distance between any two candidate lesion connected units is less than or equal to the topological break merging threshold, then a connection path is constructed between the boundary point pairs corresponding to the minimum boundary distance, and the spatial position covered by the connection path is assigned a value of 1 in the initial abnormal region mask to generate the bridged abnormal region mask.

[0175] The topological break merging threshold is used to control the maximum permissible gap between two candidate lesion connected units that are spatially considered as the same small lesion in the same magnetic anomaly region.

[0176] Perform a union update on the bridged abnormal region mask by applying the constrained candidate lesion connected components to obtain a topologically accurate abnormal region mask.

[0177] In Example 1, all spatial locations are traversed in the bridged abnormal region mask. If a spatial location belongs to any constrained candidate lesion connected component, the value of the corresponding spatial location in the bridged abnormal region mask is kept as 1. If a spatial location does not belong to any constrained candidate lesion connected component, the value of the corresponding spatial location in the bridged abnormal region mask is set to 0.

[0178] The set of spatial locations in the bridged abnormal region mask that are all connected components of the candidate lesion under any constraint and have a value of 1 constitutes the topologically accurate abnormal region mask. The topologically accurate abnormal region mask is used to represent the final spatial region of the small lesions in the magnetic anomaly region after the connected component integrity check and topological consistency constraint.

[0179] In the inference stage, Monte Carlo random inactivation inference is implemented on the improved FlowSDF multi-resolution bidirectional velocity field prediction model to obtain multiple sets of FlowSDF symbol distance fields and calculate the spatial variance. Based on the spatial variance, confidence evaluation results are generated. Dynamic morphological filtering is performed on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask.

[0180] In this embodiment, Monte Carlo random inactivation inference is performed on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference phase, including:

[0181] In each Monte Carlo random inactivation inference, a random inactivation mask is applied to the network unit of the improved FlowSDF multi-resolution bidirectional velocity field prediction model to form a Monte Carlo random inactivation inference model, and the FlowSDF symbolic distance field is output by the Monte Carlo random inactivation inference model.

[0182] In Example 1, the number of Monte Carlo random inactivation inferences is set to R, where R represents the total number of times the Monte Carlo random inactivation inferences are repeatedly executed during the inference phase, and r represents the r-th Monte Carlo random inactivation inference. The value of r ranges from 1 to R. Each FlowSDF symbolic distance field is used to represent the symbolic distance from each spatial location in the magnetic resonance reconstruction space to the boundary of the magnetic resonance abnormality region.

[0183] The mean symbol distance field of the FlowSDF symbol distance field is calculated at each spatial location in the magnetocardiogram reconstruction space, and the spatial variance field of the symbol distance is calculated through the mean symbol distance field.

[0184] In Example 1, the corresponding statistical calculations of the R groups of FlowSDF symbol distance fields in the magnetic resonance imaging space are performed on each spatial location. For any spatial location in the magnetic resonance imaging space, the R symbol distances obtained from the first to the Rth random inactivation inferences are extracted. The R symbol distances are arithmetically averaged to obtain the mean symbol distance field at the spatial location. The mean symbol distance field represents the estimated average symbol distance from the spatial location to the boundary of the magnetic resonance imaging abnormal region under repeated random inactivation conditions.

[0185] Using the mean symbol distance field at the spatial location as a benchmark, the difference between each symbol distance obtained in R random inactivation inferences at the spatial location and the mean symbol distance field is calculated. All differences are squared and summed, and then divided by R minus one to obtain the symbol distance spatial variance field at the corresponding spatial location. The symbol distance spatial variance field is used to measure the degree of dispersion of the estimation results from the spatial location to the boundary of the magnetic anomaly region under random inactivation perturbation.

[0186] Confidence assessment results are generated based on the symbolic distance space variance field.

[0187] In Example 1, the confidence assessment result at the corresponding spatial location is obtained by dividing the symbolic distance spatial variance field by the square of the confidence scale parameter, taking the negative value, and using it as the exponent term of the natural exponential function.

[0188] The confidence assessment result is used to describe the reliability of the boundary of a small lesion in the magnetic field abnormality region at that spatial location. When the symbol distance spatial variance field is large, the confidence assessment result is small, and when the symbol distance spatial variance field is small, the confidence assessment result is large.

[0189] Based on the confidence assessment results, dynamic morphological filtering is performed on the topology-fidelity outlier mask to obtain the outlier mask after dynamic morphological filtering, which is denoted as the final outlier mask.

[0190] In Example 1, a minimum and a maximum structural element radius are set. The structural element radius at a spatial location is calculated based on the confidence assessment result. The structural element radius is equal to the linear interpolation result between the minimum and maximum structural element radii. When the confidence assessment result is low, the structural element radius approaches the maximum structural element radius; when the confidence assessment result is high, the structural element radius approaches the minimum structural element radius. Based on the structural element radius, morphological opening and closing operations are performed on the topology-fidelity anomalous region mask. The morphological opening operation is used to remove isolated spur structures caused by cardiomagnetic noise or model uncertainty, and the morphological closing operation is used to fill internal hole structures caused by cardiomagnetic noise or model uncertainty. This achieves adaptive smoothing and structural correction of the anomalous region mask.

[0191] The final abnormal region mask represents the result of the abnormal region determination at a spatial location. A value of 1 indicates that the spatial location belongs to the interior of the micro lesion in the magnetic anomaly region, and a value of 0 indicates that the spatial location belongs to the exterior of the micro lesion in the magnetic anomaly region.

[0192] Based on the final abnormal region mask and the FlowSDF symbolic distance field, a set of quantitative indicators for the abnormal region is calculated to form the identification results of small lesions in the magnetic field abnormal region.

[0193] In this embodiment, the set of quantization indicators for the abnormal region is calculated based on the final abnormal region mask and the FlowSDF symbolic distance field, including:

[0194] Calculate the spatial area of ​​the abnormal region based on the final abnormal region mask;

[0195] In Example 1, the number of spatial locations that satisfy the final abnormal region mask value of 1 is counted to obtain the number of abnormal region voxels. The volume of a single grid voxel of the discrete grid for magnetic resonance reconstruction space is set, and the equivalent area unit quantity corresponding to a single voxel is derived based on the voxel volume. By multiplying the number of abnormal region voxels by the equivalent area unit quantity corresponding to a single voxel, the spatial area of ​​the abnormal region is obtained. The spatial area is used to represent the two-dimensional projection area scale of the small lesions of the magnetic resonance abnormal region in the magnetic resonance reconstruction space.

[0196] Calculate the geometric center of the abnormal region based on the final abnormal region mask;

[0197] In Example 1, a set of all spatial locations that satisfy the final abnormal region mask value of 1 is obtained. The three-dimensional spatial coordinates of each spatial location in the set are summed, and the summation result is divided by the number of abnormal region voxels to obtain the geometric center position. The geometric center position represents the spatial positioning coordinates of the small lesions in the cardiac magnetic abnormal region in the cardiac magnetic reconstruction space, and is used to represent the overall spatial distribution center of the lesions.

[0198] Based on the FlowSDF symbolic distance field and the boundary of the sub-pixel precision anomaly region, the curvature value is calculated and a set of boundary curvature distributions is formed.

[0199] In Example 1, the positions of all boundary points in the boundary of the sub-pixel precision abnormal region are used as the target spatial position set for curvature calculation.

[0200] At each boundary point, the symbolic distance function value in the FlowSDF symbolic distance field is obtained, and partial derivatives of the FlowSDF symbolic distance field are calculated along the three spatial coordinate directions to obtain the spatial gradient vector at the spatial location. The spatial gradient vector is normalized, and the spatial divergence of the normalized gradient vector is calculated at the spatial location to obtain the curvature value at the spatial location. The curvature values ​​corresponding to the boundary points in the boundaries of all sub-pixel precision abnormal regions are recorded one by one according to the spatial location to form a boundary curvature distribution set. The boundary curvature distribution set is used to represent the local curvature variation of the boundary of the small lesion in the magnetic anomaly region at different spatial locations.

[0201] Based on the FlowSDF symbol distance field, the peak value of the magnetic field gradient inside the anomaly region is calculated;

[0202] In Example 1, in all spatial locations where the final abnormal region mask value is 1, the spatial gradient vector of the FlowSDF symbol distance field at each spatial location is calculated, the magnitude of the spatial gradient vector is calculated, and the maximum value of the magnitude of the spatial gradient vector is selected from all spatial locations within the abnormal region to obtain the magnetic field gradient peak. The magnetic field gradient peak is used to represent the location intensity feature where the symbol distance changes most drastically at the spatial location within the small lesion of the cardiac magnetic abnormal region.

[0203] The spatial area, geometric center location, boundary curvature distribution set, and magnetic field gradient peak of the abnormal region are combined to form a quantitative index set of the abnormal region, which is then output together with the final abnormal region mask to form the identification result of small lesions in the cardiac magnetic abnormal region.

[0204] In this embodiment, a method for identifying magnetic field abnormality regions based on image segmentation includes:

[0205] The data acquisition and processing module acquires the raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device and performs preprocessing to obtain preprocessed magnetic field signals, and generates a signal-to-noise ratio weighted magnetic field signal dataset.

[0206] The feature fusion module fuses the adaptive position-coded feature set with the signal-to-noise ratio-weighted magnetocardiogram (MCG) signal dataset to generate a fused MCG feature tensor.

[0207] An improved FlowSDF module was developed, and an improved FlowSDF multi-resolution bidirectional velocity field prediction model was constructed based on the FlowSDF architecture to obtain the FlowSDF symbol range field.

[0208] The sub-pixel space reconstruction module calculates the corresponding gradient and curvature information, and uses an adaptive step size to iteratively update the zero isosurface position based on the gradient and curvature information to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel accuracy anomaly region.

[0209] The verification module performs topological consistency constraints and connectivity integrity checks on the boundaries of sub-pixel precision abnormal regions, and generates a topologically accurate abnormal region mask.

[0210] The filtering module performs Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference stage, obtains multiple sets of FlowSDF symbol distance fields and calculates the spatial variance, generates confidence evaluation results based on the spatial variance, and performs dynamic morphological filtering on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask.

[0211] The identification results module calculates a set of quantitative indicators for abnormal regions based on the final abnormal region mask and the FlowSDF symbolic distance field, forming the identification results of small lesions in the magnetic field abnormal region.

[0212] Example 2:

[0213] During a continuous magnetic resonance imaging (MRI) acquisition cycle, the operator performed multi-channel MRI on a subject. The subject reported occasional chest tightness after exercise, but a routine electrocardiogram (ECG) did not capture stable abnormalities. The MRI detection device was a multi-channel array, and the total number of sensor channels was set to [number missing]. Each sensor outputs a time-series measurement of the magnetic field strength of the heart. The sampling frequency was set to 1000Hz, and the effective sampling duration was 240s, thus obtaining 240,000 sampling points per channel. The implementers observed in the original multi-channel heart signal data that if the 64-channel data was directly interpolated into a one-dimensional grid and then pixel segmented, some abnormal areas would show jagged boundaries, and there would be breaks and noise clumps at the gaps between the sensors, which could not reflect the abnormal morphology of the continuous physical field.

[0214] The system acquires raw multi-channel magnetocardiogram (MCC) signal data output from a multi-channel MCC detection device and simultaneously records the spatial coordinates and time synchronization information of each sensor. The system performs noise reduction filtering, baseline drift correction, time window segmentation, and amplitude normalization on each channel to obtain a preprocessed MCC signal. The time window length is set to [value missing]. The total number of time windows is ,by Indicates the first Within a given time window, for several representative channels, the system outputs the following segments after preprocessing:

[0215] No. Channel, No. Within the time window, preprocessed magnetocardiogram signals The peak-to-peak value is approximately The root mean square of the window is approximately .

[0216] No. Channel, No. Within the time window, the peak-to-peak value of the preprocessed magnetocardiogram signal is approximately The root mean square of the window is approximately However, residual fluctuations are more pronounced at low frequencies.

[0217] No. Channel, No. Within the time window, the preprocessed magnetocardiogram signal showed a slight but stable shift in the latter part of the window, with the peak-to-peak value being approximately [missing information]. .

[0218] The system then calculates the signal-to-noise ratio weighting coefficient for each sensor channel based on the short-time energy density. During the acquisition period, the system outputs the following statistical results for some channels:

[0219] The noise reference energy density value is ,get .

[0220] The noise reference energy density value is ,get .

[0221] The noise reference energy density value is ,get .

[0222] The system multiplies the signal-to-noise ratio (SNR) weighting coefficients by the corresponding preprocessed magnetocardiogram (MCC) signals to form an SNR-weighted MCC dataset. At this point, the system can provide an intuitive view of the weighted intensity of each channel within the same time window: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Within a time window The weighted peak-to-peak value is approximately ,and The weighted peak value is approximately Based on this, the system automatically reduces the influence of the 7th channel on the subsequent geometric field evolution.

[0223] The system acquires the signal-to-noise ratio weighted magnetic field signal dataset—corresponding to the spatial coordinates of each sensor—and then... Each sensor spatial coordinate is mapped to a high-dimensional feature space, and the dimension of the high-dimensional feature space is set to 1. , obtained the The system will use the high-dimensional coordinate feature vectors of each sensor to... Multiplying by the corresponding high-dimensional coordinate feature vector yields the adaptive positional encoding feature vector. The system then applies this to each time window. Calculate the weighted magnetocardiogram signal feature scalar and fuse it with the adaptive position-coded feature vector to obtain a length of The fused magnetic signature vector.

[0224] During the data collection period, the system... Examples of time windows outputting the scalar component at the end of the fused vector for several channels:

[0225] The scalar component is approximately ;

[0226] The scalar component is approximately ;

[0227] The scalar component is approximately .

[0228] The system correlates and aggregates the fused magnetocardiogram feature vectors of all sensor channels across all time windows to form a fused magnetocardiogram feature tensor, which is then used as the conditional input for subsequent FlowSDF.

[0229] The system constructs a sensor influence weighting field based on the total number of sensor channels, signal-to-noise ratio weighting coefficients, and sensor spatial coordinates. Within the acquisition period, the system sets... To verify the reflection of the blind zone caused by sensor sparsity in the weighted field, the system selected three spatial location points (all three-dimensional spatial location variables in the magnetocardiogram reconstruction space) for log output:

[0230] In spatial location (Approximately 9mm from the nearest sensor) .

[0231] In spatial location (Located in the gap between the geometric centers of the four sensors, approximately 16mm from the nearest sensor) .

[0232] In spatial location (Near the low-weight channel area, adjacent main channel) Mostly around 0.5. .

[0233] The system sets a low-resolution grid scale and constructs a low-resolution initial symbol range field on the low-resolution grid. The initial symbol range field is a uniformly randomly distributed initial range distribution. During the acquisition period, the system sets the low-resolution grid scale to [value missing]. Initialize on the grid By applying weighted residuals with physical divergence constraints, the system records several key values ​​of the forward evolution process to prove that it is not generated out of thin air: during the continuous-time variable progression (with the divergence number set as...), The system outputs at the same location point. Statistics:

[0234] exist place, The 95th percentile value dropped from 0.62 to 0.17;

[0235] exist place, The 95th percentile value dropped from 0.81 to 0.23;

[0236] exist place, The 95th percentile value dropped from 0.88 to 0.29.

[0237] The process corresponds to the constraint effect that prevents abnormal fluctuations in the velocity field divergence near the boundary. Based on this, the system outputs a low-resolution, coarse-resolution symbolic distance field. .

[0238] The system will A high-resolution initial symbolic distance field is formed by spatial interpolation to a high-resolution grid, and symbolic distance renormalization is performed to make the symbolic distance gradient magnitude close to 1. The high-resolution grid scale is set to [value missing] during the acquisition period. The system then constructs a narrowband refinement region, first taking... The candidate region is determined, and the preset narrowband threshold is set as a ratio of the sensor spacing (the minimum adjacent distance between sensors in this device is approximately 32mm, and the system uses a ratio of 0.08). After spatial connectivity filtering, the candidate regions are refined into narrowband regions. The system log shows that the number of candidate narrowband voxels is approximately... After removing small isolated regions, the number of voxels retained is approximately .

[0239] The system constructs a high-resolution inverse velocity field prediction function within a narrow-band refinement region and applies physical divergence constraints. Through refinement using inverse evolution equations, as the continuous-time variable evolves from 1 to 0, the system performs numerical integration updates at each discrete-time step and outputs a uniformly resolved and continuously differentiable FlowSDF symbolic distance field at the final time. At this point, the system observed that near the suspected abnormal region, the zero isosurface of the FlowSDF symbol distance field was in a continuous closed shape, while using the traditional method of interpolation followed by pixel classification, the boundary corresponding to the zero isosurface showed multiple breaks and jagged corners in the same region.

[0240] The system extracts the zero isosurface from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region, calculates the spatial gradient vector and curvature on the boundary, and then constructs a sub-pixel virtual sampling grid in the neighborhood. During the acquisition period, the system sets the sub-pixel sampling interval to 0.4 mm and automatically reduces the adaptive step size in regions with high curvature. The system sets the number of iterations. The system extracted 5 boundary points on the same boundary segment as examples for sub-pixel updates and recorded them. Convergence process (unit: mm):

[0241] Point P1: Initial 6th 12th 18th time 0.02

[0242] Point P2: Initial 6th 12th 18th time 0.02

[0243] Point P3: Initial 6th 12th 18th time 0.03

[0244] Point P4: Initial 6th 12th 18th time 0.01

[0245] Point P5: Initial 6th 12th 18th time 0.02

[0246] When the number of iterations reaches the set number or The iteration terminates when the pixel size is less than the threshold (0.03 mm in this embodiment 2), obtaining the boundary of the sub-pixel precision anomaly region. The system maps the boundary of the sub-pixel precision anomaly region to a discrete grid, obtains the discrete boundary indicator function, and performs internal filling on the closed region to obtain the initial anomaly region mask. The system then performs connected component decomposition and connected component voxel count, and calculates the boundary point set for each connected component. The system obtains the total number of connected components within the acquisition period. Two connected component voxels were removed because their number of voxels was below the minimum connected component voxel threshold. The remaining four entered the candidate lesion connected component. The system further calculated the minimum boundary distance between the candidate lesion connected components and output two pairs of the most critical distances:

[0247] ; .

[0248] The system sets the topology fracture merging threshold to be [value]. Therefore, for and By constructing connection paths and bridging merging, the small abnormal regions that were originally cut into two pieces are reformed into a single connected structure. After bridging, the number of isolated noise points in the abnormal region mask decreases from 43 to 6, and the connected domain structure converges from 4 candidates to 2 candidates, one of which is the main lesion region.

[0249] The system implements Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model. The system sets the number of Monte Carlo random inactivation inferences to be [number missing]. Each inference outputs a set of FlowSDF symbolic distance fields, and successively calculates the mean symbolic distance field and the spatial variance field of the symbolic distance for the same spatial location, further generating confidence assessment results. The system extracts three spatial location points near the main lesion boundary and outputs the variance and confidence level.

[0250] Point inside the boundary ;

[0251] Points near the boundary ;

[0252] Far from the lesion point .

[0253] The system performs dynamic morphological filtering on the topology-fidelity outlier region mask based on confidence level, and sets a minimum structuring element radius. 0.6mm, maximum structural element radius The structuring element radius is obtained by linear interpolation based on confidence level. Therefore, in nearby Approaching 2.0mm or larger, used for strong burr suppression; in nearby Approximately 0.8mm is used to preserve true boundary details. After filtering, the system outputs the final abnormal area mask. The log shows that the boundary smoothness of the main lesion area is improved without losing small protrusion structures, and the pseudo islands in the non-lesion area are completely removed.

[0254] The system calculates and outputs a set of quantitative indicators for abnormal regions based on the final abnormal region mask and the FlowSDF symbolic distance field. The quantitative indicators output by the system are in a format similar to that of a clinical report.

[0255] Spatial area index: Calculated by counting the number of spatial locations with a value of 1 in the final abnormal region mask and combining this with voxel-scale conversion. The spatial area index of the main lesion area is: The spatial area index of the secondary lesion area is .

[0256] Geometric center location: The geometric center location is obtained by averaging the voxel coordinates within the main lesion area. The system outputs the three-dimensional coordinates of the geometric center position as follows: The geometric center of the secondary lesion area is located at... .

[0257] Boundary curvature distribution: Curvature is calculated point-by-point on the zero isosurface to form a curvature distribution set. The system outputs the curvature distribution of the main lesion boundaries using quantile statistics, with the 25th quantile being... The 50th percentile of curvature is The 75th percentile of curvature is The 95th percentile of curvature is It also records that the high curvature segments are concentrated in two bending areas at the boundary.

[0258] Peak magnetic field gradient: calculated within the region where the final anomaly mask value is 1. The maximum value, the system output of the peak index of the magnetic field gradient of the main lesion is Secondary lesions are This was incorporated into the quantitative index set as an indicator of the intensity of boundary changes.

[0259] To ensure the comparison was convincing, the implementers used three procedures for the same set of raw multichannel magnetocardiogram (MCC) signal data:

[0260] Traditional Method A: Linearly interpolate the 64-channel data into a two-dimensional grid image, and then use U-Net for pixel-level segmentation;

[0261] Traditional Method B: Perform cubic spline interpolation on the 64-channel data to create a two-dimensional grid image, and then use AttentionU-Net for pixel-level segmentation;

[0262] The method of this invention is as follows: FlowSDF symbolic distance field is generated according to the process of claims 1-8, and sub-pixel boundary reconstruction and uncertainty-driven dynamic morphological filtering are performed.

[0263] During the training phase, the implementer extracted time window samples from the current collection period for fine-tuning (only a small number of segments are shown to illustrate the data format). The system randomly selected three time windows from the training sample set (all different time windows of the same subject) and output their weighted feature scalar segments:

[0264] Time window sample #1 Corresponding channel 23, 7, 41, 12, 55);

[0265] Time window sample #2 ;

[0266] Time window samples .

[0267] The final comparison results (based on the main lesion area) for the same collection cycle are as follows:

[0268] Boundary serration (expressed as the density of boundary broken line corners, unit: number of corners / 100mm): Traditional A is 18.4, Traditional B is 14.1, and the present invention is 6.7;

[0269] Number of topological breaks (the number of times the same lesion is divided into multiple disconnected blocks, unit: times): Traditional A is 2, Traditional B is 1, and the present invention is 0;

[0270] Center positioning offset within the sensor gap blind zone (statistical based on the offset from the blind zone landing point to the nearest boundary point, unit: mm): Traditional A is 13.8, Traditional B is 9.6, and this invention is 4.1;

[0271] The number of pseudo-islands in the final mask (area less than) The number of isolated connected components (unit: number): Traditional A is 39, Traditional B is 21, and the present invention is 6;

[0272] Curvature distribution stability (ratio of 95th percentile to 50th percentile of curvature, the larger the value, the more obvious the boundary spikes): Traditional A is 3.12, Traditional B is 2.46, and the present invention is 2.33;

[0273] The number of residual holes (number of holes inside the mask, unit: number) after confidence-guided dynamic morphological filtering is 11 for traditional A, 7 for traditional B, and 2 for this invention.

[0274] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying magnetic field abnormality regions based on image segmentation, characterized in that, include: The raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device is collected and preprocessed to obtain the preprocessed magnetic field signal. The signal-to-noise ratio weighting coefficient is calculated for each sensor channel. The signal-to-noise ratio weighting coefficient is multiplied by the corresponding preprocessed magnetic field signal to generate a signal-to-noise ratio weighted magnetic field signal dataset. The sensor spatial coordinate information is mapped to a high-dimensional feature space. The mapping result is multiplied by the signal-to-noise ratio weighting coefficient to form an adaptive position coding feature set. This set is then fused with the signal-to-noise ratio weighted magnetocardiogram signal dataset to generate a fused magnetocardiogram feature tensor. An improved FlowSDF multi-resolution bidirectional velocity field prediction model is constructed based on the FlowSDF architecture. The fused magnetocardiogram feature tensor is used as a conditional input to obtain the FlowSDF symbolic distance field. The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region, and the corresponding gradient and curvature information is calculated. Based on the gradient and curvature information, the position of the zero isosurface is updated iteratively with an adaptive step size to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel precision anomaly region. Perform topological consistency constraints and connected component integrity checks on the boundaries of sub-pixel precision anomalous regions to generate a topologically accurate anomalous region mask. In the inference stage, Monte Carlo random inactivation inference is implemented on the improved FlowSDF multi-resolution bidirectional velocity field prediction model to obtain multiple sets of FlowSDF symbol distance fields and calculate the spatial variance. Based on the spatial variance, confidence evaluation results are generated. Dynamic morphological filtering is performed on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask. Based on the final abnormal region mask and the FlowSDF symbolic distance field, a set of quantitative indicators for the abnormal region is calculated to form the identification results of small lesions in the magnetic field abnormal region.

2. The method for identifying magnetic field anomaly regions based on image segmentation according to claim 1, characterized in that, The process of acquiring and preprocessing the raw multi-channel magnetic resonance imaging (MRCI) signal data output by the multi-channel MRCI detection device includes: Acquire raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device; Time correction is performed on the original multi-channel magnetocardiogram (MCC) signal data to obtain the time-aligned original multi-channel MCC signal at the correction time. Denoising filtering was performed on the time-aligned original multi-channel magnetocardiogram (MCC) signal to obtain the filtered MCC signal. Baseline drift correction is performed on the filtered magnetocardiogram signal to obtain the baseline drift-corrected magnetocardiogram signal; Time window segmentation is performed on the baseline drift-corrected magnetocardiogram signal to obtain time window signal segments; Amplitude normalization is performed on the time window signal segments to obtain the preprocessed magnetocardiogram signal; Calculate the signal-to-noise ratio weighting coefficient of the k-th sensor channel based on the short-time energy density; The signal-to-noise ratio (SNR) weighting coefficient is multiplied by the corresponding preprocessed magnetocardiogram (MCC) signal to obtain the SNR-weighted MMC signal. The spatial coordinates of each sensor are associated with and aggregated with the corresponding SNR-weighted MMC signal to form an SNR-weighted MMC signal dataset.

3. The method for identifying magnetic field anomaly regions based on image segmentation according to claim 1, characterized in that, The process of mapping sensor spatial coordinate information to a high-dimensional feature space includes: Obtain the spatial coordinates of each sensor that correspond one-to-one with the signal-to-noise ratio-weighted magnetocardiogram signal dataset; Map the spatial coordinates of the k-th sensor to a high-dimensional feature space to obtain the high-dimensional coordinate feature vector of the k-th sensor; Multiply the signal-to-noise ratio weighting coefficient by the corresponding high-dimensional coordinate feature vector to obtain the adaptive position coding feature vector of the k-th sensor; The weighted magnetic field signal characteristic scalar is obtained by integrating the signal-to-noise ratio weighted magnetocardiogram signal of the k-th sensor channel in the m-th time window over the time interval corresponding to the time window and then dividing it by the length of the time window. The adaptive position coding feature vector of the k-th sensor is fused with the weighted magnetocardiogram signal feature scalar of the sensor in the m-th time window to obtain the fused magnetocardiogram feature vector. The fused magnetocardiogram feature vectors of all sensor channels across all time windows are correlated and aggregated to form a fused magnetocardiogram feature tensor.

4. The method for identifying magnetic field abnormalities based on image segmentation according to claim 1, characterized in that, The improved FlowSDF multi-resolution bidirectional velocity field prediction model based on the FlowSDF architecture includes: A sensor influence weighting field is constructed based on the total number of sensor channels, the signal-to-noise ratio weighting coefficient, and the sensor spatial coordinates. Set a low-resolution grid scale and construct a low-resolution initial symbolic distance field on the low-resolution grid; Based on the FlowSDF architecture, a low-resolution positive velocity field prediction function is constructed, and a physical divergence constraint is applied to the low-resolution positive velocity field prediction function. The physical divergence constraint is a weighted constraint on the divergence consistency of the velocity field through the sensor influence weight field. In the low-resolution positive velocity field prediction function, the fused magnetic cardiometa-feature tensor is used as the conditional input to predict the low-resolution positive velocity field under different continuous time variables. By continuously evolving the low-resolution initial symbolic distance field along the low-resolution positive velocity field, the low-resolution initial symbolic distance field gradually converges into a low-resolution coarse-resolution symbolic distance field that conforms to the spatial distribution characteristics of small lesions in the magnetic cardiometa-abnormal region. The results of the low-resolution coarse-resolution symbolic distance field at the continuous time variable t=1 are improved to a high-resolution grid through spatial interpolation to form a high-resolution initial symbolic distance field. Based on the high-resolution initial symbol distance field, a narrow-band refinement region is constructed to define the boundary refinement range of tiny lesions in the magnetic anomaly region with a diameter smaller than the sensor spacing. Based on the FlowSDF architecture, a high-resolution inverse velocity field prediction function is constructed in the narrowband refinement region, and physical divergence constraints are applied to the high-resolution inverse velocity field prediction function. As the high-resolution inverse velocity field prediction function evolves from 1 to 0 in continuous time, the high-resolution initial symbolic distance field is numerically updated at each discrete time step. When the continuous time variable evolves to the termination time, a FlowSDF symbolic distance field with uniform resolution and continuous differentiability defined on the entire high-resolution grid is obtained.

5. The method for identifying magnetic field anomaly regions based on image segmentation according to claim 1, characterized in that, The step of extracting the zero isosurface from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region and calculating the corresponding gradient and curvature information includes: The zero isosurface is extracted from the FlowSDF symbolic distance field as the initial boundary of the magnetic anomaly region; The spatial gradient information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the spatial gradient vector; The curvature information of the FlowSDF symbol distance field is calculated on the initial boundary of the magnetic anomaly region to obtain the curvature; Within the neighborhood of the initial boundary of the magnetic anomaly region, a sub-pixel virtual sampling grid is constructed; Based on curvature, the adaptive step size for sub-pixel space reconstruction is calculated; Based on the adaptive step size, the boundary points of the initial boundary of the magnetic anomaly region are iteratively updated within the sub-pixel virtual sampling grid; When the number of iterations reaches the set number of iterations or the absolute value of the sign distance at the boundary point position in the nth iteration is less than the preset convergence threshold, the iteration is terminated, and all updated boundary point positions are taken as the boundaries of the sub-pixel precision abnormal region.

6. The method for identifying magnetic field anomaly regions based on image segmentation according to claim 1, characterized in that, The process of performing topological consistency constraints and connected component integrity checks on the boundaries of sub-pixel precision anomaly regions includes: In the magnetocardiogram reconstruction space, the boundary of the sub-pixel precision anomalous region is mapped to a discrete grid used to generate the anomalous region mask, and a discrete boundary indicator function is obtained. Based on the discrete boundary indicator function, the closed region enclosed by the boundary of the sub-pixel precision anomalous region is filled to obtain the initial anomalous region mask. Perform connected component decomposition based on the initial abnormal region mask to obtain a set of connected components, and mark each connected component as a connected component unit; For each connected component, calculate the number of connected component voxels and the set of connected component boundary points; Based on the number of connected voxels, a connected component integrity check is performed on the connected component set. Connected components with a number of connected voxels greater than or equal to the minimum connected component voxel threshold are retained as candidate lesion connected components. Topological consistency constraints are then applied to obtain the constrained candidate lesion connected components. If the minimum boundary distance between any two candidate lesion connected units is less than or equal to the topological break merging threshold, then a connection path is constructed between the boundary point pairs corresponding to the minimum boundary distance, and the spatial position covered by the connection path is assigned a value of 1 in the initial abnormal region mask to generate the bridged abnormal region mask. Perform a union update on the bridged abnormal region mask by applying the constrained candidate lesion connected components to obtain a topologically accurate abnormal region mask.

7. The method for identifying magnetic field abnormality regions based on image segmentation according to claim 1, characterized in that, The implementation of Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference phase includes: In each Monte Carlo random inactivation inference, a random inactivation mask is applied to the network unit of the improved FlowSDF multi-resolution bidirectional velocity field prediction model to form a Monte Carlo random inactivation inference model, and the FlowSDF symbolic distance field is output by the Monte Carlo random inactivation inference model. The mean symbol distance field of the FlowSDF symbol distance field is calculated at each spatial location in the magnetocardiogram reconstruction space, and the spatial variance field of the symbol distance is calculated through the mean symbol distance field. Confidence assessment results are generated based on the symbolic distance space variance field. Based on the confidence assessment results, dynamic morphological filtering is performed on the topology-fidelity outlier mask to obtain the outlier mask after dynamic morphological filtering, which is denoted as the final outlier mask.

8. The method for identifying magnetic field abnormality regions based on image segmentation according to claim 1, characterized in that, The set of quantization indicators for anomaly regions calculated based on the final anomaly region mask and the FlowSDF symbolic distance field includes: Calculate the spatial area of ​​the abnormal region based on the final abnormal region mask; Calculate the geometric center of the abnormal region based on the final abnormal region mask; Based on the FlowSDF symbolic distance field and the boundary of the sub-pixel precision anomaly region, the curvature value is calculated and a set of boundary curvature distributions is formed. Based on the FlowSDF symbol distance field, the peak value of the magnetic field gradient inside the anomaly region is calculated; The spatial area, geometric center location, boundary curvature distribution set, and magnetic field gradient peak of the abnormal region are combined to form a quantitative index set of the abnormal region, which is then output together with the final abnormal region mask to form the identification result of small lesions in the cardiac magnetic abnormal region.

9. A system for identifying magnetic field abnormalities based on image segmentation, used to execute the method for identifying magnetic field abnormalities based on image segmentation as described in any one of claims 1-8, characterized in that, include: The data acquisition and processing module acquires the raw multi-channel magnetic field signal data output by the multi-channel magnetic field detection device and performs preprocessing to obtain preprocessed magnetic field signals, and generates a signal-to-noise ratio weighted magnetic field signal dataset. The feature fusion module fuses the adaptive position-coded feature set with the signal-to-noise ratio-weighted magnetocardiogram (MCG) signal dataset to generate a fused MCG feature tensor. An improved FlowSDF module was developed, and an improved FlowSDF multi-resolution bidirectional velocity field prediction model was constructed based on the FlowSDF architecture to obtain the FlowSDF symbol range field. The sub-pixel space reconstruction module calculates the corresponding gradient and curvature information, and uses an adaptive step size to iteratively update the zero isosurface position based on the gradient and curvature information to perform sub-pixel space reconstruction and obtain the boundary of the sub-pixel accuracy anomaly region. The verification module performs topological consistency constraints and connectivity integrity checks on the boundaries of sub-pixel precision abnormal regions, and generates a topologically accurate abnormal region mask. The filtering module performs Monte Carlo random inactivation inference on the improved FlowSDF multi-resolution bidirectional velocity field prediction model during the inference stage, obtains multiple sets of FlowSDF symbol distance fields and calculates the spatial variance, generates confidence evaluation results based on the spatial variance, and performs dynamic morphological filtering on the anomaly region mask through the confidence evaluation results to obtain the final anomaly region mask. The identification results module calculates a set of quantitative indicators for abnormal regions based on the final abnormal region mask and the FlowSDF symbolic distance field, forming the identification results of small lesions in the magnetic field abnormal region.

Images