Method for detecting coronary heart disease based on magnetocardiogram image analysis and lesion region identification

Abstract

This invention discloses a method for detecting coronary artery disease (CAD) based on magnetocardiogram (MCC) image analysis and lesion region identification, belonging to the field of CAD technology. The method involves obtaining a baseline-calibrated MCC spatial image dataset; generating a corresponding MCC frequency domain spectral image dataset; obtaining a high-frequency enhanced frequency domain spectral image dataset; obtaining a frequency-domain enhanced MCC spatial image dataset and constructing a spatial-frequency domain joint enhancement feature tensor; obtaining a patient-level adaptively fine-tuned encoded feature map, which is then input into a decoder with an improved TTT-UNet network structure to obtain lesion candidate feature maps; generating a corrected lesion saliency probability distribution map; obtaining a MCC lesion region mask; and outputting a MCC lesion identification result map for the diagnosis of ischemic areas in CAD. This invention achieves dynamic adaptation to individual differences, significantly improving the robustness and stability of MCC lesion identification across different patients.

Description

Technical Field

[0001] This invention relates to the field of coronary artery disease technology, and in particular to a method for detecting coronary artery disease based on magnetic resonance imaging analysis and lesion region identification. Background Technology

[0002] Magnetocardiography (MCG) is a medical imaging technique that detects extremely weak magnetic field signals generated by the electrical activity of the human heart, thereby enabling non-contact monitoring of the myocardial electrophysiological state. Due to its non-invasive, radiation-free nature and ability to reflect abnormal changes in myocardial electrophysiology, MCG has gained increasing attention in recent years for clinical diagnostic applications such as ischemic area identification in coronary artery disease, arrhythmia detection, and myocardial lesion assessment. By analyzing the spatial distribution of MCG images, abnormal magnetic field regions corresponding to localized abnormal electrical activity in the myocardium can be identified, thus assisting physicians in locating and assessing ischemic areas in coronary artery disease.

[0003] In existing technologies, methods for identifying lesions in magnetocardiogram (MCG) images mainly include signal processing-based methods and deep learning-based methods. Signal processing-based methods typically identify abnormal regions by filtering, performing spectral analysis, or extracting statistical features from the MCG signal, and then combining this with threshold rules or template matching. However, due to the inherent characteristics of MCG signals, such as extremely weak amplitude, strong noise interference, and significant individual differences, relying solely on single spatial or frequency domain features for analysis often fails to stably represent the true abnormal magnetic field distribution of the myocardium, resulting in low accuracy in lesion localization.

[0004] With the development of deep learning technology, some existing techniques have begun to apply convolutional neural networks to magnetocardiogram (MCC) image analysis. These deep network models are trained to segment lesions or detect abnormal regions in MCC images. However, most existing deep learning methods directly input MCC images as ordinary medical images into the network for processing, without fully considering the physical correlation between the spatial magnetic field distribution and frequency domain abnormal energy in the MCC signal. Furthermore, due to differences in thoracic cavity structure, cardiac spatial location, and baseline drift among different subjects, significant distribution variations exist in MCC images among different individuals.

[0005] Furthermore, in the process of segmenting lesion regions in magnetocardiogram (MCC) images, existing technologies typically rely on simple probability thresholding or conventional post-processing methods to generate lesion region masks. However, in practical applications, there are a large number of non-lesion magnetic field disturbances and local abnormal responses caused by noise in MCC images. Relying solely on a single probability determination can easily lead to blurred lesion region boundaries or false detections. Summary of the Invention

[0006] One objective of this invention is to propose a method for detecting coronary heart disease based on magnetocardiographic image analysis and lesion region identification. This invention achieves dynamic adaptation to individual differences and significantly improves the robustness and stability of magnetocardiographic image lesion identification among different patients.

[0007] The method for detecting coronary heart disease based on magnetocardiographic image analysis and lesion region identification according to embodiments of the present invention includes:

[0008] Raw magnetic field image data of the chest cavity surface of the subject were acquired and preprocessed to obtain a baseline-calibrated magnetic field spatial image data set;

[0009] A two-dimensional frequency domain transformation is performed on the baseline calibration magnetic-cardiogram spatial image dataset to generate the corresponding magnetic-cardiogram frequency domain spectral image dataset.

[0010] The spectral energy distribution of the magnetocardiogram frequency domain spectral image dataset is calculated, and spectral weights are applied to the high-frequency region to highlight high-frequency abnormal energy, resulting in a high-frequency enhanced frequency domain spectral image dataset.

[0011] An inverse frequency domain transformation is performed on the high-frequency enhanced frequency domain spectral image dataset to obtain a frequency domain enhanced magnetocardiogram spatial image dataset, which is then aligned and fused with the baseline calibrated magnetocardiogram spatial image dataset in the channel dimension to construct a joint spatial-frequency domain enhancement feature tensor.

[0012] The spatial-frequency domain joint enhanced feature tensor is input into the encoder of the improved TTT-UNet network structure. During the inference phase, when the test is triggered, the adaptive fine-tuning module is used to construct the self-supervised auxiliary task loss using the spatial-frequency domain joint enhanced feature tensor. The hidden layer weights of the encoder are updated in real time through backpropagation. The encoded feature map after patient-level adaptive fine-tuning is then input into the decoder of the improved TTT-UNet network structure to perform layer-by-layer upsampling and cross-layer feature fusion to obtain the lesion candidate feature map.

[0013] The saliency probability distribution map of lesions is calculated based on the lesion candidate feature map, and the saliency probability distribution map of lesions is corrected at the pixel level to generate the corrected saliency probability distribution map of lesions.

[0014] The lesion saliency probability distribution map after correction is segmented into regions according to a preset saliency threshold to obtain a lesion region mask in the magnetocardiogram image;

[0015] Based on the lesion region mask of the magnetocardiogram image, the spatial location of the lesion, the boundary contour of the lesion, and the salience probability information of the lesion are reconstructed, and the magnetocardiogram image lesion identification result map for the diagnosis of ischemic areas of coronary heart disease is output.

[0016] Optionally, the preprocessing includes time synchronization correction, environmental magnetic noise suppression, and individual baseline stabilization.

[0017] Optionally, the two-dimensional frequency domain transformation of the baseline-calibrated magnetocardiogram spatial image dataset includes:

[0018] A two-dimensional discrete frequency domain mapping relationship is established for each frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset to obtain the corresponding complex magnetocardiogram frequency domain spectrum image.

[0019] For the complex magnetic field frequency domain spectrum image corresponding to the baseline-calibrated magnetic field spatial image of the qth frame, the complex magnetic field frequency domain spectrum value at each frequency domain coordinate is decomposed into real part spectrum value and imaginary part spectrum value;

[0020] For each frequency domain coordinate position, calculate the corresponding amplitude spectrum value and phase spectrum value for the real part and imaginary part spectrum values;

[0021] Perform a frequency domain centering rearrangement operation on the amplitude spectrum values ​​to obtain the centered amplitude spectrum values;

[0022] The centered amplitude spectrum value and the corresponding phase spectrum value are jointly represented to obtain the frequency domain characterization result of the baseline-calibrated magnetocardiogram spatial image of the qth frame. The frequency domain characterization results of all frames of baseline-calibrated magnetocardiogram spatial images are then organized into a set according to the image frame number to construct a magnetocardiogram frequency domain spectrum image data set.

[0023] Optionally, the step of calculating the spectral energy distribution for the magnetocardiogram frequency domain spectral image dataset and applying spectral weights in the high-frequency region to highlight high-frequency anomalous energy includes:

[0024] Based on the frequency domain characterization results in the magnetocardiogram frequency domain spectral image dataset, a spectral energy distribution is established based on the centered amplitude spectral image to obtain the spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image.

[0025] A frequency-domain radial distance distribution is established based on the frequency-domain center position of the centered amplitude spectrum image, and the frequency-domain radial distance value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained.

[0026] Based on the frequency domain radial distance value, the high-frequency region determination result is constructed, and the high-frequency region mask value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained.

[0027] Spectral weights are applied to the frequency domain coordinates where the mask value in the high-frequency region is equal to 1, to obtain the high-frequency spectral weight values ​​corresponding to the baseline-calibrated magnetocardiogram spatial image.

[0028] The spectral energy value is enhanced by weighting the spectral energy value with high-frequency spectral weights to obtain the high-frequency enhanced spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image.

[0029] Based on the high-frequency enhanced spectral energy value, the corresponding high-frequency enhanced amplitude spectral value is recovered and jointly represented with the phase spectral value to obtain the high-frequency enhanced frequency domain spectral image corresponding to the baseline calibrated magnetocardiogram spatial image;

[0030] All frames of high-frequency enhanced frequency domain spectral images are organized into a set according to the image frame number to construct a high-frequency enhanced frequency domain spectral image data set.

[0031] Optionally, performing an inverse frequency domain transform on the high-frequency enhanced frequency domain spectral image data set includes:

[0032] For each frame of high-frequency enhanced frequency domain spectrum image in the high-frequency enhanced frequency domain spectrum image dataset, the complex high-frequency enhanced frequency domain spectrum value at the corresponding frequency domain coordinate position is reconstructed based on the high-frequency enhanced amplitude spectrum value and phase spectrum value.

[0033] A two-dimensional inverse discrete frequency domain transform is performed on the complex high-frequency enhanced frequency domain spectral values ​​to obtain the corresponding frequency domain enhanced magnetocardiogram spatial image;

[0034] The real part of the inverse frequency domain transformation result is taken from the frequency domain enhanced magnetocardiogram spatial image to obtain the real-valued frequency domain enhanced magnetocardiogram spatial image corresponding to the baseline calibrated magnetocardiogram spatial image;

[0035] All real-value frequency-domain enhanced magnetocardiogram spatial images are grouped and organized according to the image frame number to construct a frequency-domain enhanced magnetocardiogram spatial image dataset;

[0036] For each frame of real-valued frequency-domain enhanced magnetocardiogram spatial image in the frequency-domain enhanced magnetocardiogram spatial image dataset and the corresponding frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset, a one-to-one spatial coordinate correspondence channel alignment relationship is established to obtain a real-valued frequency-domain enhanced magnetocardiogram spatial image with complete channel alignment.

[0037] The channel-aligned real-valued frequency-domain enhanced magnetocardiogram spatial image and the baseline-calibrated magnetocardiogram spatial image are fused in the channel dimension to obtain the spatial-frequency joint enhanced feature tensor of the corresponding frame;

[0038] All frame spatial-frequency domain joint enhancement feature tensors are organized into sets according to the image frame number to construct a set of spatial-frequency domain joint enhancement feature tensors.

[0039] Optionally, the step of inputting the spatial-frequency domain joint enhanced feature tensor into the encoder of the improved TTT-UNet network structure, and the adaptive fine-tuning module when triggering testing during the inference phase, utilizes the spatial-frequency domain joint enhanced feature tensor to construct a self-supervised auxiliary task loss, and performs real-time backpropagation updates on the hidden layer weights of the encoder to obtain the encoded feature map after patient-level adaptive fine-tuning, which is then input into the decoder of the improved TTT-UNet network structure, includes:

[0040] The spatial-frequency domain joint enhancement feature tensor is input into the encoder of the improved TTT-UNet network structure, and a shallow joint response feature map is constructed based on the spatial domain channel and the frequency domain enhanced spatial channel.

[0041] Based on the shallow joint response feature map, a patient-level baseline drift representation vector is established, and the channel scaling parameters and channel translation parameters of the encoder hidden layer are generated based on the patient-level baseline drift representation vector.

[0042] Patient-level adaptive modulation of encoder hidden layer features is performed using channel scaling and channel translation parameters to obtain encoded feature maps containing patient-level baseline compensation information.

[0043] When the test is triggered during the inference stage of the improved TTT-UNet network structure, the adaptive fine-tuning module is used to construct a mask reconstruction consistency self-supervised auxiliary task based on the spatial-frequency domain joint enhanced feature tensor, and the self-supervised task loss is obtained.

[0044] The weight parameters of the encoder hidden layer and the weight parameters of the patient-level adaptive modulation parameters are updated in real time by backpropagation based on the self-supervised auxiliary task loss. The update is performed in a preset number of iterations or until the self-supervised auxiliary task loss value is less than the preset convergence threshold, so as to obtain the encoded feature map after patient-level adaptive fine-tuning.

[0045] The encoded feature map of patient-level baseline compensation information and the encoded feature map after patient-level adaptive fine-tuning are input into the decoder of the improved TTT-UNet network structure, layer-by-layer upsampling is performed, and cross-layer feature fusion is performed using the encoded feature map of the corresponding layer of the encoder after patient-level adaptive modulation to obtain the decoded fused feature map.

[0046] The lesion candidate response is mapped onto the decoded and fused feature map to obtain the lesion candidate feature map.

[0047] Optionally, the step of calculating the lesion saliency probability distribution map based on the lesion candidate feature map and performing pixel-level correction on the lesion saliency probability distribution map includes:

[0048] Based on the candidate feature map of lesions, the initial lesion significance probability value is calculated;

[0049] Based on the feature differences between the spatial domain channel and the frequency domain enhanced spatial channel at the same spatial coordinates in the spatial-frequency domain joint enhanced feature tensor, a spatial-frequency domain consistency deviation value is constructed.

[0050] Based on the spatial-frequency domain consistency deviation value, construct the spatial-frequency domain consistency constraint weight value;

[0051] Based on the initial lesion significance probability value, construct the significance probability value of neighboring lesions;

[0052] The initial lesion saliency probability value and the saliency probability value of neighboring lesions are jointly corrected at the pixel level using the spatial-frequency domain consistency constraint weight value to generate the corrected lesion saliency probability value.

[0053] The corrected lesion saliency probability values ​​at all spatial coordinates are grouped and organized according to the image frame number to construct a corrected lesion saliency probability distribution map.

[0054] Optionally, the step of segmenting the corrected lesion significance probability distribution map into regions according to a preset significance threshold includes:

[0055] For the corrected lesion significance probability distribution map, the corrected lesion significance probability value is read at the spatial coordinate (x,y), and the corrected lesion significance probability value is compared with the preset significance threshold to obtain the lesion region judgment value at the spatial coordinate (x,y).

[0056] All lesion region judgment values ​​are organized into binary regions according to spatial coordinate distribution to obtain a lesion region mask for magnetocardiogram images.

[0057] The mask value of the lesion region mask in the magnetic resonance imaging at spatial coordinates (x, y) is equal to the lesion region determination value at spatial coordinates (x, y). When the mask value is equal to 1, it means that the current spatial coordinate position belongs to the magnetic resonance lesion region. When the mask value is equal to 0, it means that the current spatial coordinate position belongs to the non-lesion region.

[0058] Optionally, the step of reconstructing the spatial location of the lesion, the boundary contour of the lesion, and the salience probability information of the lesion based on the lesion region mask of the magnetocardiogram image includes:

[0059] Based on the lesion region mask of magnetocardiogram image, extract the set of spatial location coordinates of the lesion;

[0060] Calculate the center coordinates of the lesion's spatial location based on the set of lesion spatial location coordinates;

[0061] Based on the lesion region mask of magnetocardiogram image, the set of lesion boundary contour coordinates is extracted by morphological boundary extraction;

[0062] Based on the corrected lesion saliency probability distribution map and the lesion region mask of the magnetocardiogram image, the lesion saliency probability information is extracted;

[0063] Based on the set of spatial location coordinates of lesions, the set of boundary contour coordinates of lesions, and the saliency probability information of lesions, a magnetic resonance imaging lesion identification result map corresponding to the baseline-calibrated magnetic resonance spatial image is constructed.

[0064] The beneficial effects of this invention are:

[0065] (1) This invention achieves coupled expression of magnetic field distribution information in the spatial domain and abnormal energy information in the frequency domain of magnetic field images by constructing a spatial-frequency domain joint enhancement feature tensor, which significantly improves the expressive ability of magnetic field lesion features. By performing a two-dimensional frequency domain transformation on the baseline-calibrated magnetic field spatial image data set, a magnetic field frequency domain spectrum image data set is constructed. By calculating the spectral energy distribution, spectral weights are applied to the high-frequency region to highlight high-frequency abnormal energy, thereby obtaining a high-frequency enhanced frequency domain spectrum image data set. The frequency domain enhanced magnetic field spatial image data set is obtained by inverse frequency domain transformation, and the frequency domain enhanced magnetic field spatial image data set is combined with the baseline-calibrated magnetic field spatial image data set. By aligning and fusing the dataset along the channel dimension, a spatial-frequency domain joint enhanced feature tensor is constructed. This allows the network to simultaneously acquire the spatial distribution structure of the magnetic field and high-frequency abnormal perturbation features during the input stage. Since myocardial ischemia areas usually correspond to local magnetic field abnormalities accompanied by enhanced high-frequency perturbations, the spatial-frequency domain joint enhanced feature tensor of this invention can highlight the abnormal response of the real lesion area at the feature expression level. This enables the improved TTT-UNet network structure to extract cardiac magnetic lesion features more accurately in the subsequent encoding process. Compared with traditional methods that only use spatial domain images, it can significantly improve the stability and accuracy of cardiac magnetic lesion identification.

[0066] (2) The present invention introduces a patient-level test-time adaptive fine-tuning mechanism in the improved TTT-UNet network structure. Through patient-level baseline drift representation and channel adaptive modulation, the model achieves dynamic adaptation to the differences in the magnetic field baseline of different subjects. The improved TTT-UNet network structure introduces a test-time adaptive fine-tuning module. By performing global statistics on the spatial-frequency domain joint enhancement feature tensor, a patient-level baseline drift representation vector is constructed. This representation vector is then used to generate the channel scaling parameters and channel translation parameters of the encoder hidden layer. Patient-level adaptive modulation is then performed on the encoder hidden layer features to compensate for the differences in the amplitude and spatial distribution of the magnetic field baseline between different subjects. The improved TTT-UNet network structure can achieve dynamic adaptation to individual differences while maintaining the expressive power of deep networks, significantly improving the robustness and stability of magnetic field image lesion identification among different patients.

[0067] (3) This invention achieves refined correction of the saliency probability distribution of lesions by constructing a pixel-level probability correction mechanism with spatial-frequency domain consistency constraints, thereby improving the accuracy of the boundary positioning of the cardiac lesion region. By using the weight value of spatial-frequency domain consistency constraints to perform joint pixel-level correction on the initial lesion saliency probability value and the saliency probability value of neighboring lesions, it effectively suppresses false abnormal responses caused by noise, and can more accurately restore the spatial boundary structure of the real cardiac lesion region, thereby improving the consistency between the lesion region segmentation result and the real myocardial ischemia region. Attached Figure Description

[0068] 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:

[0069] Figure 1 This is a flowchart of the coronary heart disease detection method based on magnetocardiographic image analysis and lesion region identification proposed in this invention;

[0070] Figure 2 This is a block diagram of the improved TTT-UNet network structure in the coronary heart disease detection method based on magnetocardiographic image analysis and lesion region identification proposed in this invention. Detailed Implementation

[0071] Example 1: Reference Figures 1-2 Coronary artery disease detection methods based on magnetocardiographic image analysis and lesion region identification include:

[0072] Raw magnetic field image data of the chest cavity surface of the subject were acquired and preprocessed to obtain a baseline-calibrated magnetic field spatial image data set;

[0073] In this embodiment, the preprocessing includes time synchronization correction, environmental magnetic noise suppression, and individual baseline stabilization.

[0074] A two-dimensional frequency domain transformation is performed on the baseline calibration magnetic-cardiogram spatial image dataset to generate the corresponding magnetic-cardiogram frequency domain spectral image dataset.

[0075] In this embodiment, a two-dimensional frequency domain transformation is performed on the baseline calibration magnetocardiogram spatial image data set, including:

[0076] A two-dimensional discrete frequency domain mapping relationship is established for each frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset to obtain the corresponding complex magnetocardiogram frequency domain spectrum image.

[0077] For the complex magnetic field frequency domain spectrum image corresponding to the baseline-calibrated magnetic field spatial image of the qth frame, the complex magnetic field frequency domain spectrum value at each frequency domain coordinate is decomposed into real part spectrum value and imaginary part spectrum value;

[0078] For each frequency domain coordinate position, calculate the corresponding amplitude spectrum value and phase spectrum value for the real part and imaginary part spectrum values;

[0079] In Example 1, the amplitude spectrum value is obtained by squaring the real part spectrum value and the imaginary part spectrum value, adding them together, and then performing a square root operation. The amplitude spectrum value is used to represent the energy intensity of the magnetocardiographic frequency component corresponding to the frequency domain coordinate position.

[0080] The phase spectrum value is obtained by calculating the four-quadrant arctangent function between the imaginary part spectrum value and the real part spectrum value. The phase spectrum value is used to represent the phase angle of the magnetic frequency component corresponding to the frequency domain coordinate position.

[0081] Perform a frequency domain centering rearrangement operation on the amplitude spectrum values ​​to obtain the centered amplitude spectrum values;

[0082] In Example 1, the low-frequency component in the amplitude spectrum located in the upper left corner of the frequency domain is moved to the center of the frequency domain image, and the high-frequency component in the amplitude spectrum located in the center of the frequency domain is moved to the edge of the frequency domain image.

[0083] The centered amplitude spectrum value and the corresponding phase spectrum value are jointly represented to obtain the frequency domain characterization result of the baseline-calibrated magnetocardiogram spatial image of the qth frame. The frequency domain characterization results of all frames of baseline-calibrated magnetocardiogram spatial images are then organized into a set according to the image frame number to construct a magnetocardiogram frequency domain spectrum image data set.

[0084] The spectral energy distribution of the magnetocardiogram frequency domain spectral image dataset is calculated, and spectral weights are applied to the high-frequency region to highlight high-frequency abnormal energy, resulting in a high-frequency enhanced frequency domain spectral image dataset.

[0085] In this embodiment, the spectral energy distribution of the magnetocardiogram frequency domain spectral image dataset is calculated, and spectral weights are applied in the high-frequency region to highlight high-frequency abnormal energy, including:

[0086] Based on the frequency domain characterization results in the magnetocardiogram frequency domain spectral image dataset, a spectral energy distribution is established based on the centered amplitude spectral image to obtain the spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image.

[0087] In Example 1, the spectral energy value is used to represent the energy intensity of the magnetocardiogram frequency component at the frequency domain coordinate position, which is obtained by squaring the amplitude spectrum value of the centered amplitude spectrum image at the corresponding frequency domain coordinate position.

[0088] A frequency-domain radial distance distribution is established based on the frequency-domain center position of the centered amplitude spectrum image, and the frequency-domain radial distance value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained.

[0089] In Example 1, the distance difference between the frequency domain coordinates (u,v) and the frequency domain center position in the horizontal frequency domain direction and the distance difference in the vertical frequency domain direction are calculated. The distance difference in the horizontal frequency domain direction and the distance difference in the vertical frequency domain direction are squared respectively. The two squared distance differences are summed and the square root of the sum is calculated to obtain the frequency domain radial distance value.

[0090] Based on the frequency domain radial distance value, the high-frequency region determination result is constructed, and the high-frequency region mask value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained.

[0091] In Example 1, the high-frequency region mask value is used to indicate whether the frequency domain coordinates (u,v) belong to the high-frequency region:

[0092] When the frequency domain radial distance value corresponding to the baseline calibrated magnetocardiogram spatial image is greater than or equal to the preset high-frequency region determination threshold, the high-frequency region mask value corresponding to the current frequency domain coordinates (u,v) is set to 1.

[0093] When the frequency domain radial distance value corresponding to the baseline calibrated magnetocardiogram spatial image is less than the preset high-frequency region determination threshold, the high-frequency region mask value corresponding to the current frequency domain coordinates (u,v) is set to 0.

[0094] Spectral weights are applied to the frequency domain coordinates where the mask value in the high-frequency region is equal to 1, to obtain the high-frequency spectral weight values ​​corresponding to the baseline-calibrated magnetocardiogram spatial image.

[0095] In Example 1, the difference between the frequency domain radial distance value and the preset high-frequency region determination threshold is calculated, the normalization ratio between the difference and the maximum frequency domain radial distance value is calculated, the normalization ratio is multiplied by the preset high-frequency enhancement coefficient, and the calculation result is added to the constant 1 to obtain the high-frequency spectrum weight value.

[0096] Among them, the maximum radial distance value in the frequency domain represents the maximum distance between the center position and the corner position in the frequency domain, and the high-frequency enhancement coefficient is used to control the enhancement amplitude of spectral energy in the high-frequency region.

[0097] The spectral energy value is enhanced by weighting the spectral energy value with high-frequency spectral weights to obtain the high-frequency enhanced spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image.

[0098] Based on the high-frequency enhanced spectral energy value, the corresponding high-frequency enhanced amplitude spectral value is recovered and jointly represented with the phase spectral value to obtain the high-frequency enhanced frequency domain spectral image corresponding to the baseline calibrated magnetocardiogram spatial image;

[0099] All frames of high-frequency enhanced frequency domain spectral images are organized into a set according to the image frame number to construct a high-frequency enhanced frequency domain spectral image data set.

[0100] An inverse frequency domain transformation is performed on the high-frequency enhanced frequency domain spectral image dataset to obtain a frequency domain enhanced magnetocardiogram spatial image dataset, which is then aligned and fused with the baseline calibrated magnetocardiogram spatial image dataset in the channel dimension to construct a joint spatial-frequency domain enhancement feature tensor.

[0101] In this embodiment, performing an inverse frequency domain transform on the high-frequency enhanced frequency domain spectral image data set includes:

[0102] For each frame of high-frequency enhanced frequency domain spectrum image in the high-frequency enhanced frequency domain spectrum image dataset, the complex high-frequency enhanced frequency domain spectrum value at the corresponding frequency domain coordinate position is reconstructed based on the high-frequency enhanced amplitude spectrum value and phase spectrum value.

[0103] In Example 1, the complex high-frequency enhanced frequency domain spectrum value at the frequency domain coordinates (u,v) corresponding to the baseline calibration magnetocardiogram spatial image of the qth frame is obtained by reconstructing the high-frequency enhanced amplitude spectrum value and the phase spectrum value at the frequency domain coordinates (u,v) using complex polar coordinates. The real part of the complex high-frequency enhanced frequency domain spectrum value is obtained by multiplying the high-frequency enhanced amplitude spectrum value by the cosine of the phase spectrum value, and the imaginary part of the complex high-frequency enhanced frequency domain spectrum value is obtained by multiplying the high-frequency enhanced amplitude spectrum value by the sine of the phase spectrum value. The real part and the imaginary part are then combined to form the complex high-frequency enhanced frequency domain spectrum value.

[0104] A two-dimensional inverse discrete frequency domain transform is performed on the complex high-frequency enhanced frequency domain spectral values ​​to obtain the corresponding frequency domain enhanced magnetocardiogram spatial image;

[0105] The real part of the inverse frequency domain transformation result is taken from the frequency domain enhanced magnetocardiogram spatial image to obtain the real-valued frequency domain enhanced magnetocardiogram spatial image corresponding to the baseline calibrated magnetocardiogram spatial image;

[0106] In Example 1, for the frequency-domain enhanced magnetocardiogram spatial image corresponding to the baseline-calibrated magnetocardiogram spatial image of the qth frame, the real-valued frequency-domain enhanced magnetocardiogram spatial image is obtained by taking the real part of the complex result of the frequency-domain enhanced magnetocardiogram spatial image at spatial coordinates (x,y).

[0107] All real-value frequency-domain enhanced magnetocardiogram spatial images are grouped and organized according to the image frame number to construct a frequency-domain enhanced magnetocardiogram spatial image dataset;

[0108] For each frame of real-valued frequency-domain enhanced magnetocardiogram spatial image in the frequency-domain enhanced magnetocardiogram spatial image dataset and the corresponding frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset, a one-to-one spatial coordinate correspondence channel alignment relationship is established to obtain a real-valued frequency-domain enhanced magnetocardiogram spatial image with complete channel alignment.

[0109] In Example 1, for the spatial coordinates (x, y) corresponding to the baseline calibration magnetocardiogram spatial image of the qth frame, the magnetic field strength value of the real-valued frequency domain enhanced magnetocardiogram spatial image at spatial coordinates (x, y) is aligned with the magnetic field strength value of the baseline calibration magnetocardiogram spatial image at spatial coordinates (x, y) to ensure that the sampling positions of the two are completely consistent in the horizontal discrete spatial coordinates and the vertical discrete spatial coordinates.

[0110] The channel-aligned real-valued frequency-domain enhanced magnetocardiogram spatial image and the baseline-calibrated magnetocardiogram spatial image are fused in the channel dimension to obtain the spatial-frequency joint enhanced feature tensor of the corresponding frame;

[0111] In Example 1, for the baseline-calibrated magnetocardiogram spatial image of frame q, a spatial-frequency domain joint enhancement feature tensor containing two channels is established at spatial coordinates (x,y). The first channel is the magnetic field strength value of the baseline-calibrated magnetocardiogram spatial image at spatial coordinates (x,y), and the second channel is the magnetic field strength value of the real-valued frequency-domain enhanced magnetocardiogram spatial image at spatial coordinates (x,y). The spatial-frequency domain joint enhancement feature tensor is obtained by combining the two channels in the channel order.

[0112] All frame spatial-frequency domain joint enhancement feature tensors are organized into sets according to the image frame number to construct a set of spatial-frequency domain joint enhancement feature tensors.

[0113] The spatial-frequency domain joint enhanced feature tensor is input into the encoder of the improved TTT-UNet network structure. During the inference phase, when the test is triggered, the adaptive fine-tuning module is used to construct the self-supervised auxiliary task loss using the spatial-frequency domain joint enhanced feature tensor. The hidden layer weights of the encoder are updated in real time through backpropagation. The encoded feature map after patient-level adaptive fine-tuning is then input into the decoder of the improved TTT-UNet network structure to perform layer-by-layer upsampling and cross-layer feature fusion to obtain the lesion candidate feature map.

[0114] In this embodiment, the spatial-frequency domain joint enhanced feature tensor is input into the encoder of the improved TTT-UNet network structure. During the inference phase, when testing is triggered, an adaptive fine-tuning module is used to construct a self-supervised auxiliary task loss using the spatial-frequency domain joint enhanced feature tensor. The hidden layer weights of the encoder are then updated in real-time via backpropagation. The resulting patient-level adaptively fine-tuned encoded feature map is then input into the decoder of the improved TTT-UNet network structure, including:

[0115] The spatial-frequency domain joint enhancement feature tensor is input into the encoder of the improved TTT-UNet network structure, and a shallow joint response feature map is constructed based on the spatial domain channel and the frequency domain enhanced spatial channel.

[0116] In Example 1, a local convolutional receptive field is established at spatial coordinates (x, y) for the spatial-frequency joint enhancement feature tensor of the q-th frame. The values ​​of all spatial-frequency joint enhancement feature tensors within the neighborhood of spatial coordinates (x, y) are weighted and summed according to the shallow joint convolution kernel weight parameters to obtain a shallow joint response feature map. The shallow joint response feature map is used to represent the coupling response between the magnetic field strength distribution of the baseline-calibrated magnetocardiogram spatial image and the magnetic field strength distribution of the real-valued frequency-domain enhanced magnetocardiogram spatial image in the local spatial neighborhood.

[0117] ;

[0118] in, Indicates the first Frame baseline calibration of magnetocardiogram spatial images in spatial coordinates First The shallow joint response eigenvalues ​​of the shallow joint feature channels. This indicates the spatial offset of shallow joint convolution kernels. The connection point The input channel and the first Convolution weight parameters for shallow joint feature channels, This indicates the image frame number in the baseline calibration magnetocardiogram spatial image dataset. This indicates the spatial position index of the baseline-calibrated magnetocardiogram spatial image in the horizontal discrete spatial coordinate direction. This indicates the spatial position index of the baseline-calibrated magnetocardiogram spatial image in the longitudinal discrete spatial coordinate direction. This represents the shallow joint feature channel index, used to distinguish different shallow joint feature channels output by the encoder of the improved TTT-UNet network architecture. This represents the input channel index in the joint space-frequency domain enhanced feature tensor. This represents the spatial domain channel corresponding to the baseline-calibrated magnetocardiogram spatial image. This represents the frequency domain enhancement spatial channel corresponding to the real-valued frequency domain enhanced magnetocardiogram spatial image. This represents the spatial offset index of the convolution kernel in the horizontal spatial direction. This represents the spatial offset index of the convolution kernel in the vertical spatial direction. Indicates the radius of the shallow joint convolution kernel. Indicates the first Frame space-frequency domain joint enhanced feature tensor in spatial coordinates Location, No. Feature values ​​on each channel.

[0119] Based on the shallow joint response feature map, a patient-level baseline drift representation vector is established, and the channel scaling parameters and channel translation parameters of the encoder hidden layer are generated based on the patient-level baseline drift representation vector.

[0120] In Example 1, global spatial statistical calculations are performed on the shallow joint response feature map over the entire spatial region to obtain the patient-level baseline drift characterization vector. The patient-level baseline drift characterization value is used to characterize the overall statistical distribution characteristics of the differences in the chest cavity geometry, cardiac anatomical position shift, and individual cardiac magnetic baseline differences of the current subject in the spatial-frequency domain joint response features.

[0121] Each representation value in the patient-level baseline drift representation vector is multiplied by its corresponding scaling weight parameter and then summed. The summation is then added to the scaling bias parameter to obtain the channel scaling parameter. The channel scaling parameter is then constrained to the range of (0,1) by the Sigmoid function to ensure the numerical stability of the adaptive modulation process. The channel scaling parameter is used to adaptively adjust the response amplitude of different feature channels in the encoder hidden layer to compensate for the differences in the magnetic field baseline amplitude between different subjects.

[0122] The channel translation parameter is solved in the same way as the channel scaling parameter. The channel translation parameter is used to adaptively offset the response baseline of different feature channels in the encoder hidden layer, thereby reducing the drift of the magnetic field baseline caused by differences in human body size, sensor distance, and respiratory rhythm.

[0123] Traditional TTT-UNet blindly updates global network weights directly using only the loss function during the fine-tuning phase of testing. Due to significant differences in chest geometry, heart position, and respiratory rhythm among different subjects, severe baseline drift occurs in individual magnetocardiogram images. Traditional methods are prone to gradient explosion or model collapse during single-sample fine-tuning. This implementation introduces a patient-level baseline drift representation vector into the improved TTT-UNet encoder and generates channel scaling and translation parameters through nonlinear mapping. It performs feature-level linear modulation on the hidden layer features, filtering out DC bias and amplitude attenuation caused by differences in body size and sensor distance in advance during the bottom-level forward propagation stage. This provides an absolutely stable numerical space for adaptive fine-tuning and significantly improves the robustness of the model when generalizing across patients.

[0124] Patient-level adaptive modulation of encoder hidden layer features is performed using channel scaling and channel translation parameters to obtain encoded feature maps containing patient-level baseline compensation information.

[0125] In Example 1, the feature value of the original feature map of the encoder hidden layer at spatial coordinates (x,y) is multiplied with the corresponding channel scaling parameter and added with the corresponding channel translation parameter to obtain the encoded feature map after patient-level adaptive modulation. The encoded feature value is used to weaken the interference of individual baseline drift on the local magnetic field abnormal response of the cardiac magnetic lesion and enhance the magnetic field abnormal response feature corresponding to the real myocardial ischemia area.

[0126] The hidden feature value of the encoder hidden layer original feature map at spatial coordinates (x,y) is obtained by multiplying all feature values ​​of the shallow joint response feature map within the corresponding convolutional receptive field by the weight parameters of the encoder hidden layer convolution kernel and summing them. The encoder hidden layer original feature map is used to represent the cardiac lesion representation features of the spatial-frequency domain joint response features at a higher semantic level.

[0127] When the test is triggered during the inference stage of the improved TTT-UNet network structure, the adaptive fine-tuning module is used to construct a mask reconstruction consistency self-supervised auxiliary task based on the spatial-frequency domain joint enhanced feature tensor, and the self-supervised task loss is obtained.

[0128] In Example 1, a self-supervised mask value is adaptively generated based on the local energy distribution of the spatial channel at the spatial coordinate (x,y) position, which is enhanced in the frequency domain. The probability of the self-supervised mask value being set to 0 is positively correlated with the high-frequency energy intensity at the spatial position. When the self-supervised mask value is equal to 1, it means that the spatial-frequency joint enhancement features at the corresponding spatial coordinate position are preserved. When the self-supervised mask value is equal to 0, it means that the spatial-frequency joint enhancement features at the corresponding spatial coordinate position are masked. The eigenvalue of the spatial-frequency joint enhancement feature tensor at the spatial coordinate (x,y) is multiplied with the corresponding self-supervised mask value to obtain the mask input tensor.

[0129] For the mask input tensor, after feature extraction by the encoder, the same patient-level adaptive modulation is performed on it using the channel scaling parameter and the channel translation parameter. Then, the reconstruction tensor is obtained by mapping through the auxiliary reconstruction head, and a mask reconstruction consistency self-supervised auxiliary task loss is constructed.

[0130] ;

[0131] in, Indicates the first The loss value of the self-supervised auxiliary task corresponding to the frame baseline calibration magnetocardiogram spatial image. This represents the total number of sampling points in the horizontal discrete spatial coordinate direction of the baseline-calibrated magnetocardiogram. This represents the total number of sampling points in the longitudinal discrete spatial coordinate direction of the baseline-calibrated magnetocardiogram. This represents the self-supervised mask value of the baseline-calibrated magnetocardiogram spatial image of the q-th frame at spatial coordinates (x, y). Indicates the first Frame baseline calibration of magnetocardiogram spatial images in spatial coordinates First Reconstructed feature values ​​on each channel Represents the spatial-frequency-domain consistency constraint coefficients. Indicates the first Frame baseline calibration of magnetocardiogram spatial images in spatial coordinates First The squared reconstruction error values ​​of each channel are used to measure the degree of reconstruction deviation of the improved TTT-UNet network structure in the occluded spatial-frequency joint enhancement feature region. Indicates the first Frame baseline calibration of magnetocardiogram spatial images in spatial coordinates The squared difference between the spatial domain reconstructed features and the frequency domain enhanced spatial reconstructed features is used to measure the consistency deviation between the spatial domain magnetic field features and the frequency domain enhanced magnetic field features.

[0132] Traditional TTT-UNet self-supervised tasks generally employ random spatial masks and a single spatial domain mean square error loss. This isotropic reconstruction mechanism, when faced with magnetocardiogram images, prioritizes fitting the dominant low-frequency strong background magnetic field of healthy ventricular repolarization, completely ignoring the extremely weak high-frequency distortion signals caused by early ischemia in coronary heart disease, leading to serious missed diagnoses of small lesions. This implementation abandons random masks and adopts a mask matrix based on high-frequency energy adaptive generation, forcing the network to reconstruct high-frequency abnormal regions. During backpropagation fine-tuning, the loss function forcibly forces the spatial domain feature representation to nonlinearly approximate the high-frequency enhanced feature representation in the frequency domain. This makes the improved TTT-UNet no longer just an image reconstructor, but evolves into a lesion feature directional amplifier, improving the significance and detection rate of extremely weak ischemic lesions in complex magnetic field backgrounds.

[0133] The weight parameters of the encoder hidden layer and the weight parameters of the patient-level adaptive modulation parameters are updated in real time by backpropagation based on the self-supervised auxiliary task loss. The update is performed in a preset number of iterations or until the self-supervised auxiliary task loss value is less than the preset convergence threshold, so as to obtain the encoded feature map after patient-level adaptive fine-tuning.

[0134] In Example 1, the gradient of the self-supervised auxiliary task loss with respect to the encoder hidden layer weight parameters is calculated. The gradient is multiplied by the adaptive fine-tuning learning rate during testing. The product is then subtracted from the current encoder hidden layer weight parameters to obtain the updated encoder hidden layer weight parameters.

[0135] The same update method is used to update the weight parameters that map the patient-level baseline drift representation vector to the channel scaling parameter and the weight parameters that map the patient-level baseline drift representation vector to the channel translation parameter. Using the updated encoder hidden layer weight parameters and the generated weight parameters, the forward calculation process is re-executed on the input spatial-frequency domain joint enhancement feature tensor to generate the recalculated encoded feature map of the patient-level baseline compensation information and the deepest layer encoded feature map after patient-level adaptive fine-tuning. Both are then input into the decoder of the improved TTT-UNet network structure.

[0136] The encoded feature map of patient-level baseline compensation information and the encoded feature map after patient-level adaptive fine-tuning are input into the decoder of the improved TTT-UNet network structure, layer-by-layer upsampling is performed, and cross-layer feature fusion is performed using the encoded feature map of the corresponding layer of the encoder after patient-level adaptive modulation to obtain the decoded fused feature map.

[0137] ;

[0138] in, Indicates the first Frame baseline calibration of magnetocardiogram spatial images in spatial coordinates First Decoding fusion feature values ​​of each decoding feature channel, This represents the deepest layer of encoded feature maps after patient-level adaptive fine-tuning. Mapped to the Decoding weight parameters for each decoding feature channel, Encoded feature map representing recalculated patient-level baseline compensation information Mapped to the Cross-layer feature fusion weight parameters for each decoded feature channel. This indicates a layer-by-layer upsampling operation. This indicates the total number of hidden feature channels in the encoder. This indicates the decoding feature channel index.

[0139] The lesion candidate response is mapped onto the decoded and fused feature map to obtain the lesion candidate feature map.

[0140] Traditional TTT-UNet architectures, after weight updates upon completion of testing, often directly sum the shallow encoded features extracted before the update (old weights) with the deep decoded features extracted after the update (new weights) via skip connections. This misaligned fusion of temporal features disrupts the continuity of the magnetic field spatial topology, generating numerous false lesion contours (artifacts) at the decoding output. This implementation defines a reset mechanism after the self-supervised fine-tuning iteration meets the convergence condition, re-executing the forward computation process using the updated encoder weights. By outputting the recalculated cross-layer fusion feature map, it ensures the absolute temporal consistency of shallow spatial details and deep pathological semantics in terms of weight states when the U-Net decoder performs layer-by-layer upsampling fusion. This eliminates the breeding ground for artifacts from the underlying logic, ensuring extremely high fidelity in the spatial location and boundary contours of the final lesion candidate response feature map.

[0141] The saliency probability distribution map of lesions is calculated based on the lesion candidate feature map, and the saliency probability distribution map of lesions is corrected at the pixel level to generate the corrected saliency probability distribution map of lesions.

[0142] In this embodiment, a lesion saliency probability distribution map is calculated based on the lesion candidate feature map, and pixel-level correction is performed on the lesion saliency probability distribution map, including:

[0143] Based on the candidate feature map of lesions, the initial lesion significance probability value is calculated;

[0144] In Example 1, the candidate feature value of the lesion is reversed and then the natural exponent operation is performed and added to the constant 1 to form the denominator. The constant 1 is used as the numerator to obtain the initial lesion significance probability value at the spatial coordinates (x,y). The initial lesion significance probability value is used to represent the lesion significance probability after the joint response of the local magnetic field abnormality of the heart and the high frequency disturbance in the frequency domain at the current spatial coordinate position.

[0145] Based on the feature differences between the spatial domain channel and the frequency domain enhanced spatial channel at the same spatial coordinates in the spatial-frequency domain joint enhanced feature tensor, a spatial-frequency domain consistency deviation value is constructed.

[0146] In Example 1, for the joint spatial-frequency domain enhancement feature tensor, the feature values ​​corresponding to the spatial domain channel and the feature values ​​corresponding to the frequency domain enhancement spatial channel are read at the spatial coordinates (x,y). By calculating the absolute value of the difference between the feature values ​​corresponding to the spatial domain channel and the feature values ​​corresponding to the frequency domain enhancement spatial channel, the spatial-frequency domain consistency deviation value at the spatial coordinates (x,y) is obtained.

[0147] Based on the spatial-frequency domain consistency deviation value, construct the spatial-frequency domain consistency constraint weight value;

[0148] In Example 1, the spatial-frequency domain consistency deviation value is calculated globally to obtain the global average consistency deviation value.

[0149] The normalized consistency deviation value is obtained by dividing the spatial-frequency domain consistency deviation value by the sum of the global average consistency deviation value and the spatial-frequency domain consistency stability constant. The normalized consistency deviation value is then inversely calculated by performing a natural exponentiation operation to obtain the spatial-frequency domain consistency constraint weight value. The spatial-frequency domain consistency constraint weight value is used to represent the consistency strength between the spatial domain magnetic field response and the frequency domain enhanced spatial magnetic field response at the current spatial coordinate position.

[0150] Based on the initial lesion significance probability value, construct the significance probability value of neighboring lesions;

[0151] In Example 1, a pixel-level correction neighborhood window is established around the spatial coordinates (x, y). Within the range of the pixel-level correction neighborhood window, the initial lesion saliency probability value corresponding to each neighborhood spatial coordinate position is read. For each neighborhood spatial coordinate position within the pixel-level correction neighborhood window, the corresponding initial lesion saliency probability value is multiplied by the corresponding neighborhood correction weight parameter. All product results are summed to obtain the neighborhood lesion saliency probability value at the spatial coordinates (x, y), which is used to represent the average saliency response level of the local magnetic lesion candidate region around the current spatial coordinate position.

[0152] The neighborhood correction weight parameter is used to adjust the contribution of different neighborhood positions around the current spatial coordinate position to the pixel-level correction result. It assigns high smoothing weights to regions with gentle changes in spatial magnetic field strength, and rapidly decays the smoothing weights when crossing steep gradient boundaries of the dipole magnetic field of the cardiac lesion. During probabilistic correction, it achieves strict edge-preserving smoothing of the physical boundaries of weak coronary lesions. For the baseline calibration of the q-th frame, the neighborhood correction weight parameter is used to determine the saliency probability value of the neighborhood lesion at spatial coordinates (x, y) in the cardiac spatial image. Dynamic generation of enhanced feature tensors from the space-frequency domain:

[0153] ;

[0154] in, Represents the spatial coordinates of the neighborhood. Represents the magnetic field gradient sensitivity coefficient. This is the local normalization factor.

[0155] The initial lesion saliency probability value and the saliency probability value of neighboring lesions are jointly corrected at the pixel level using the spatial-frequency domain consistency constraint weight value to generate the corrected lesion saliency probability value.

[0156] In Example 1, the initial lesion significance probability value at spatial coordinates (x, y) is multiplied by the spatial-frequency domain consistency constraint weight value to obtain the original significance contribution value; the neighborhood lesion significance probability value is multiplied by the result of constant 1 minus the spatial-frequency domain consistency constraint weight value to obtain the neighborhood correction contribution value; the original significance contribution value and the neighborhood correction contribution value are added to obtain the corrected lesion significance probability value at spatial coordinates (x, y).

[0157] The corrected lesion saliency probability value is used to achieve continuous adaptive smooth control. When the consistency between the spatial domain magnetic field response and the frequency domain enhanced spatial magnetic field response is stronger than the threshold, the original lesion saliency judgment is given a higher retention weight. When the consistency between the two is weaker than the threshold, the neighboring lesion saliency probability value is given a higher correction weight, thereby achieving pixel-level dynamic flexible correction of the lesion saliency judgment at the current spatial coordinate position.

[0158] The corrected lesion saliency probability values ​​at all spatial coordinates are grouped and organized according to the image frame number to construct a corrected lesion saliency probability distribution map.

[0159] This implementation method constructs a spatial-frequency domain consistency deviation value as a flexible confidence gating and combines it with anisotropic neighborhood weights dynamically driven by local spatial magnetic field gradients to perform joint pixel-level correction on the initial lesion probability map output by the network. This breaks through the technical limitations of blind smoothing or reliance on hard thresholds in traditional medical image post-processing. It uses the multidimensional differences of the physical field at the input end to dynamically self-evaluate the reliability of the prediction. It can not only effectively filter out false lesion artifacts caused by complex cardiac magnetic background disturbances (adaptively introduce neighborhood correction in areas with weak consistency), but also rapidly attenuate the smoothing force when crossing the steep boundary of the cardiac magnetic dipole. This achieves edge-preserving smoothing of the true physical boundary of extremely weak ischemic lesions in coronary heart disease, and outputs a cardiac magnetic lesion saliency heatmap with accurate boundaries, extremely high confidence, and strong physical interpretability for clinical diagnosis.

[0160] The lesion saliency probability distribution map after correction is segmented into regions according to a preset saliency threshold to obtain a lesion region mask in the magnetocardiogram image;

[0161] In this embodiment, the corrected lesion significance probability distribution map is segmented into regions according to a preset significance threshold, including:

[0162] For the corrected lesion significance probability distribution map, the corrected lesion significance probability value is read at the spatial coordinate (x,y), and the corrected lesion significance probability value is compared with the preset significance threshold to obtain the lesion region judgment value at the spatial coordinate (x,y).

[0163] In Example 1, the lesion region determination value of the baseline-calibrated magnetic field image of the qth frame at spatial coordinates (x, y) is obtained by comparing the corrected lesion significance probability value of the baseline-calibrated magnetic field image of the qth frame at spatial coordinates (x, y) with a preset significance threshold.

[0164] When the corrected lesion significance probability value is greater than or equal to the preset significance threshold, the lesion region judgment value at the current spatial coordinates (x,y) is set to 1;

[0165] When the corrected lesion significance probability value is less than the preset significance threshold, the lesion region determination value at the current spatial coordinates (x,y) is set to 0;

[0166] The lesion region determination value is used to indicate whether the current spatial coordinate position is determined to be a cardiac lesion region.

[0167] All lesion region judgment values ​​are organized into binary regions according to spatial coordinate distribution to obtain a lesion region mask for magnetocardiogram images.

[0168] The mask value of the lesion region mask in the magnetic resonance imaging at spatial coordinates (x, y) is equal to the lesion region determination value at spatial coordinates (x, y). When the mask value is equal to 1, it means that the current spatial coordinate position belongs to the magnetic resonance lesion region. When the mask value is equal to 0, it means that the current spatial coordinate position belongs to the non-lesion region.

[0169] The lesion region masks of the magnetic field images corresponding to all frame baseline-calibrated magnetic field spatial images are grouped and organized according to the image frame number to obtain the magnetic field image lesion region mask set.

[0170] Based on the lesion region mask of the magnetocardiogram image, the spatial location of the lesion, the boundary contour of the lesion, and the salience probability information of the lesion are reconstructed, and the magnetocardiogram image lesion identification result map for the diagnosis of ischemic areas of coronary heart disease is output.

[0171] In this embodiment, the spatial location, boundary contour, and salience probability information of the lesion are reconstructed based on the lesion region mask of the magnetocardiogram image, including:

[0172] Based on the lesion region mask of magnetocardiogram image, extract the set of spatial location coordinates of the lesion;

[0173] In Example 1, the mask value of the lesion region in the magnetocardiogram is read point by point within the entire spatial coordinate range. When the mask value of the lesion region in the magnetocardiogram at a certain spatial coordinate position is equal to 1, the corresponding spatial coordinate position is recorded as the spatial position coordinate of the lesion. All spatial coordinate positions that satisfy the mask value of the lesion region in the magnetocardiogram equal to 1 are set together according to the spatial coordinate form to obtain the set of spatial position coordinates of the lesion corresponding to the baseline-calibrated magnetocardiogram spatial image of the qth frame.

[0174] Calculate the center coordinates of the lesion's spatial location based on the set of lesion spatial location coordinates;

[0175] Based on the lesion region mask of magnetocardiogram image, the set of lesion boundary contour coordinates is extracted by morphological boundary extraction;

[0176] Based on the corrected lesion saliency probability distribution map and the lesion region mask of the magnetocardiogram image, the lesion saliency probability information is extracted;

[0177] In Example 1, when the mask value of the lesion region in the magnetocardiogram at a certain spatial coordinate position is equal to 1, the corrected lesion saliency probability value at the corresponding spatial coordinate position is recorded as lesion saliency probability information. All corrected lesion saliency probability values ​​that satisfy the mask value of the lesion region in the magnetocardiogram equal to 1 are aggregated and organized according to the spatial coordinate order to obtain the lesion saliency probability information set corresponding to the baseline-calibrated magnetocardiogram spatial image of the qth frame. The lesion saliency probability information set is used to represent the lesion saliency probability distribution at each spatial coordinate position within the ischemic area of ​​coronary heart disease.

[0178] Based on the set of spatial location coordinates of lesions, the set of boundary contour coordinates of lesions, and the saliency probability information of lesions, a magnetic resonance imaging lesion identification result map corresponding to the baseline-calibrated magnetic resonance spatial image is constructed.

[0179] In Example 1, for the baseline-calibrated magnetic field image of the qth frame, the set of spatial location coordinates of the lesion, the set of boundary contour coordinates of the lesion, and the saliency probability information of the lesion are read point by point within the entire spatial coordinate range;

[0180] When a certain spatial coordinate location belongs to the set of spatial coordinate locations of the lesion, the display intensity of the lesion area is assigned in the lesion identification result map of the magnetic heart image;

[0181] When a certain spatial coordinate position belongs to the set of lesion boundary contour coordinates, the lesion boundary contour display intensity is assigned in the lesion recognition result image of the magnetic heart image.

[0182] When the corrected lesion saliency probability value corresponding to a certain spatial coordinate position is greater than zero, the lesion saliency probability display intensity is assigned in the lesion identification result map of the magnetocardiogram image;

[0183] By performing joint mapping processing on the three display intensities, the lesion identification result map of the baseline-calibrated magnetic field image of the qth frame is obtained. The lesion identification result map of the magnetic field image is used to jointly represent the lesion region attribution, lesion boundary contour attribution, and lesion saliency probability intensity of the ischemic area of ​​coronary heart disease at the current spatial coordinate position.

[0184] Example 2: During the operation of a cardiac magnetic detection system in a medical testing institution, staff members conducted routine cardiovascular function tests on multiple subjects. The detection system uses a 64-channel cardiac magnetic sensor array to collect cardiac magnetic signals on the surface of the human chest cavity through a magnetically shielded environment and reconstruct the cardiac magnetic signals into a two-dimensional cardiac magnetic image. The system collects continuous cardiac magnetic signal data of a subject during a single test and automatically analyzes it using the method of this invention.

[0185] At the start of the test, the system acquired the subject's raw magnetocardiogram (MCC) signal data. The raw data included 120 seconds of continuous magnetic field signal recordings, sampled 1000 times per second, for a total of 120,000 sampling points. A two-dimensional MCC image sequence was generated using a spatial reconstruction algorithm, with approximately 110 MCC images corresponding to each cardiac cycle, each image being 128×128 pixels in size. During the data preprocessing stage, the system detected a significant abnormal magnetic field region in one set of MCC images. For example, in one frame, the magnetic field strength near spatial coordinates (54, 72) reached 3.8 × 10⁻¹² T, while the average magnetic field strength in normal areas was 1.1 × 10⁻¹² T. This abnormal signal was recorded as a suspected lesion magnetic field response.

[0186] The system performs baseline calibration on the raw magnetocardiogram (MCC) images. During processing, the system detected an ambient magnetic noise with a mean value of 0.6 × 10⁻¹² T. After filtering using a magnetic noise suppression algorithm, the calibrated baseline magnetic field fluctuation range was reduced to ±0.2 × 10⁻¹² T. After calibration, the system generates a baseline-calibrated MCC spatial image dataset. In this dataset, the system recorded a magnetic field strength of 3.5 × 10⁻¹² T at spatial coordinates (52, 74) and 3.6 × 10⁻¹² T at spatial coordinates (55, 75) in the 38th frame image, both of which showed significant anomalies.

[0187] The system then performed a two-dimensional frequency domain transformation on the baseline-calibrated magnetocardiogram spatial image dataset. During the frequency domain analysis, the system statistically analyzed the spectral energy distribution and recorded that the spectral energy at the frequency domain coordinates (22, 19) reached 2.4 × 10⁻²³, while the average spectral energy in the normal region was 6.7 × 10⁻². 4 The system determines that there is an abnormal high-frequency energy response in the frequency domain region and automatically marks the region as a candidate region for high-frequency enhancement.

[0188] The system continued to perform high-frequency enhancement processing on the frequency domain data. During the enhancement process, the high-frequency enhancement coefficient was set to 2.1, and the system detected that the enhanced frequency domain energy value increased to 5.0 × 10⁻²³. The system then performed an inverse frequency domain transformation on the enhanced frequency domain spectrum image to obtain a frequency-enhanced magnetic field spatial image. At spatial coordinates (54, 73), the magnetic field response increased from the original 3.5 × 10⁻¹²T to 4.1 × 10⁻¹²T, while the average value of the normal area remained around 1.2 × 10⁻¹²T, indicating a significant improvement in signal contrast in the lesion area.

[0189] The system fuses the frequency-domain enhanced magnetic field image with the baseline-calibrated magnetic field image along the channel dimension to construct a joint spatial-frequency domain enhancement feature tensor. In this feature tensor, the system records a spatial domain magnetic field value of 3.6 × 10⁻¹² T and a frequency-domain enhanced magnetic field value of 4.0 × 10⁻¹² T at spatial coordinates (53, 74), while the corresponding values ​​in the normal region are 1.2 × 10⁻¹² T and 1.3 × 10⁻¹² T, respectively.

[0190] After entering the network analysis phase, the system inputs the spatial-frequency domain joint enhancement feature tensor into the improved TTT-UNet network structure. The system generates a shallow joint response feature map during the encoder phase. During the computation, the system detects that the shallow joint response feature value at spatial coordinates (54, 74) reaches 2.83, while the average value in the normal region is 0.91. The system marks this feature as a potential anomalous region feature.

[0191] The system calculates the patient-level baseline drift representation vector. During the statistical process, the system detected that the average baseline magnetic field of the chest region of the examined subject was 1.05 × 10⁻¹²T, while the average value of the training samples was 1.18 × 10⁻¹²T. Therefore, the system automatically generated a channel scaling parameter of 0.92 and a channel translation parameter of 0.07. Through this modulation mechanism, the encoder hidden layer features are dynamically adjusted to compensate for baseline differences between examined subjects.

[0192] During the inference phase, the system adaptively fine-tunes the module when testing is triggered and constructs a self-supervised mask reconstruction task. In one random masking process, the system masks the joint feature values ​​of the region with spatial coordinates (50,72) to (60,78) and requires the network to reconstruct the features of this region. After three iterations, the self-supervised loss decreases from 0.084 to 0.019, and the network parameters gradually adapt to the current distribution of the tested object data.

[0193] After encoding and fine-tuning, the system generates a candidate feature map of the lesion using a decoder. In this feature map, the system detects a candidate feature value of 4.12 for the lesion at spatial coordinates (54, 73), while the average value for the surrounding area is 0.88. The system then calculates the lesion significance probability based on the candidate feature map. The calculation results show that the significance probability at spatial coordinates (54, 73) reaches 0.96, while the probabilities for normal areas are all less than 0.25.

[0194] The system utilizes spatial-frequency domain consistency constraints to perform pixel-level correction of saliency probabilities. When calculating consistency bias, the system detected a spatial domain magnetic field value of 3.6 × 10⁻¹² T and a frequency-domain enhanced magnetic field value of 4.0 × 10⁻¹² T at spatial coordinates (54, 73), with a difference of 0.4 × 10⁻¹² T, corresponding to a consistency weight of 0.92. After correction, the saliency probability at this location was adjusted from 0.96 to 0.94, while the probability in some noisy regions decreased from 0.41 to 0.18.

[0195] During the region segmentation stage, the system sets a saliency threshold of 0.55. Regions with a probability greater than this threshold are identified as lesion areas. The system ultimately identifies a continuous lesion region between spatial coordinates (52,72) and (57,76), with a total area of ​​approximately 23 pixels. Subsequently, the system extracts the lesion center location as (54,74) and generates the lesion boundary contour.

[0196] The system ultimately generated a magnetocardiogram (MCC) image showing the identified lesion, and marked the area as a suspected area of ​​myocardial ischemia in the report. Based on the system report, staff conducted further clinical examinations on the subject, and the results were consistent with the MCC identification.

[0197] To verify the effectiveness of the method of the present invention, the system conducted comparative experiments on the same dataset using the traditional UNet method, the traditional frequency domain enhancement method, and the method of the present invention. The experimental data are shown in Table 1 below, where the number of training samples is 260, the number of test samples is 80, and the total number of magnetocardiogram image frames is 8800.

[0198] Table 1 Statistical analysis of experimental results

[0199] method Lesion identification accuracy Dice coefficient Average positioning error Traditional UNet 82.4% 0.80 7.6mm Frequency Domain Enhancement UNet 86.1% 0.84 6.3mm Method of the present invention 93.7% 0.92 3.5mm

[0200] Experimental results show that the method of the present invention is significantly superior to traditional methods in terms of lesion identification accuracy, localization accuracy, and early lesion detection capability.

[0201] 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 detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification, characterized in that, include: Raw magnetic field image data of the chest cavity surface of the subject were acquired and preprocessed to obtain a baseline-calibrated magnetic field spatial image data set; A two-dimensional frequency domain transformation is performed on the baseline calibration magnetic-cardiogram spatial image dataset to generate the corresponding magnetic-cardiogram frequency domain spectral image dataset. The spectral energy distribution of the magnetocardiogram frequency domain spectral image dataset is calculated, and spectral weights are applied to the high-frequency region to highlight high-frequency abnormal energy, resulting in a high-frequency enhanced frequency domain spectral image dataset. An inverse frequency domain transformation is performed on the high-frequency enhanced frequency domain spectral image dataset to obtain a frequency domain enhanced magnetocardiogram spatial image dataset, which is then aligned and fused with the baseline calibrated magnetocardiogram spatial image dataset in the channel dimension to construct a joint spatial-frequency domain enhancement feature tensor. The spatial-frequency domain joint enhanced feature tensor is input into the encoder of the improved TTT-UNet network structure. During the inference phase, when the test is triggered, the adaptive fine-tuning module is used to construct the self-supervised auxiliary task loss using the spatial-frequency domain joint enhanced feature tensor. The hidden layer weights of the encoder are updated in real time through backpropagation. The encoded feature map after patient-level adaptive fine-tuning is then input into the decoder of the improved TTT-UNet network structure to perform layer-by-layer upsampling and cross-layer feature fusion to obtain the lesion candidate feature map. The saliency probability distribution map of lesions is calculated based on the lesion candidate feature map, and the saliency probability distribution map of lesions is corrected at the pixel level to generate the corrected saliency probability distribution map of lesions. The lesion saliency probability distribution map after correction is segmented into regions according to a preset saliency threshold to obtain a lesion region mask in the magnetocardiogram image; Based on the lesion region mask of the magnetocardiogram image, the spatial location of the lesion, the boundary contour of the lesion, and the salience probability information of the lesion are reconstructed, and the magnetocardiogram image lesion identification result map for the diagnosis of ischemic areas of coronary heart disease is output.

2. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The preprocessing includes time synchronization correction, environmental magnetic noise suppression, and individual baseline stabilization.

3. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The two-dimensional frequency domain transformation of the baseline-calibrated magnetocardiogram spatial image data set includes: A two-dimensional discrete frequency domain mapping relationship is established for each frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset to obtain the corresponding complex magnetocardiogram frequency domain spectrum image. For the complex magnetic field frequency domain spectrum image corresponding to the baseline-calibrated magnetic field spatial image of the qth frame, the complex magnetic field frequency domain spectrum value at each frequency domain coordinate is decomposed into real part spectrum value and imaginary part spectrum value; For each frequency domain coordinate position, calculate the corresponding amplitude spectrum value and phase spectrum value for the real part and imaginary part spectrum values; Perform a frequency domain centering rearrangement operation on the amplitude spectrum values ​​to obtain the centered amplitude spectrum values; The centered amplitude spectrum value and the corresponding phase spectrum value are jointly represented to obtain the frequency domain characterization result of the baseline-calibrated magnetocardiogram spatial image of the qth frame. The frequency domain characterization results of all frames of baseline-calibrated magnetocardiogram spatial images are then organized into a set according to the image frame number to construct a magnetocardiogram frequency domain spectrum image data set.

4. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The calculation of spectral energy distribution for the magnetocardiogram frequency domain spectral image dataset, and the application of spectral weights in the high-frequency region to highlight high-frequency anomalous energy, includes: Based on the frequency domain characterization results in the magnetocardiogram frequency domain spectral image dataset, a spectral energy distribution is established based on the centered amplitude spectral image to obtain the spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image. A frequency-domain radial distance distribution is established based on the frequency-domain center position of the centered amplitude spectrum image, and the frequency-domain radial distance value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained. Based on the frequency domain radial distance value, the high-frequency region determination result is constructed, and the high-frequency region mask value corresponding to the baseline calibrated magnetocardiogram spatial image is obtained. Spectral weights are applied to the frequency domain coordinates where the mask value in the high-frequency region is equal to 1, to obtain the high-frequency spectral weight values ​​corresponding to the baseline-calibrated magnetocardiogram spatial image. The spectral energy value is enhanced by weighting the spectral energy value with high-frequency spectral weights to obtain the high-frequency enhanced spectral energy value corresponding to the baseline calibrated magnetocardiogram spatial image. Based on the high-frequency enhanced spectral energy value, the corresponding high-frequency enhanced amplitude spectral value is recovered and jointly represented with the phase spectral value to obtain the high-frequency enhanced frequency domain spectral image corresponding to the baseline calibrated magnetocardiogram spatial image; All frames of high-frequency enhanced frequency domain spectral images are organized into a set according to the image frame number to construct a high-frequency enhanced frequency domain spectral image data set.

5. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The inverse frequency domain transform of the high-frequency enhanced frequency domain spectral image data set includes: For each frame of high-frequency enhanced frequency domain spectrum image in the high-frequency enhanced frequency domain spectrum image dataset, the complex high-frequency enhanced frequency domain spectrum value at the corresponding frequency domain coordinate position is reconstructed based on the high-frequency enhanced amplitude spectrum value and phase spectrum value. A two-dimensional inverse discrete frequency domain transform is performed on the complex high-frequency enhanced frequency domain spectral values ​​to obtain the corresponding frequency domain enhanced magnetocardiogram spatial image; The real part of the inverse frequency domain transformation result is taken from the frequency domain enhanced magnetocardiogram spatial image to obtain the real-valued frequency domain enhanced magnetocardiogram spatial image corresponding to the baseline calibrated magnetocardiogram spatial image; All real-value frequency-domain enhanced magnetocardiogram spatial images are grouped and organized according to the image frame number to construct a frequency-domain enhanced magnetocardiogram spatial image dataset; For each frame of real-valued frequency-domain enhanced magnetocardiogram spatial image in the frequency-domain enhanced magnetocardiogram spatial image dataset and the corresponding frame of baseline-calibrated magnetocardiogram spatial image in the baseline-calibrated magnetocardiogram spatial image dataset, a one-to-one spatial coordinate correspondence channel alignment relationship is established to obtain a real-valued frequency-domain enhanced magnetocardiogram spatial image with complete channel alignment. The channel-aligned real-valued frequency-domain enhanced magnetocardiogram spatial image and the baseline-calibrated magnetocardiogram spatial image are fused in the channel dimension to obtain the spatial-frequency joint enhanced feature tensor of the corresponding frame; All frame spatial-frequency domain joint enhancement feature tensors are organized into sets according to the image frame number to construct a set of spatial-frequency domain joint enhancement feature tensors.

6. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The process involves inputting a spatial-frequency domain joint enhanced feature tensor into the encoder of the improved TTT-UNet network structure. During the inference phase, an adaptive fine-tuning module is triggered to construct a self-supervised auxiliary task loss using the spatial-frequency domain joint enhanced feature tensor. The hidden layer weights of the encoder are then updated in real-time via backpropagation. The resulting patient-level adaptively fine-tuned encoded feature map is then input into the decoder of the improved TTT-UNet network structure. This includes: The spatial-frequency domain joint enhancement feature tensor is input into the encoder of the improved TTT-UNet network structure, and a shallow joint response feature map is constructed based on the spatial domain channel and the frequency domain enhanced spatial channel. Based on the shallow joint response feature map, a patient-level baseline drift representation vector is established, and the channel scaling parameters and channel translation parameters of the encoder hidden layer are generated based on the patient-level baseline drift representation vector. Patient-level adaptive modulation of encoder hidden layer features is performed using channel scaling and channel translation parameters to obtain encoded feature maps containing patient-level baseline compensation information. When the test is triggered during the inference stage of the improved TTT-UNet network structure, the adaptive fine-tuning module is used to construct a mask reconstruction consistency self-supervised auxiliary task based on the spatial-frequency domain joint enhanced feature tensor, and the self-supervised task loss is obtained. The weight parameters of the encoder hidden layer and the weight parameters of the patient-level adaptive modulation parameters are updated in real time by backpropagation based on the self-supervised auxiliary task loss. The update is performed in a preset number of iterations or until the self-supervised auxiliary task loss value is less than the preset convergence threshold, so as to obtain the encoded feature map after patient-level adaptive fine-tuning. The encoded feature map of patient-level baseline compensation information and the encoded feature map after patient-level adaptive fine-tuning are input into the decoder of the improved TTT-UNet network structure, layer-by-layer upsampling is performed, and cross-layer feature fusion is performed using the encoded feature map of the corresponding layer of the encoder after patient-level adaptive modulation to obtain the decoded fused feature map. The lesion candidate response is mapped onto the decoded and fused feature map to obtain the lesion candidate feature map.

7. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The calculation of the lesion saliency probability distribution map based on the lesion candidate feature map, and the pixel-level correction of the lesion saliency probability distribution map, includes: Based on the candidate feature map of lesions, the initial lesion significance probability value is calculated; Based on the feature differences between the spatial domain channel and the frequency domain enhanced spatial channel at the same spatial coordinates in the spatial-frequency domain joint enhanced feature tensor, a spatial-frequency domain consistency deviation value is constructed. Based on the spatial-frequency domain consistency deviation value, construct the spatial-frequency domain consistency constraint weight value; Based on the initial lesion significance probability value, construct the significance probability value of neighboring lesions; The initial lesion saliency probability value and the saliency probability value of neighboring lesions are jointly corrected at the pixel level using the spatial-frequency domain consistency constraint weight value to generate the corrected lesion saliency probability value. The corrected lesion saliency probability values ​​at all spatial coordinates are grouped and organized according to the image frame number to construct a corrected lesion saliency probability distribution map.

8. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The step of segmenting the corrected lesion significance probability distribution map into regions according to a preset significance threshold includes: For the corrected lesion significance probability distribution map, the corrected lesion significance probability value is read at the spatial coordinate (x,y), and the corrected lesion significance probability value is compared with the preset significance threshold to obtain the lesion region judgment value at the spatial coordinate (x,y). All lesion region judgment values ​​are organized into binary regions according to spatial coordinate distribution to obtain a lesion region mask for magnetocardiogram images. The mask value of the lesion region mask in the magnetic resonance imaging at spatial coordinates (x, y) is equal to the lesion region determination value at spatial coordinates (x, y). When the mask value is equal to 1, it means that the current spatial coordinate position belongs to the magnetic resonance lesion region. When the mask value is equal to 0, it means that the current spatial coordinate position belongs to the non-lesion region.

9. The method for detecting coronary heart disease based on magnetic resonance imaging analysis and lesion region identification according to claim 1, characterized in that, The process of reconstructing the spatial location, boundary contour, and salience probability information of the lesion based on the lesion region mask of the magnetocardiogram includes: Based on the lesion region mask of magnetocardiogram image, extract the set of spatial location coordinates of the lesion; Calculate the center coordinates of the lesion's spatial location based on the set of lesion spatial location coordinates; Based on the lesion region mask of magnetocardiogram image, the set of lesion boundary contour coordinates is extracted by morphological boundary extraction; Based on the corrected lesion saliency probability distribution map and the lesion region mask of the magnetocardiogram image, the lesion saliency probability information is extracted; Based on the set of spatial location coordinates of lesions, the set of boundary contour coordinates of lesions, and the saliency probability information of lesions, a magnetic resonance imaging lesion identification result map corresponding to the baseline-calibrated magnetic resonance spatial image is constructed.

Images