Method for auxiliary diagnosis of migraine and evaluation of curative effect of magnetic resonance meningeal lymphatic parameters based on pattern recognition

Abstract

This invention discloses a method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition and magnetic resonance imaging (MRI) meningeal lymphatic vessel parameters. It relates to the field of medical imaging diagnostic technology, including the acquisition of high-resolution MRI images of the brain from multiple sequences of subjects, and simultaneous collection of clinical assessment information such as headache diaries, symptom scores, and medication records. An improved region growing algorithm is used to segment and parameterize meningeal lymphatic vessels, extracting corresponding morphological and functional quantitative feature sets. The clinical assessment information is then temporally structured and encoded to generate a clinical temporal feature vector. These two types of features are then fed into a multimodal pattern recognition model for joint computation to obtain the migraine diagnosis classification probability and efficacy change trajectory, thereby outputting standardized auxiliary diagnostic opinions and efficacy evaluation reports. This method achieves refined image feature extraction and fusion computation of clinical temporal data, optimizing the data-driven analysis of migraine diagnosis and treatment.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging diagnostic technology, specifically a method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters. Background Technology

[0002] Currently, clinical diagnosis of migraines largely relies on subjective questioning by healthcare professionals and symptom scales completed by patients. Brain MRI images are only used to screen for organic brain lesions, and the industry does not routinely include imaging features related to meningeal lymphatic vessels in its analysis. Traditional brain image segmentation generally uses standard region growing algorithms to divide tissue regions. Clinically collected headache-related written records, symptom scores, and daily medication records have long been archived in a raw and fragmented form, used only as offline diagnostic references, and have not been incorporated into intelligent data analysis processes.

[0003] Standard region growing algorithms, with their fixed decision conditions, cannot adapt to the delicate and dispersed structure of meningeal lymphatic vessels. This leads to issues such as missed regions and misaligned boundaries during image segmentation, and makes it impossible to extract quantitative parameters reflecting the actual state of the meningeal lymphatic vessels in batches. Furthermore, the image analysis system is disconnected from the clinical diagnostic information system. Relying solely on image features only allows for basic structural observation and cannot provide in-depth analysis considering the patient's disease progression. Conventional intelligent analysis modes only output fixed symptom judgments and cannot generate relevant content reflecting disease progression trends. The lack of dynamic state assessment during diagnosis and treatment results in a relatively thin overall analytical dimension. Therefore, it is necessary to adjust the image segmentation operation logic, standardize the conversion of fragmented clinical information, and establish a multi-type data fusion analysis path. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters, including:

[0006] Acquire a high-resolution magnetic resonance imaging dataset of the subjects, the high-resolution magnetic resonance imaging dataset including multi-sequence brain images and clinical assessment information acquired simultaneously with the images, the clinical assessment information including headache diaries, symptom scores and medication records;

[0007] The multi-sequence brain images are subjected to meningeal lymphatic vessel segmentation and parameterization processing to extract a quantitative feature set reflecting the morphology and function of the meningeal lymphatic vessels. The meningeal lymphatic vessel segmentation and parameterization processing is implemented based on an improved region growing algorithm.

[0008] The clinical assessment information is time-series structured encoded to generate a clinical time-series feature vector;

[0009] The quantitative feature set and the clinical time-series feature vector are input into a multimodal pattern recognition model for joint analysis to generate classification probabilities for auxiliary diagnosis of migraine and efficacy change trajectories for efficacy evaluation.

[0010] Based on the classification probability and the trajectory of efficacy changes, a structured auxiliary diagnostic opinion and efficacy evaluation report are output.

[0011] Furthermore, the multi-sequence brain images are subjected to meningeal lymphatic vessel segmentation and parameterization processing to extract a quantitative feature set reflecting the morphology and function of the meningeal lymphatic vessels, including:

[0012] The multi-sequence brain images are preprocessed, including head motion correction, spatial normalization, and registration between different sequence images, to obtain multimodal image data in standard space;

[0013] In the multimodal image data under the standard space, based on the intensity and texture features of a specific sequence of images, the improved region growing algorithm is used to iteratively grow a complete three-dimensional structure of meningeal lymphatic vessels starting from a preset seed point. The improved region growing algorithm introduces prior local blood vessel orientation and anatomical structure constraints during the growth process.

[0014] For the segmented three-dimensional meningeal lymphatic vessel structure, three-dimensional morphological analysis and functional parameter calculation based on image intensity are performed to extract multiple quantitative features, including total lymphatic vessel volume, surface area, length distribution, branching complexity, average wall thickness, average signal intensity of the lumen, and signal intensity gradient along the vessel.

[0015] All extracted quantitative features are normalized and integrated to form the quantitative feature set.

[0016] Furthermore, the improved region growing algorithm incorporates prior knowledge of local blood vessel orientation and anatomical structure constraints during the growing process, including:

[0017] In each growth iteration of the improved region growing algorithm, the local orientation consistency between the candidate growth point and the already grown points in its neighborhood is calculated.

[0018] The expected vascular course range at the current location is obtained from the preset meningeal lymphatic anatomical atlas. Candidate points whose local course consistency is higher than a preset threshold and are within the expected vascular course range are given higher growth priority.

[0019] Simultaneously, information on the boundaries of anatomical structures that are prohibited from being crossed near the current position is obtained from the anatomical atlas, and candidate points located outside the boundaries of the anatomical structures are completely excluded during the growth process;

[0020] By combining the prior knowledge of local vascular orientation with anatomical constraints, the regional growth process is guided more precisely along the actual anatomical path of the meningeal lymphatic vessels.

[0021] Furthermore, in each growth iteration of the improved region growing algorithm, the local orientation consistency between the candidate growth point and its neighboring already grown points is calculated, including:

[0022] In the three-dimensional image space, a local spherical neighborhood is defined with the candidate growth point as the center;

[0023] Calculate the principal direction of the spatial point set formed by the positions of all grown points in the neighborhood;

[0024] Calculate the direction vector from the candidate growth point to the centroid of the spatial point set;

[0025] Calculate the cosine of the angle between the main direction and the direction vector, and use the cosine of the angle as a quantitative measure of the consistency of the local orientation. The closer the cosine value is to 1, the higher the consistency.

[0026] Furthermore, the preset meningeal lymphatic vessel anatomical atlas is constructed by probabilistic fusion based on the results of standard spatial meningeal lymphatic vessel segmentation from multiple healthy subjects.

[0027] Furthermore, the clinical assessment information is time-series structured encoded to generate a clinical time-series feature vector, including:

[0028] The headache diary was analyzed to extract the onset time, duration, pain intensity, and temporal sequence of accompanying symptoms of headache attacks.

[0029] The symptom scores were analyzed, and the score sequences of the standardized migraine symptom scale collected at different time points were extracted.

[0030] The medication records are analyzed to extract the time sequence of drug name, administration time, and dosage.

[0031] The temporal sequence of headache attacks, the score sequence of the symptom scale, and the temporal record of medication are time-aligned and unified onto the same time coordinate axis.

[0032] At each aligned time point, the headache attack status characteristics, symptom score characteristics, and medication status characteristics are combined into a multidimensional feature point, which is then arranged in chronological order to form the clinical time series feature vector.

[0033] Furthermore, the quantified feature set and the clinical time-series feature vector are input into a multimodal pattern recognition model for joint analysis to generate classification probabilities for auxiliary diagnosis of migraine and efficacy change trajectories for efficacy evaluation, including:

[0034] The multimodal pattern recognition model includes a static feature encoding branch, a temporal feature encoding branch, and a fusion decision branch;

[0035] The static feature encoding branch receives the quantized feature set and encodes it into a high-dimensional static feature representation through a multilayer perceptron.

[0036] The temporal feature encoding branch receives the clinical temporal feature vector, captures its temporal dependencies through a recurrent neural network, and encodes it into a temporal context feature representation.

[0037] The fusion decision branch receives the high-dimensional static feature representation and the temporal context feature representation, calculates their interaction information through a cross-attention mechanism, and generates a fusion feature representation.

[0038] The fused feature representations are respectively input into a parallel diagnostic classifier and a therapeutic trajectory regressor. The diagnostic classifier outputs the classification probability that the subject is classified as a migraine patient or a healthy control. The therapeutic trajectory regressor outputs a predicted trajectory characterizing the changes in meningeal lymphatic vessel parameters over time as the therapeutic change trajectory.

[0039] Further, the temporal feature encoding branch receives the clinical temporal feature vector and captures its temporal dependencies through a recurrent neural network, including:

[0040] The recurrent neural network adopts a gated recurrent unit structure to process each multidimensional feature point in the clinical time sequence feature vector arranged in chronological order in turn.

[0041] At each time step, the gated loop unit calculates the update gate and the reset gate based on the input features of the current time step and the hidden state of the previous time step.

[0042] The reset gate controls how much information about the hidden state of the previous time step is forgotten, and the update gate controls how much new information about the current time step is incorporated.

[0043] Through the gating mechanism consisting of the update gate and the reset gate, the recurrent neural network can capture the long-term temporal dependence and dynamic evolution pattern of clinical features under the influence of headache onset, relief, and medication events. Finally, the hidden state of the last time step is used as the temporal context feature representation.

[0044] Furthermore, the therapeutic trajectory regressor outputs a predicted trajectory characterizing the changes in meningeal lymphatic vessel parameters over time as the therapeutic effect change trajectory, including:

[0045] The therapeutic trajectory regressor is a neural network-based regression model. Its input is the fused feature representation, and its output is a parameter sequence. The parameter sequence defines a functional form describing the changes of key parameters of the meningeal lymphatic vessels over time.

[0046] The function form includes linear trend, exponential decay, or S-curve, with the specific form determined by the network weights.

[0047] During the training phase, subject data with multiple imaging follow-ups were used, with the baseline fusion feature representation as input and the actual measured changes in meningeal lymphatic parameters during the follow-up period as supervision signals, to train the therapeutic trajectory regressor so that it could predict future parameter change trends based on baseline information.

[0048] In the application phase, the fusion feature representation of the subject to be evaluated is input into the trained efficacy trajectory regressor to obtain the predicted efficacy change trajectory.

[0049] Furthermore, based on the aforementioned classification probabilities and efficacy change trajectories, structured auxiliary diagnostic opinions and efficacy evaluation reports are output, including:

[0050] The classification probability is compared with a preset diagnostic threshold. If the classification probability is greater than or equal to the diagnostic threshold, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated; otherwise, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated.

[0051] Analyze the morphology of the therapeutic effect change trajectory, calculate the slope, curvature or the time point when the trajectory tends to stabilize, and use it as a quantitative indicator of the dynamics of the therapeutic effect.

[0052] From the preset efficacy evaluation rule base, the corresponding efficacy level description text is matched according to the shape and quantitative indicators of the efficacy change trajectory;

[0053] The auxiliary diagnostic opinions, the quantitative indicators of the efficacy change trajectory, and the matching efficacy level description text are integrated according to the preset report template format to generate the structured auxiliary diagnostic opinions and efficacy evaluation report containing diagnostic conclusions and efficacy analysis sections.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] The internal computational logic of the region growing algorithm was optimized, and the triggering conditions and boundary convergence rules for region growing were redefined to adapt to the spatial distribution morphology of the meningeal lymphatic vessels. During the segmentation process of multi-sequence brain images obtained from image scanning, the complete distribution area of ​​the meningeal lymphatic vessels can be accurately located, and tissue edges and fine branch structures can be completely segmented. The quantitative features generated after parameterization cover multiple physiological representation dimensions, allowing for the complete presentation of subtle tissue state differences hidden in the original images. Multiple effective information within brain images can be extracted and decomposed, expanding the analytical depth of image data. The state differences of meningeal lymphatic vessels among different individuals can be intuitively reflected through quantitative data, breaking the current situation of one-sided feature extraction in traditional image segmentation methods.

[0056] Following a linear temporal arrangement rule, various clinical assessment information is coded in a temporal structure. Disordered medical records are then systematically transformed along a timeline, converting fragmented medical information into uniformly formatted feature vector data. The systematically generated clinical temporal features can be directly integrated with imaging quantitative features, enabling collaborative computation between these two types of data within the same framework. During this collaborative computation, disease category differentiation and dynamic progression of disease status can be simultaneously performed, resulting in temporally continuous state changes. Medical assessment is no longer limited to static judgments within a single time period; it can present the overall trend of disease fluctuations along a temporal trajectory, ensuring continuity and completeness in the assessment content. This broadens the scope of data analysis and overcomes the limitations of single-dimensional data analysis. Attached Figure Description

[0057] Figure 1 This is a flowchart of the migraine auxiliary diagnosis and efficacy evaluation method based on pattern recognition magnetic resonance meningolymphatic vessel parameters as described in this invention.

[0058] Figure 2 A flowchart for introducing local blood vessel orientation priors and anatomical structure constraints to improve the region growing algorithm;

[0059] Figure 3 A flowchart illustrating the operation of a multimodal pattern recognition model. Detailed Implementation

[0060] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] See Figure 1The specific implementation method of migraine auxiliary diagnosis and efficacy evaluation based on pattern recognition magnetic resonance meningolymphatic parameters is as follows: A high-resolution magnetic resonance imaging dataset of the subject is acquired. This dataset includes multi-sequence brain images and clinical assessment information acquired simultaneously with the images. The clinical assessment information includes a headache diary, symptom scores, and medication records. In practice, a 3.0T or higher field strength magnetic resonance scanner is used to acquire T1-weighted, T2-weighted, and magnetic resonance lymphangiography sequences of brain images. Simultaneously, at least one month's headache diary, daily headache impact test scores, and acute and preventative medication records of the subject are collected synchronously through an electronic medical record system or a dedicated mobile terminal.

[0062] Meningeal lymphatic vessel segmentation and parameterization are performed on multi-sequence brain images to extract a quantitative feature set reflecting the morphology and function of the meningeal lymphatic vessels. This segmentation and parameterization is based on an improved region growing algorithm. Specifically, head motion correction and brain tissue extraction are performed on the multi-sequence images. Using the improved region growing algorithm, starting from pre-labeled seed points in the meningeal region, three-dimensional segmentation is performed by combining image intensity thresholds and local texture consistency criteria. After obtaining the meningeal lymphatic vessel structure, parameters such as its total volume, average diameter, number of branch points, and signal intensity changes along the lymphatic vessel path are calculated.

[0063] The clinical assessment information is then temporally structured and encoded to generate a clinical temporal feature vector. Specifically, the onset date, duration, and pain level in the headache diary are converted into time-series data, symptom scores are aligned according to the collection date, and medication records are encoded as discrete event sequences. The quantified feature set and the clinical temporal feature vector are input into a multimodal pattern recognition model for joint analysis. This model employs a multi-branch neural network structure based on an attention mechanism. The static feature branch processes lymphatic parameters, while the temporal branch processes the clinical event sequence. After fusion, a classification probability for auxiliary diagnosis of migraine and a trajectory of efficacy changes for efficacy evaluation are generated. Based on the classification probability and the trajectory of efficacy changes, a structured auxiliary diagnostic opinion and efficacy evaluation report are output. The report includes the diagnostic conclusion, confidence level, predicted trends in lymphatic parameter changes, and corresponding efficacy levels.

[0064] In one embodiment of the present invention, when performing meningeal lymphatic vessel segmentation and parameterization on multi-sequence brain images and extracting a quantitative feature set reflecting the morphology and function of meningeal lymphatic vessels, the multi-sequence brain images are preprocessed. The preprocessing includes head motion correction, spatial standardization and registration between different sequence images to obtain multimodal image data in standard space.

[0065] The preprocessing steps specifically employ rigid body transformation to correct for head movement during the scanning process. A nonlinear registration algorithm maps individual images to a standard brain template space, and T1-weighted images are cross-registered with magnetic resonance lymphangiography sequences to ensure anatomical consistency. In multimodal image data within the standard space, based on the intensity and texture features of specific image sequences, an improved region growing algorithm iteratively grows complete three-dimensional structures of meningeal lymphatic vessels from a pre-defined seed point. This improved algorithm incorporates prior knowledge of local vascular orientation and anatomical constraints during the growth process. Specifically, a magnetic resonance lymphangiography sequence is selected as the growth reference image. In this sequence, meningeal lymphatic vessels exhibit high signal characteristics. The pre-defined seed point is located in the dura mater region near the superior sagittal sinus and transverse sinus. The growth algorithm, centered on the seed point, determines whether to include new voxels based on the signal intensity of neighboring voxels and the texture homogeneity index calculated from the local gray-level co-occurrence matrix.

[0066] For the segmented 3D meningeal lymphatic vessel structure, 3D morphological analysis and image intensity-based functional parameter calculations were performed to extract multiple quantitative features, including total lymphatic vessel volume, surface area, length distribution, branch complexity, average wall thickness, average lumen signal intensity, and signal intensity gradient along the vessel. Morphological analysis yielded the total volume by multiplying voxel counts by voxel volume, calculated the surface area using an isosurface reconstruction algorithm, and extracted the lengths of each branch using skeletonization, statistically analyzing their distribution. Branch complexity was characterized by the ratio of the number of branch points to the number of terminal points. In functional parameter calculations, the average wall thickness was calculated from the lymphatic vessel boundary inwards using a distance transform algorithm, the average lumen signal intensity was directly taken as the average voxel signal within the segmented region, and the signal intensity gradient along the vessel was obtained by discretizing the lymphatic vessel centerline and calculating the rate of signal change between adjacent sampling points. All extracted quantitative features were normalized and integrated to form a quantitative feature set. Normalization was performed using the z-score method based on a reference distribution from a healthy control population.

[0067] In practice, when performing meningeal lymphatic vessel segmentation and parameterization on multi-sequence brain images to extract a quantitative feature set reflecting the morphology and function of meningeal lymphatic vessels, the multi-sequence brain images are preprocessed. The preprocessing includes head motion correction, spatial standardization, and registration between different sequence images to obtain multimodal image data in a standard space. The preprocessing steps use rigid body transformation to correct the head motion displacement of a migraine subject. The individual images are mapped to the standard brain template space of the Montreal Neurological Institute through a nonlinear registration algorithm. The T1-weighted images are then mutually registered with the magnetic resonance lymphangiography sequence. The registered multi-sequence images are resampled to 1 mm isotropic voxels.

[0068] In multimodal image data in standard space, based on the intensity and texture features of magnetic resonance lymphangiography sequences, an improved region growing algorithm is used to iteratively grow complete three-dimensional structures of meningeal lymphatic vessels starting from a preset seed point. The improved region growing algorithm introduces prior knowledge of local vascular orientation and anatomical constraints during the growth process. Meningeal lymphatic vessels in magnetic resonance lymphangiography sequences exhibit high signal characteristics. The preset seed point for meningeal lymphatic vessels is located in the dura mater region near the superior sagittal sinus and transverse sinus. The growth algorithm determines whether to include new voxels based on the signal intensity of neighboring voxels and the texture homogeneity index calculated from the local gray-level co-occurrence matrix. The signal intensity threshold is set to 80% to 120% of the typical signal mean of meningeal lymphatic vessels in the magnetic resonance lymphangiography sequence image, and the threshold for the texture homogeneity index is set to 0.7.

[0069] Three-dimensional morphological analysis and image intensity-based functional parameter calculations were performed on the segmented three-dimensional meningeal lymphatic vessel structure. Multiple quantitative features were extracted, including total lymphatic vessel volume, surface area, length distribution, branch complexity, average wall thickness, average lumen signal intensity, and signal intensity gradient along the vessel. Morphological analysis yielded the total lymphatic vessel volume by multiplying the voxel count by the volume of each voxel. The surface area was calculated using an isosurface reconstruction algorithm. The lengths of each branch were extracted using skeletonization, and their distribution was statistically analyzed. Branch complexity was represented by the ratio of the number of branch points to the number of terminal points. In the functional parameter calculations, the average wall thickness was calculated from the boundary of the three-dimensional meningeal lymphatic vessel structure inwards using a distance transform algorithm. The average lumen signal intensity was taken as the mean signal of all voxels within the segmented region in the magnetic resonance lymphangiography sequence. The signal intensity gradient along the vessel was obtained by discretizing the meningeal lymphatic vessel centerline into equally spaced sampling points and calculating the rate of change of signal between adjacent sampling points. The formula for calculating the rate of change is:

[0070]

[0071] in: This represents the signal intensity gradient between the kth adjacent sampling points. This represents the average signal intensity in the lumen at the (k+1)th sampling point. This represents the average signal strength in the lumen at the k-th sampling point. This represents the Euclidean distance between the k-th sampling point and the (k+1)-th sampling point. All extracted quantitative features are normalized and integrated to form a quantitative feature set. Normalization is performed using the z-score method, corrected based on the quantitative feature distribution of a reference population consisting of 30 healthy control subjects. In some embodiments, the length distribution is processed by extracting all branch lengths and then calculating the median and interquartile range as two independent features included in the quantitative feature set. Optionally, when there is local signal inhomogeneity in the subject's magnetic resonance lymphangiography sequence, median filtering is performed on the voxel signals within the segmented region before extracting the average signal intensity of the lumen. The median filtering window size is set to 3×3×3 voxels. It can be understood that a branch point is defined as a voxel connecting three or more branches in the three-dimensional structural framework of the meningeal lymphatic vessels, and a terminal point is defined as a voxel connecting only one branch in the framework.

[0072] In one embodiment of the present invention, the improved region growing algorithm introduces prior knowledge of local blood vessel orientation and anatomical structure constraints during the growing process, see reference. Figure 2 In each iteration of the improved region growing algorithm, the local orientation consistency between the candidate growth point and its neighboring already grown points is calculated. In the 3D image space, a local spherical neighborhood is defined centered on the candidate growth point, with the neighborhood radius set to 1.5 to 2.5 mm based on the image resolution and expected lymphatic vessel diameter. The principal direction of the spatial point set formed by the positions of all grown points within this neighborhood is calculated. The principal direction is obtained by performing eigenvalue decomposition on the covariance matrix of the spatial point set and taking the eigenvector corresponding to the largest eigenvalue.

[0073] The direction vector from the candidate growth point to the centroid of the spatial point set is calculated. The centroid is obtained by the arithmetic mean of the coordinates of the already grown points in the neighborhood. The cosine of the angle between the principal direction and the direction vector is calculated and used as a quantitative measure of local orientation consistency. The closer the cosine value is to 1, the higher the consistency. The expected vascular orientation range of the current location is obtained from a preset meningeal lymphatic anatomical atlas. Candidate points with local orientation consistency higher than a preset threshold and within the expected vascular orientation range are given higher growth priority. In specific implementation, the cosine value threshold is set to be no less than 0.85, and the expected vascular orientation range is defined as having an angle of no more than 30 degrees with the lymphatic vessel orientation at the corresponding location in the atlas. At the same time, the boundary information of anatomical structures that are prohibited from being crossed near the current location is obtained from the atlas. Candidate points located outside the boundaries of anatomical structures are completely excluded during the growth process. This boundary information includes the walls of large vessels, sinus venosities, brain parenchyma boundaries, and bony structures. By combining the prior knowledge of local vascular orientation with anatomical constraints, the regional growth process is guided to more accurately follow the actual anatomical path of the meningeal lymphatic vessels. The pre-defined meningeal lymphatic vessel anatomical atlas is constructed by probabilistic fusion based on the meningeal lymphatic vessel segmentation results in standard space from multiple healthy subjects. Specifically, high-resolution meningeal lymphatic vessel images of more than 30 healthy volunteers were collected, and fine segmentation was manually corrected one by one in standard space. After superimposing the segmentation results, the probability of meningeal lymphatic vessels appearing at each voxel location was calculated. Voxels with a probability greater than 50% constitute the basic structure of the atlas, and the local orientation distribution of each location was further extracted.

[0074] In practice, the improved region growing algorithm incorporates prior knowledge of local vascular orientation and anatomical constraints during the growth process. In each growth iteration of the improved region growing algorithm, the consistency of the local orientation of the candidate growth point with that of the already grown points in its neighborhood is calculated. In a specific implementation, for a standard spatial magnetic resonance lymphangiography sequence of a migraine subject, the current candidate growth point is located in the dura mater region adjacent to the superior sagittal sinus. A local spherical neighborhood is defined in the three-dimensional image space with the candidate growth point as the center, and the radius of the local spherical neighborhood is set to 2 mm.

[0075] The principal direction of the spatial point set, which comprises the positions of all grown points within the local spherical neighborhood, is calculated. This principal direction is obtained by performing eigenvalue decomposition on the covariance matrix of the spatial point set and taking the eigenvector corresponding to the largest eigenvalue. The direction vector from the candidate grown point to the centroid of the spatial point set is calculated. The centroid is obtained by the arithmetic mean of the coordinates of the grown points within the local spherical neighborhood. The cosine of the angle between the principal direction and the direction vector is calculated. This cosine is used as a quantitative measure of local orientation consistency; a cosine value closer to 1 indicates higher consistency. The formula for calculating the cosine is:

[0076]

[0077] in: This represents the cosine of the angle between the principal direction and the direction vector. This represents the direction vector from the candidate growth point to the centroid of the point set in space. It represents the principal direction vector of the set of all grown points in the local spherical neighborhood.

[0078] In some embodiments, the expected vascular orientation range at the current location is obtained from a preset meningeal lymphatic anatomical atlas. Candidate points with local orientation consistency higher than a preset threshold and within the expected vascular orientation range are given higher growth priority. The preset threshold is set to a cosine value of the angle greater than or equal to 0.85. The expected vascular orientation range is defined as the angle between the vascular orientation and the corresponding position in the meningeal lymphatic anatomical atlas not exceeding 30 degrees. The meningeal lymphatic anatomical atlas stores an expected orientation direction vector and its allowed deviation range for each spatial location. Optionally, information on anatomical structures that are prohibited from being crossed near the current location is also obtained from the meningeal lymphatic anatomical atlas. During the growth process, candidate points located outside the anatomical structure boundaries are completely excluded. The information on anatomical structures that are prohibited from being crossed includes the boundaries of large vessel walls, venous sinus walls, brain parenchyma, and bony structures. This boundary information is stored as a binary mask in the meningeal lymphatic anatomical atlas, with regions having a mask value of 1 indicating prohibited growth.

[0079] It is understandable that by combining prior knowledge of local vascular orientation with anatomical constraints, the region growth process is guided to more accurately follow the actual anatomical path of the meningeal lymphatic vessels. In each growth iteration, candidate points with high local orientation consistency and within the expected vascular orientation range are preferentially included in the growth region, while any candidate points located outside the boundaries of prohibited anatomical structures are rejected. In some embodiments, the preset meningeal lymphatic vessel anatomical atlas is constructed by probabilistic fusion based on the meningeal lymphatic vessel segmentation results in standard space of multiple healthy subjects. Specifically, high-resolution meningeal lymphatic vessel images of 40 healthy volunteers are acquired, and fine segmentation with manual correction is performed one by one in standard space. The segmentation results are superimposed, and the probability of meningeal lymphatic vessels appearing at each voxel location is calculated. Voxels with a probability greater than 50% constitute the basic structure of the meningeal lymphatic vessel anatomical atlas, and the local orientation distribution and the boundary information of prohibited anatomical structures at each location are further extracted from the probabilistic fusion results.

[0080] In practice, for candidate growth points adjacent to the superior sagittal sinus, the meningeal lymphatic anatomy atlas shows that the expected vascular orientation is inclined at approximately 15 degrees from medial to lateral, with an allowable deviation range of ±10 degrees. The cosine of the angle between the principal direction of the set of already grown points within the local spherical neighborhood and the direction vector pointing from the candidate growth point to the centroid is calculated to be 0.92, which is higher than the preset threshold of 0.85 but within the allowable deviation range. Therefore, this candidate growth point is given a high growth priority. Optionally, for candidate growth points located outside the boundary of the venous sinus wall, even if their local orientation consistency is higher than the preset threshold, the venous sinus wall boundary mask value in the meningeal lymphatic anatomy atlas is 1, and the candidate point is completely excluded. It can be understood that the radius of the local spherical neighborhood can be adjusted to 2.5 mm when the image resolution is low to ensure that the neighborhood contains a sufficient number of already grown points for principal direction calculation. When the number of already grown points in the neighborhood is less than 5, local orientation consistency is not calculated, and the judgment is directly based on the intensity threshold.

[0081] In one embodiment of the present invention, when generating clinical temporal feature vectors by temporal structured encoding of clinical assessment information, headache diaries are parsed to extract the temporal sequence of headache onset time, duration, pain intensity, and accompanying symptoms. Headache diaries are typically recorded in electronic form. Each record is parsed using natural language processing or structured query analysis to extract the date and timestamp, headache start and end times, pain intensity value using a 0-10 numerical rating scale, and the presence or absence of accompanying symptoms such as nausea, photophobia, and phonophobia. Symptom scores are then analyzed to extract score sequences from standardized migraine symptom scales collected at different time points. These scales include migraine disability assessment scales or headache impact test scales, with each assessment record including the corresponding collection date, total score, and scores for each sub-item.

[0082] Medication records were analyzed to extract time-series data on drug names, administration times, and dosages. These records included acute-phase treatments such as triptans and nonsteroidal anti-inflammatory drugs (NSAIDs), as well as prophylactic treatments such as beta-blockers or antiepileptic drugs. The time-series records of headache attacks, symptom scale scores, and medication administration were time-aligned onto a single time axis, with the calendar day as the smallest unit. Multiple headache attacks recorded within a single day were sorted by the onset time. For assessment data from multiple consecutive days, linear interpolation was used to fill in missing time points. At each aligned time point, headache attack status characteristics, symptom score characteristics, and medication status characteristics were combined into a multidimensional feature point. Headache attack status characteristics included whether an attack occurred that day, the number of attacks, and the cumulative pain intensity score. Symptom score characteristics directly used the raw scores from each scale. Medication status characteristics included whether medication was used that day, the medication category code, and the standardized dose value. All multidimensional feature points are arranged in chronological order to form a clinical time-series feature vector. The length of this vector is determined based on the available duration of clinical records, typically ranging from 30 to 90 days.

[0083] In practice, when generating clinical temporal feature vectors by temporal structured encoding of clinical assessment information, the headache diary of a migraine subject was analyzed to extract the temporal sequence of headache onset time, duration, pain intensity, and accompanying symptoms. The headache diary was recorded in spreadsheet format, with each entry including a date and timestamp (format: year-month-day-hour-minute), headache start time, headache end time, pain intensity value using a 0-10 numerical rating scale, and presence or absence of accompanying symptoms (nausea marked as 1 or 0, photophobia marked as 1 or 0, phonophobia marked as 1 or 0). For headache attacks lasting more than 24 hours, the end time was recorded as the same time the following day. Symptom scores were analyzed to extract the score sequence of the standardized migraine symptom scale collected at different time points. The standardized migraine symptom scale used was the Migraine Disability Assessment Scale. Each assessment record included the collection date and total score (range: 0-270). The subject underwent three assessments: at baseline, on day 14, and on day 28.

[0084] Medication records were analyzed to extract time-series records of drug names, administration times, and dosages. These records included acute-phase treatment drugs (rizatriptan, 10 mg per dose) and prophylactic treatment drugs (propranolol, 40 mg daily). Each record contained the date and time of administration, along with the corresponding drug name and dosage. The time-series records of headache attacks, symptom scale scores, and medication administration were time-aligned and unified onto the same time axis. The time axis used the calendar day as the smallest unit, with each day from 00:00 to 24:00 as a time point. For multiple headache attacks recorded within a day, they were sorted according to the onset time and mapped to multiple sub-time points on the corresponding date. These sub-time points were discretized in hours. For assessment data from multiple consecutive days, linear interpolation was used to supplement missing time point scores. The interpolation formula is as follows:

[0085]

[0086] in: This represents the symptom scale score at the time point t to be interpolated. and These represent two adjacent time points with known scores. Indicates a point in time Symptom scale scores at the location. Indicates a point in time Symptom scale scores at each aligned time point. In some embodiments, at each aligned time point, headache attack status characteristics, symptom score characteristics, and medication status characteristics are combined into a multidimensional feature point. Headache attack status characteristics include whether an attack occurred that day (1 for yes, 0 for no), number of attacks (cumulative number of attacks within a day), and cumulative pain intensity score (sum of pain intensity values ​​for each attack that day). Symptom score characteristics directly use the raw scores of the Migraine Disability Assessment Scale (scores after interpolation each day). Medication status characteristics include whether acute-phase treatment drugs were used that day (1 for yes, 0 for no), standardized value of acute-phase treatment drug dosage (actual dose divided by standard single dose of 10 mg), whether prophylactic treatment drugs were used that day (1 for yes, 0 for no), and standardized value of prophylactic treatment drug dosage (actual dose divided by standard daily dose of 40 mg).

[0087] All multidimensional feature points are arranged in chronological order to form a clinical time-series feature vector. The time length of this vector is determined to be 30 days based on the available clinical record duration, that is, 30 time points from day 1 to day 30. The multidimensional feature points at each time point have 8 dimensions (3 dimensions for headache attack state features, 1 dimension for symptom score features, and 4 dimensions for medication status features). Therefore, the clinical time-series feature vector is represented as a 30×8 matrix.

[0088] In specific implementation, for entries with incomplete headache onset time records, the default headache onset time is 8:00 AM on the day of recording; for entries with missing headache end time, the default duration is 6 hours. Optionally, when a subject records multiple accompanying symptoms simultaneously, the three binary features of nausea, photophobia, and phonophobia are merged into a three-dimensional vector as an accompanying symptom feature, and stored alongside the pain intensity in the headache attack status feature. It can be understood that when a medication name in a medication record cannot be mapped to a standardized drug library, the dosage standardization factor for that drug is preset to 1.0, and the medication record is marked as an undefined category in subsequent analysis. In some embodiments, for time periods with no headache attack records for several consecutive days, the "whether an attack occurred that day," the "number of attacks," and the "cumulative pain intensity score" in the headache attack status feature are set to 0, while the symptom score feature retains the interpolated score value, and the medication status feature is set according to the actual medication record. Optionally, when the symptom score collection time point completely overlaps with the headache attack time point, the original score is used directly without interpolation. It is understandable that the clinical time-series feature vector output as a numerical matrix can be directly used by the time-series feature encoding branch of the multimodal pattern recognition model.

[0089] In one embodiment of the present invention, see [reference] Figure 3The quantitative feature set and clinical temporal feature vectors are input into a multimodal pattern recognition model for joint analysis, generating classification probabilities for auxiliary diagnosis of migraine and efficacy change trajectories for efficacy evaluation. The multimodal pattern recognition model includes a static feature encoding branch, a temporal feature encoding branch, and a fusion decision branch. The static feature encoding branch receives the quantitative feature set and encodes it into a high-dimensional static feature representation using a multilayer perceptron. The multilayer perceptron consists of an input layer, two hidden layers, and an output layer. The number of neurons in each hidden layer is 128 and 64, respectively, using a modified linear unit as the activation function. The number of nodes in the input layer equals the number of features in the quantitative feature set, and the output layer obtains a 128-dimensional high-dimensional static feature representation.

[0090] The temporal feature encoding branch receives clinical temporal feature vectors and captures their temporal dependencies through a recurrent neural network (RNN), encoding them into a temporal context feature representation. This RNN employs a gated recurrent unit (ROU) structure, sequentially processing each multidimensional feature point arranged chronologically in the clinical temporal feature vector. At each time step, the gated RNU calculates update and reset gates based on the input features of the current time step and the hidden state of the previous time step. The reset gate controls how much information from the hidden state of the previous time step is forgotten, while the update gate controls how much new information from the current time step is incorporated. Through this gating mechanism of update and reset gates, the RNN can capture the long-term temporal dependencies and dynamic evolution patterns of clinical features under the influence of headache onset, relief, and medication events. Finally, the hidden state of the last time step is used as the temporal context feature representation, and the dimension of this hidden state is set to 128.

[0091] The fusion decision branch receives high-dimensional static feature representations and temporal context feature representations. It calculates the interaction information between the two through a cross-attention mechanism to generate a fusion feature representation. In the cross-attention mechanism, the high-dimensional static feature representation is used as the query, and the temporal context feature representation is used as the key and value. After calculating attention weights, a weighted sum is obtained to obtain a 128-dimensional fusion feature representation. This fusion feature representation is then input into a parallel diagnostic classifier and a treatment trajectory regressor. The diagnostic classifier outputs the probability of classifying the subject as a migraine patient or a healthy control. The diagnostic classifier uses a fully connected network with one hidden layer, and the output layer uses a sigmoid activation function to generate probability values ​​between 0 and 1. The treatment trajectory regressor outputs a predicted trajectory representing the changes in meningeal lymphatic vessel parameters over time as the treatment effect trajectory. This treatment trajectory regressor is a neural network-based regression model. Its input is the fusion feature representation, and its output is a parameter sequence. This parameter sequence defines the functional form describing the changes in key meningeal lymphatic vessel parameters over time. The functional form can be linear, exponentially decaying, or an S-shaped curve, with the specific form determined by the network weights. The therapeutic trajectory regressor outputs five parameters, representing the function type code, initial value, rate of change, asymptotic value, and inflection point time, respectively.

[0092] During the training phase, data from subjects with multiple imaging follow-ups were used. The baseline fusion feature representation was used as input, and changes in meningeal lymphatic vessel parameters measured during the follow-up period were used as supervisory signals to train the efficacy trajectory regressor, enabling it to predict future parameter change trends based on baseline information. In the application phase, the fusion feature representation of the subject to be evaluated was input into the trained efficacy trajectory regressor to obtain the predicted efficacy trajectory.

[0093] In practice, the quantitative feature set and clinical temporal feature vector are input into a multimodal pattern recognition model for joint analysis to generate classification probabilities for migraine auxiliary diagnosis and efficacy change trajectories for efficacy evaluation. The multimodal pattern recognition model includes a static feature encoding branch, a temporal feature encoding branch, and a fusion decision branch. The static feature encoding branch receives a quantitative feature set of a migraine subject (including eight features: total lymphatic vessel volume, surface area, median of length distribution, interquartile range of length distribution, branch complexity, average wall thickness, average signal intensity of lumen, and mean of signal intensity gradient along the vessel). The quantitative feature set is encoded into a high-dimensional static feature representation through a multilayer perceptron. The multilayer perceptron consists of an input layer, two hidden layers, and an output layer. The input layer has 8 nodes, the first hidden layer has 128 neurons and uses a modified linear unit activation function, the second hidden layer has 64 neurons and uses a modified linear unit activation function, and the output layer obtains a 128-dimensional high-dimensional static feature representation.

[0094] The temporal feature encoding branch receives clinical temporal feature vectors (30×8 matrices) and captures the temporal dependencies of these vectors using a recurrent neural network, encoding them as temporal context feature representations. The recurrent neural network employs a gated recurrent unit structure, sequentially processing each multidimensional feature point (8-dimensional feature points at each time point) arranged chronologically in the clinical temporal feature vector. At each time step, the gated recurrent unit calculates update and reset gates based on the input features of the current time step and the hidden state of the previous time step. The reset gate controls how much information from the hidden state of the previous time step is forgotten, and the update gate controls how much new information from the current time step is incorporated. The formula for calculating the reset gate is:

[0095]

[0096] in: Indicates time step The reset gate output vector at that location, This represents the sigmoid activation function. This represents the input weight matrix of the reset gate. Indicates time step The input is a multidimensional feature point (8-dimensional vector). This represents the hidden state weight matrix of the reset gate. This represents the hidden state (a 128-dimensional vector) of the previous time step. This represents the bias vector of the reset gate.

[0097] In some embodiments, through a gating mechanism consisting of update and reset gates, the recurrent neural network can capture the long-term temporal dependence and dynamic evolution patterns of clinical features under the influence of headache onset, relief, and medication events. The hidden state of the last time step is used as a temporal context feature representation, and the dimension of the hidden state is set to 128. For the clinical temporal feature vectors of 30 time steps, the gated recurrent unit processes the time steps sequentially. to Initial hidden state Set as the zero vector, the hidden state output after time step 30. As a temporal context feature representation. Optionally, the hidden state dimension of the gated recurrent unit is consistent with the dimension of the high-dimensional static feature representation (both are 128-dimensional). The fusion decision branch receives the high-dimensional static feature representation (128-dimensional) and the temporal context feature representation (128-dimensional), and calculates the interaction information between the high-dimensional static feature representation and the temporal context feature representation through a cross-attention mechanism to generate a fused feature representation. In the cross-attention mechanism, the high-dimensional static feature representation is used as the query vector, and the temporal context feature representation is used as the key vector and value vector to obtain the 128-dimensional fused feature representation.

[0098] The fused feature representations are input into a parallel diagnostic classifier and a treatment trajectory regressor, respectively. The diagnostic classifier outputs the classification probability of a subject being classified as a migraine patient or a healthy control. The diagnostic classifier uses a fully connected network with one hidden layer (32 neurons) and a modified linear unit activation function. The output layer uses a sigmoid activation function to generate classification probability values ​​between 0 and 1. The treatment trajectory regressor outputs a predicted trajectory representing the changes in meningeal lymphatic parameters over time as the treatment change trajectory. The treatment trajectory regressor is a neural network-based regression model. Its input is the fused feature representation, and its output is a parameter sequence. This parameter sequence defines the functional form describing the changes in key meningeal lymphatic parameters over time. The functional form can be linear, exponentially decaying, or a sigmoid curve, with the specific form determined by the network weights. The treatment trajectory regressor outputs five parameters representing the function type encoding, initial value, rate of change, asymptotic value, and inflection point time, respectively. In the training phase, subject data with multiple imaging follow-ups are used, with the baseline fused feature representation as input and the actual measured changes in meningeal lymphatic parameters during the follow-up period as the supervision signal to train the treatment trajectory regressor. In the application phase, the fusion feature representation of the subjects to be evaluated is input into the trained efficacy trajectory regressor to obtain the predicted efficacy change trajectory.

[0099] In one embodiment of the present invention, when outputting structured auxiliary diagnostic opinions and efficacy evaluation reports based on classification probabilities and efficacy change trajectories, the classification probabilities are compared with a preset diagnostic threshold. If the classification probability is greater than or equal to the diagnostic threshold, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated; otherwise, an auxiliary diagnostic opinion indicating a lack of tendency towards migraine is generated. The diagnostic threshold is set to 0.75 based on the receiver operating characteristic (ROC) curve analysis. The morphology of the efficacy change trajectory is analyzed, and the slope, curvature, or time point of stabilization of the trajectory is calculated as a quantitative indicator of efficacy dynamics. For linear trend functions, the slope value is calculated; a positive slope indicates improvement in lymphatic vessel function, and a negative slope indicates deterioration. For exponential decay functions, the decay constant is calculated and converted into a half-life as the time point of stabilization. For S-shaped curves, the time corresponding to the point of maximum curvature is calculated as the inflection point of efficacy.

[0100] From a pre-defined efficacy assessment rule base, the system matches the efficacy trajectory morphology with quantitative indicators to corresponding efficacy level descriptions. The rule base defines five levels: significant improvement, mild improvement, stable, mild deterioration, and significant deterioration. Each level corresponds to a different slope range or a time threshold interval for stabilization. The auxiliary diagnostic opinions, quantitative indicators of the efficacy trajectory, and matching efficacy level descriptions are integrated according to a pre-defined report template format to generate a structured auxiliary diagnostic opinion and efficacy assessment report containing diagnostic conclusions and efficacy analysis sections. The report template includes a subject basic information section, an image acquisition parameter section, a diagnostic conclusion section displaying classification probabilities and auxiliary diagnostic opinions, an efficacy analysis section displaying the predicted trajectory curve and corresponding efficacy level, and a clinical recommendation section.

[0101] In practice, structured auxiliary diagnostic opinions and efficacy evaluation reports are output based on classification probabilities and efficacy change trajectories. For a migraine subject, the multimodal pattern recognition model outputs a classification probability of 0.87, and the output efficacy change trajectory is in the form of an exponential decay function, with the function expression being: ,in Indicates at a point in time Predicted total volume of meningeal lymphatic vessels at (unit: month) location. Indicates the initial attenuation magnitude (cubic millimeters). This represents the decay rate constant (per month). Representing asymptotic values ​​(cubic millimeters), the subject's parameter sequence was (2, 125.6, -0.08, 42.3, 0), corresponding to... cubic millimeters per month, cubic millimeters.

[0102] In practice, the classification probability of 0.87 is compared with the preset diagnostic threshold of 0.75. Since the classification probability of 0.87 is greater than the diagnostic threshold of 0.75, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated. The morphology of the efficacy trajectory is analyzed, and the slope, curvature, or time point of stabilization of the trajectory are calculated as quantitative indicators of the efficacy dynamics, for an exponential decay function. The point at which things tend to stabilize Defined as when the function value changes from the initial value Decrease to asymptotic value The difference is the time corresponding to 5% of the initial attenuation amplitude, calculated using the following formula:

[0103]

[0104] in: Indicates the point in time when things tend to stabilize (in months). This represents the decay rate constant (per month). Representing the natural logarithm function, substituting the subject's... Monthly, calculated Months.

[0105] In some embodiments, based on a preset efficacy assessment rule base, the corresponding efficacy level description text is matched with the morphology of the efficacy change trajectory and quantitative indicators. The efficacy assessment rule base defines five levels: significant improvement, mild improvement, stable, mild deterioration, and significant deterioration. For an exponential decay function, when the decay rate is constant... Matching “significant improvement” monthly, when Matching "mild improvement" monthly, when Matching "stable" monthly, when Matching "mild deterioration" monthly, when Monthly matching "significant deterioration", subjects' Monthly settlement The scope and matching efficacy level description text are "mild improvement". Optionally, for linear trend functions, the matching rules in the efficacy assessment rule base are based on the slope value (unit: cubic millimeters per month); for S-curve functions, the matching rules are based on the inflection point time (unit: month) and the asymptotic change magnitude. The auxiliary diagnostic opinion (predisposition to migraine), the quantitative indicators of the efficacy change trajectory (the time point to stabilization is 37.45 months), and the matching efficacy level description text (mild improvement) are integrated according to a preset report template format to generate a structured auxiliary diagnostic opinion and efficacy assessment report containing diagnostic conclusions and efficacy analysis sections.

[0106] It is understood that the preset report template includes a subject basic information field (name, age, gender, examination number), an image acquisition parameter field (MRI equipment model, sequence type, scan date), a diagnostic conclusion field (displaying a classification probability of 0.87 and an auxiliary diagnostic opinion indicating a tendency towards migraine), an efficacy analysis field (displaying an exponential decay curve of the predicted total volume of meningeal lymphatic vessels over time, a decay rate constant of 0.08 per month, a stabilization point of 37.45 months, and an efficacy level of "mild improvement"), and a clinical recommendation field (providing recommendations to continue the current treatment plan and conduct regular follow-ups based on the efficacy level). In some embodiments, when the classification probability is exactly equal to the diagnostic threshold of 0.75, an auxiliary diagnostic opinion of "critical state" is generated, and a follow-up examination is recommended in the short term. Optionally, when the efficacy trajectory is a linear trend, the slope in the quantitative indicators of efficacy dynamics is directly taken as the slope value of the linear function; when the efficacy trajectory is an S-shaped curve, the curvature in the quantitative indicators of efficacy dynamics is taken as the first derivative value at the inflection point of the S-shaped curve, and the stabilization point is taken as the time point when the S-shaped curve reaches 95% of its final asymptotic value.

[0107] In practice, the efficacy assessment rule base is stored in the system in the form of a lookup table. Each efficacy level corresponds to a set of numerical interval conditions and a predefined descriptive text. The system selects the corresponding numerical interval conditions for comparison based on the function type of the efficacy change trajectory. If a match is found, the descriptive text is extracted. It can be understood that the efficacy analysis section in the report template also includes a text box to display the raw output values ​​of the efficacy change trajectory parameter sequence for clinicians to review.

[0108] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters, characterized in that, The method includes: Acquire a high-resolution magnetic resonance imaging dataset of the subjects, the high-resolution magnetic resonance imaging dataset including multi-sequence brain images and clinical assessment information acquired simultaneously with the images, the clinical assessment information including headache diaries, symptom scores and medication records; The multi-sequence brain images are subjected to meningeal lymphatic vessel segmentation and parameterization processing to extract a quantitative feature set reflecting the morphology and function of the meningeal lymphatic vessels. The meningeal lymphatic vessel segmentation and parameterization processing is implemented based on an improved region growing algorithm. The clinical assessment information is time-series structured encoded to generate a clinical time-series feature vector; The quantitative feature set and the clinical time-series feature vector are input into a multimodal pattern recognition model for joint analysis to generate classification probabilities for auxiliary diagnosis of migraine and efficacy change trajectories for efficacy evaluation. Based on the classification probability and the trajectory of efficacy changes, a structured auxiliary diagnostic opinion and efficacy evaluation report are output.

2. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 1, characterized in that, The multi-sequence brain images were segmented and parameterized to extract a set of quantitative features reflecting the morphology and function of the meningeal lymphatic vessels, including: The multi-sequence brain images are preprocessed, including head motion correction, spatial normalization, and registration between different sequence images, to obtain multimodal image data in standard space; In the multimodal image data under the standard space, based on the intensity and texture features of a specific sequence of images, the improved region growing algorithm is used to iteratively grow a complete three-dimensional structure of meningeal lymphatic vessels starting from a preset seed point. The improved region growing algorithm introduces prior local blood vessel orientation and anatomical structure constraints during the growth process. For the segmented three-dimensional meningeal lymphatic vessel structure, three-dimensional morphological analysis and functional parameter calculation based on image intensity are performed to extract multiple quantitative features, including total lymphatic vessel volume, surface area, length distribution, branching complexity, average wall thickness, average signal intensity of the lumen, and signal intensity gradient along the vessel. All extracted quantitative features are normalized and integrated to form the quantitative feature set.

3. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 2, characterized in that, The improved region growing algorithm incorporates prior knowledge of local blood vessel orientation and anatomical structure constraints during the growing process, including: In each growth iteration of the improved region growing algorithm, the local orientation consistency between the candidate growth point and the already grown points in its neighborhood is calculated. The expected vascular course range at the current location is obtained from the preset meningeal lymphatic anatomical atlas. Candidate points whose local course consistency is higher than a preset threshold and are within the expected vascular course range are given higher growth priority. Simultaneously, information on the boundaries of anatomical structures that are prohibited from being crossed near the current position is obtained from the anatomical atlas, and candidate points located outside the boundaries of the anatomical structures are completely excluded during the growth process; By combining the prior knowledge of local vascular orientation with anatomical constraints, the regional growth process is guided more precisely along the actual anatomical path of the meningeal lymphatic vessels.

4. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 3, characterized in that, In each growth iteration of the improved region growing algorithm, the local orientation consistency between the candidate growth point and its neighboring already grown points is calculated, including: In the three-dimensional image space, a local spherical neighborhood is defined with the candidate growth point as the center; Calculate the principal direction of the spatial point set formed by the positions of all grown points in the neighborhood; Calculate the direction vector from the candidate growth point to the centroid of the spatial point set; Calculate the cosine of the angle between the main direction and the direction vector, and use the cosine of the angle as a quantitative measure of the consistency of the local orientation. The closer the cosine value is to 1, the higher the consistency.

5. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 3, characterized in that, The preset meningeal lymphatic vessel anatomical atlas is constructed by probabilistic fusion based on the meningeal lymphatic vessel segmentation results in standard space of multiple healthy subjects.

6. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 1, characterized in that, The clinical assessment information is time-series structured encoded to generate a clinical time-series feature vector, including: The headache diary was analyzed to extract the onset time, duration, pain intensity, and temporal sequence of accompanying symptoms of headache attacks. The symptom scores were analyzed, and the score sequences of the standardized migraine symptom scale collected at different time points were extracted. The medication records are analyzed to extract the time sequence of drug name, administration time, and dosage. The temporal sequence of headache attacks, the score sequence of the symptom scale, and the temporal record of medication are time-aligned and unified onto the same time coordinate axis. At each aligned time point, the headache attack status characteristics, symptom score characteristics, and medication status characteristics are combined into a multidimensional feature point, which is then arranged in chronological order to form the clinical time series feature vector.

7. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 1, characterized in that, The quantified feature set and the clinical time-series feature vector are input into a multimodal pattern recognition model for joint analysis to generate classification probabilities for auxiliary diagnosis of migraine and efficacy change trajectories for efficacy evaluation, including: The multimodal pattern recognition model includes a static feature encoding branch, a temporal feature encoding branch, and a fusion decision branch; The static feature encoding branch receives the quantized feature set and encodes it into a high-dimensional static feature representation through a multilayer perceptron. The temporal feature encoding branch receives the clinical temporal feature vector, captures its temporal dependencies through a recurrent neural network, and encodes it into a temporal context feature representation. The fusion decision branch receives the high-dimensional static feature representation and the temporal context feature representation, calculates their interaction information through a cross-attention mechanism, and generates a fusion feature representation. The fused feature representations are respectively input into a parallel diagnostic classifier and a therapeutic trajectory regressor. The diagnostic classifier outputs the classification probability that the subject is classified as a migraine patient or a healthy control. The therapeutic trajectory regressor outputs a predicted trajectory characterizing the changes in meningeal lymphatic vessel parameters over time as the therapeutic change trajectory.

8. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 7, characterized in that, The temporal feature encoding branch receives the clinical temporal feature vector and captures its temporal dependencies through a recurrent neural network, including: The recurrent neural network adopts a gated recurrent unit structure to process each multidimensional feature point in the clinical time sequence feature vector arranged in chronological order in turn. At each time step, the gated loop unit calculates the update gate and the reset gate based on the input features of the current time step and the hidden state of the previous time step. The reset gate controls how much information about the hidden state of the previous time step is forgotten, and the update gate controls how much new information about the current time step is incorporated. Through the gating mechanism consisting of the update gate and the reset gate, the recurrent neural network can capture the long-term temporal dependence and dynamic evolution pattern of clinical features under the influence of headache onset, relief, and medication events. Finally, the hidden state of the last time step is used as the temporal context feature representation.

9. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 7, characterized in that, The therapeutic trajectory regressor outputs a predicted trajectory characterizing the changes in meningeal lymphatic vessel parameters over time as the therapeutic change trajectory, including: The therapeutic trajectory regressor is a neural network-based regression model. Its input is the fused feature representation, and its output is a parameter sequence. The parameter sequence defines a functional form describing the changes of key parameters of the meningeal lymphatic vessels over time. The function form includes linear trend, exponential decay, or S-curve, with the specific form determined by the network weights. During the training phase, subject data with multiple imaging follow-ups were used, with the baseline fusion feature representation as input and the actual measured changes in meningeal lymphatic parameters during the follow-up period as supervision signals, to train the therapeutic trajectory regressor so that it could predict future parameter change trends based on baseline information. In the application phase, the fusion feature representation of the subject to be evaluated is input into the trained efficacy trajectory regressor to obtain the predicted efficacy change trajectory.

10. The method for auxiliary diagnosis and efficacy evaluation of migraine based on pattern recognition magnetic resonance meningeal lymphatic vessel parameters according to claim 1, characterized in that, Based on the classification probabilities and efficacy change trajectories, a structured auxiliary diagnostic opinion and efficacy evaluation report are output, including: The classification probability is compared with a preset diagnostic threshold. If the classification probability is greater than or equal to the diagnostic threshold, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated; otherwise, an auxiliary diagnostic opinion indicating a tendency towards migraine is generated. Analyze the morphology of the therapeutic effect change trajectory, calculate the slope, curvature or the time point when the trajectory tends to stabilize, and use it as a quantitative indicator of the dynamics of the therapeutic effect. From the preset efficacy evaluation rule base, the corresponding efficacy level description text is matched according to the shape and quantitative indicators of the efficacy change trajectory; The auxiliary diagnostic opinions, the quantitative indicators of the efficacy change trajectory, and the matching efficacy level description text are integrated according to the preset report template format to generate the structured auxiliary diagnostic opinions and efficacy evaluation report containing diagnostic conclusions and efficacy analysis sections.

Images