A Deep Learning-Based Automatic Segmentation Method and System for Multimodal Medical Image Lesions

Abstract

This invention discloses a method and system for automatic lesion segmentation in multimodal medical images based on deep learning, belonging to the field of medical image processing. The method includes: Step 1, acquiring cross-center unlabeled multimodal magnetic resonance images, performing dual-channel asymmetric encoding and decoding based on the magnetic resonance signal equation, decoupling the image into latent variables of acquisition parameters and intrinsic tissue parameter maps, and obtaining the decoupled intrinsic tissue parameter map by combining the latent variables of acquisition parameters and the intrinsic tissue parameter map with the voxel-fidelity constraint of the synthesized image and the original image; Step 2, constructing an anchor point dictionary containing fingerprints of lesion and normal tissue parameters using the labeled lesion regions and tissue parameters of normal tissue in the intrinsic tissue parameter map. This invention can solve the problem of small-sample overfitting caused by the scarcity of high-quality labeled clinical samples, avoid the model capturing false feature associations, and adapt to intelligent segmentation scenarios with small clinical samples.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and more specifically, to a method and system for automatic segmentation of multimodal medical image lesions based on deep learning. Background Technology

[0002] Multimodal magnetic resonance imaging (MRI) lesion segmentation is a technique for clinical disease diagnosis, preoperative planning, and postoperative evaluation. By fusing complementary features from multiple image sequences, it can extract lesion morphology, texture, and functional information. Compared with single-modal image analysis methods, it has a stronger lesion representation capability and has been widely used in intelligent diagnosis and treatment scenarios for various diseases such as tumors and brain lesions. With the development of deep learning technology, deep learning-based multimodal medical image segmentation models, with their excellent feature extraction and pixel-level recognition performance, are gradually replacing traditional manual segmentation and traditional machine vision algorithms, becoming the mainstream solution for intelligent medical image analysis.

[0003] In practical cross-center clinical deployments, multimodal pixel-level lesion annotation is extremely challenging. It requires precise registration of multiple image sequences and meticulous delineation of lesion regions based on the clinical experience of professional physicians. The high manpower and time costs of annotation result in a severe shortage of high-quality labeled samples available for clinical use, leaving models to be trained on small samples for extended periods. Under these conditions, existing deep learning models are highly susceptible to capturing spurious associations in images, using non-lesion features such as scanner texture and imaging noise as shortcuts for discrimination, severely impacting the model's ability to identify true lesions.

[0004] On the other hand, there are significant differences in the field strength, imaging parameters, scanning protocols, and image reconstruction algorithms of MRI equipment in different medical institutions. This causes images from each center to form equipment-specific textures and fingerprints, resulting in cross-center distribution shifts. Existing domain adaptation techniques are difficult to adapt to the feature coupling characteristics of multimodal scenarios, cannot synchronously align the highly entangled joint distribution of multimodal images, and cannot effectively eliminate inter-domain differences. During model training, the model continuously overfits to center-specific scanning parameters, constantly reinforcing spurious feature associations. This results in excellent accuracy when the model is trained on single-center data, but when transferred to new target center data, drift and spurious feature overfitting problems occur, and the cross-center generalization ability is insufficient, which greatly restricts the large-scale clinical application of deep learning segmentation models. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention aims to provide a method and system for automatic segmentation of multimodal medical image lesions based on deep learning. This method can solve the problem of overfitting in small samples caused by the scarcity of high-quality labeled clinical samples, avoid the model capturing false feature associations, and adapt to intelligent segmentation scenarios with small clinical samples.

[0006] To solve the above problems, the present invention adopts the following technical solution:

[0007] Firstly, a multimodal medical image lesion automatic segmentation method based on deep learning includes:

[0008] Step 1: Acquire cross-center unlabeled multimodal magnetic resonance images, perform dual-channel asymmetric encoding and decoding based on the magnetic resonance signal equation, decouple the images into latent variables of acquisition parameters and tissue intrinsic parameter maps, and obtain the decoupled tissue intrinsic parameter maps by combining the latent variables of acquisition parameters and tissue intrinsic parameter maps with the voxel fidelity constraint of the synthesized images and the original images.

[0009] Step 2: Using the tissue parameters of the labeled lesion region and normal tissue in the tissue intrinsic parameter map, construct an anchor point dictionary containing the parameter fingerprints of lesions and normal tissue. The anchor point dictionary is established based on the manifold topological continuity of the intrinsic parameter maps of different centers.

[0010] Step 3: Use the cross-center parameter fingerprint anchor point dictionary to search and obtain the initial pseudo mask. Use the parameter gradient of the organization intrinsic parameter map as the driving force to make the boundary of the initial pseudo mask converge to the parameter abrupt change position to obtain the pseudo label.

[0011] Step 4: Use pseudo-labels to construct a mapping relationship with the tissue intrinsic parameter map as input and the lesion segmentation result as output. The mapping relationship uses the tissue relaxation parameter as the decision feature.

[0012] Step 5: Obtain multimodal magnetic resonance images of the target center, obtain the intrinsic tissue parameter map of the target center through decoupling, and obtain the lesion segmentation result using the mapping relationship;

[0013] Step 6: Using the obtained lesion segmentation results as a shape prior, apply shape consistency constraints with the shape prior to the decoupled output of the tissue intrinsic parameter map to correct the tissue intrinsic parameter map.

[0014] Further, step 1 includes:

[0015] The cross-center multimodal magnetic resonance images are output as tissue intrinsic parameter maps and acquisition parameter latent variable maps through the tissue channel and acquisition channel, respectively. The images are synthesized based on the magnetic resonance signal equation and subjected to voxel fidelity constraints to achieve decoupling.

[0016] A center-specific acquisition offset manifold is constructed based on the latent variable map of the acquisition parameters. The latent variable map of the acquisition parameters of the sample to be processed is normalized on the manifold to obtain a standardized latent variable map. The tissue intrinsic parameter map and the standardized latent variable map are synthesized into a virtual image by the magnetic resonance signal equation and compared with the original image to obtain the residual. The residual is used to inversely compensate the tissue intrinsic parameter map.

[0017] Further, step 2 includes:

[0018] On the intrinsic parameter map of the tissue, pixels with parameter gradient exceeding the threshold are removed based on the lesion label, and candidate detection points in the low gradient stable region are retained. The candidate detection points are clustered based on the intrinsic parameter vector, and the cluster center is taken as the parameter detection point and assigned a tissue semantic label.

[0019] The intrinsic parameter vectors of the parameter detection points of cross-center lesions and normal tissues are used to construct a point cloud. A nearest neighbor graph is constructed and iteratively diffused and smoothed until convergence to obtain a common parameter manifold. On the manifold, the density peak points of the detection points of each tissue class are searched and proposed as anchor parameters. An anchor dictionary is constructed from the anchor parameters and their semantic labels.

[0020] The difference in average proximity of each anchor parameter in the anchor dictionary to the lesion-type anchor and the normal tissue-type anchor is calculated as the contrast score. Based on the contrast score, the anchor dictionary is reorganized into first-level anchors with high contrast scores and second-level anchors with medium contrast scores. The second-level anchors participate in the response calculation of the pixel based on the proximity relationship between the pixel and the dominant region of the first-level anchor of a certain category in the parameter space.

[0021] Further, step 3 includes:

[0022] Using the first-level anchors of the hierarchical anchor dictionary, perform parametric distance matching on the tissue intrinsic parameter map, mark high-confidence candidate points, and for the remaining pixels in the pending state, activate the second-level anchors of the corresponding category to participate in response calculation based on their proximity to the dominant region of the first-level anchor of a certain category in the parameter space, and generate the initial lesion response map.

[0023] On the intrinsic parameter map, the gradient structure tensor of the multidimensional intrinsic parameter vector is calculated. Based on the relative relationship between the maximum and minimum eigenvalues ​​of the gradient structure tensor, a local metric for the width of the parameter transition band is determined, and a spatial adaptive evolution velocity map is constructed.

[0024] Furthermore, step 3 also includes:

[0025] Using the closed surface intercepted by the given response threshold on the initial lesion response map as the boundary of the zero level set, the hierarchical anchor point dictionary is retrieved at the boundary points, and the contrast score of the nearest neighbor anchor point is converted into a local driving force. This is then superimposed point by point with the spatial adaptive evolution velocity map to form a dual driving field. The driving boundary converges at the local maximum of the parameter gradient along the boundary evolution direction, thus obtaining pseudo-labels.

[0026] For points on the converged pseudo-label boundary, check whether the magnitude of the parameter gradient in the intrinsic parameter map is the maximum value in the local neighborhood of the point. For boundary points that do not reach the maximum value in the local neighborhood, adjust them along the normal direction to the position of the gradient maximum. Preserve the topology through surface interpolation to obtain regular pseudo-labels.

[0027] Further, step 4 includes:

[0028] The regular pseudo-labels are superimposed with the intrinsic parameter map of the tissue. The intrinsic parameter vectors of pixels marked as lesions and pixels marked as normal tissue are extracted by pseudo-labels. These vectors form the lesion parameter set and the normal parameter set, respectively. The lesion parameter set and the normal parameter set are then subjected to density subdivision and recursive separation in the multidimensional intrinsic parameter space. The intrinsic parameter dimensions are traversed and a single-dimensional threshold for separating the lesion parameter set and the normal parameter set is selected to generate a multidimensional threshold logical path tree.

[0029] The decision accuracy of each branch in the multidimensional threshold logic path tree is verified by using the intrinsic parameter maps of other cross-center samples and known pseudo-labels. Branches with accuracy lower than the preset cross-center consistency threshold are pruned, and equivalent replacement continuous threshold conditions are merged, retaining the mapping composed of intrinsic parameter thresholds.

[0030] Further, step 5 includes:

[0031] Acquire multimodal magnetic resonance images of the target center, normalize the decoupled acquisition parameter latent variable map into a standardized acquisition latent variable map on the acquisition offset manifold, synthesize the image with the decoupled tissue intrinsic parameter map using the magnetic resonance signal equation, and compare it with the original image to correct the tissue intrinsic parameter map with residuals;

[0032] The modified intrinsic parameter map of the tissue is used to determine the initial segmentation mask through a multi-dimensional threshold logic path tree. The boundary pixels of the initial segmentation mask are extracted, and the minimum parameter distance between the intrinsic parameter vector of the mask and the lesion anchor point and the normal tissue anchor point is retrieved from the hierarchical anchor point dictionary. The boundary pixel category is adjusted based on the minimum parameter distance and contrast score.

[0033] Further, step 6 includes:

[0034] A tubular analysis zone is constructed based on the boundary of the lesion segmentation mask. Within the tubular analysis zone, the parameter profile of the intrinsic tissue parameter map along the normal direction is calculated. The deviation of each parameter profile is compared with the interpolated prediction profile of the adjacent reliable profile. Pixels corresponding to the profiles whose deviation exceeds the abnormal threshold set based on the local parameter variance within the tubular analysis zone are marked as artifact pixels.

[0035] For the marked artifact pixels, search for samples with matching normal distance and expected tissue category in the non-artifact region, extract the centroid of its intrinsic parameter vector as restoration candidate value, project the restoration candidate value onto the common parameter manifold to obtain manifold regularized restoration value, write the restoration value into the artifact pixel to obtain the corrected tissue intrinsic parameter map.

[0036] Furthermore, step 6 also includes:

[0037] The intrinsic parameter map of the corrected tissue is compared with the intrinsic parameter map of the tissue before correction corresponding to the marked artifact pixel positions to obtain the artifact component map. The artifact component map is then spatially cross-correlation identified with the obtained latent variable map of the acquisition parameters. The components related to the artifact components in the latent variable map of the acquisition parameters are extracted as leakage artifact sources, and the contribution weight of the leakage artifact sources in the synthesis of the magnetic resonance signal equation is adjusted.

[0038] Secondly, the present invention also provides an automatic segmentation system for multimodal medical images based on deep learning, comprising: an intrinsic decoupling module, used to decouple cross-center multimodal magnetic resonance images into a tissue intrinsic parameter map and an acquisition parameter latent variable map, and to achieve decoupling through voxel fidelity constraints between synthetic images and original images based on magnetic resonance signal equations;

[0039] The anchor point construction module is used to extract tissue intrinsic parameter fingerprints of lesions and normal tissues by utilizing lesion annotations on the tissue intrinsic parameter map, and to construct a cross-center parameter fingerprint anchor point dictionary.

[0040] The pseudo-label generation module is used to retrieve the initial pseudo-mask by using the cross-center parameter fingerprint anchor point dictionary. The initial pseudo-mask boundary converges to the parameter abrupt change position by using the parameter gradient of the organization intrinsic parameter map as the driving force to obtain the pseudo-label.

[0041] The mapping construction module is used to construct a mapping from the tissue intrinsic parameter map to the lesion segmentation result using the tissue intrinsic parameter map and the obtained pseudo-annotations. The mapping uses the tissue intrinsic parameters of each pixel in the tissue intrinsic parameter map as the decision basis.

[0042] The segmentation output module is used to acquire multimodal magnetic resonance images of the target center, obtain tissue intrinsic parameter maps after decoupling, and input the tissue intrinsic parameter maps into the mapping to obtain lesion segmentation results;

[0043] The artifact correction module is used to correct the obtained tissue intrinsic parameter map based on the lesion segmentation results, so as to suppress tissue intrinsic parameter artifacts that are inconsistent with the segmentation results.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] (1) This scheme decouples the intrinsic tissue information of multimodal magnetic resonance images from the equipment acquisition information, gets rid of the dependence of segmentation decision on the texture of the imaging equipment and imaging noise, and generates high-quality pseudo-annotations based on the hierarchical anchor point dictionary, effectively solving the problem of small sample overfitting caused by the scarcity of high-quality clinical annotation samples, avoiding the model from capturing false feature associations, and adapting to the clinical small sample intelligent segmentation scenario.

[0046] (2) This scheme constructs a cross-center common parameter manifold and a standardized anchor point dictionary, normalizes the multi-center acquisition parameters, constructs a segmentation mapping based on the inherent tissue parameters of the human body, effectively adapts to the differences in field strength and scanning protocol of MRI equipment in different medical institutions, eliminates cross-center domain distribution offset, solves the defect of traditional domain adaptive technology that cannot adapt to multimodal feature coupling, and greatly improves cross-center generalization ability.

[0047] (3) This scheme adopts a dual-drive field boundary evolution and gradient extreme value boundary regularization mechanism, combined with a multi-dimensional threshold logic path tree to achieve refined pixel discrimination, and specifically optimizes the segmentation effect of the transition region of the lesion boundary, effectively improving the problems of blurred boundaries, misclassification of pixels, and distortion of contours in traditional segmentation methods.

[0048] (4) This scheme establishes a closed-loop mechanism for artifact correction and source weight optimization, accurately corrects artifacts of intrinsic parameters of the organization, traces the source of parameter leakage artifacts and adaptively adjusts the signal synthesis weight, reduces the probability of parameter distortion from the dual dimensions of end correction and source suppression, and continuously optimizes the cross-center adaptation stability of the model. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0050] Figure 1 This is a flowchart of the method for automatic segmentation of multimodal medical image lesions based on deep learning according to the present invention;

[0051] Figure 2 This is a flowchart of the image decoupling and acquisition parameter normalization process of the present invention;

[0052] Figure 3 This is a flowchart of the anchor point dictionary construction and pseudo-annotation generation process of the present invention;

[0053] Figure 4 This is a data flow diagram between the various modules of the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] Example 1

[0056] Please see Figures 1 to 3 A deep learning-based automatic segmentation method for multimodal medical image lesions is proposed. This method includes: Step 1, acquiring cross-center unlabeled multimodal magnetic resonance images, performing dual-channel asymmetric encoding and decoding based on the magnetic resonance signal equation, decoupling the image into latent variables of acquisition parameters and intrinsic tissue parameters, and obtaining the decoupled intrinsic tissue parameters by combining the latent variables of acquisition parameters and the intrinsic tissue parameters with the voxel-fidelity constraint between the synthesized image and the original image. The specific operations are as follows:

[0057] This invention is adapted to clinical multimodal MRI imaging scenarios such as tumors and brain lesions acquired by mainstream 1.5T and 3.0T medical MRI equipment in multiple centers. The aim is to solve the technical defect of high coupling between the inherent pathological features of human tissues and the interference features of equipment acquisition in the original multimodal MRI images. The information composition of a single voxel in the clinically acquired cross-center multimodal MRI images is complex. It contains the inherent physiological and pathological parameters of human tissues that do not change with the acquisition equipment and scanning protocol, as well as the equipment-specific acquisition interference information caused by the field strength, imaging parameters and reconstruction algorithms of different medical institutions. The deep coupling of these two types of information is the root cause of false feature learning in the model and cross-center distribution shift.

[0058] Based on the fundamental laws governing the changes in physical signals in medical magnetic resonance imaging (MRI), an industry-standard voxel grayscale error of 0 to 5 HU is set as the fidelity constraint control range. This parameter is set according to the national standard for clinical diagnostic accuracy of medical MRI images. With voxel-level image fidelity as the core constraint, the input cross-center multimodal MRI images are subjected to refined information decomposition processing. Through the decoupling logic constrained by the physical signal equation, the mixed and complex original images are accurately decomposed into two sets of independent image data without feature coupling: an intrinsic tissue parameter map that retains only the inherent attribute information of human tissue, and an acquisition parameter latent variable map that fully retains the specific interference information of the equipment acquisition. The full-process voxel fidelity constraint mechanism can effectively ensure that the decoupling operation does not destroy the lesion spatial location, tissue texture, and structural integrity of the original image, achieving distortion-free information decoupling and separating the inherent tissue features and equipment acquisition interference features from the image.

[0059] Step 1 further includes the following steps:

[0060] Step 11: The cross-center multimodal magnetic resonance images are output as tissue intrinsic parameter maps and acquisition parameter latent variable maps via the tissue channel and acquisition channel, respectively. Voxel-fidelity constraints are then applied to the synthesized images based on the magnetic resonance signal equation to achieve decoupling. The specific operations are as follows:

[0061] This step is the specific implementation of the overall decoupling logic in Step 1. First, the batch-acquired cross-center multimodal magnetic resonance images are imported pixel by pixel into the parallel tissue feature analysis channel and acquisition feature analysis channel. The dual channels adopt unbiased static feature mapping logic, without introducing additional weight optimization and feature selection mechanisms. The separation and extraction of the two types of features are completed entirely based on the physical composition logic of the magnetic resonance signal, avoiding human feature interference. Among them, the tissue feature analysis channel is specifically responsible for analyzing the information in the image that represents the inherent properties of human tissue. It extracts inherent physiological parameters such as longitudinal relaxation time, transverse relaxation time, and proton density of tissue pixel by pixel, and generates a tissue intrinsic parameter map whose dimensions match the parameters of the clinical routine multimodal magnetic resonance scanning sequence. The parameters of this map only reflect the real physiological and pathological state of human tissue and are independent of external acquisition conditions.

[0062] The feature analysis channel operates independently, specifically capturing unique features in images generated by variable acquisition conditions such as device field strength, scan slice thickness, repetition time, and echo time. It generates a high-dimensional latent variable map of acquisition parameters, completely recording all device acquisition offset information corresponding to a single sample. After completing the dual-channel feature output, forward image reconstruction is performed based on the general magnetic resonance signal equation for clinical medicine. Then, a new simulated image is synthesized using the currently decoupled tissue intrinsic parameters and acquisition latent variable parameters. The magnetic resonance signal equation is a spin echo (SE) sequence, specifically: ;

[0063] In the formula: Voxel magnetic resonance signal intensity under specific acquisition parameters is the gray value of a single voxel in the original magnetic resonance image, and is the core basis for voxel fidelity constraint comparison. Proton density is an intrinsic parameter of tissue, determined solely by the hydrogen proton concentration in human tissue and does not change with external collection conditions. The longitudinal relaxation time is an intrinsic parameter of the tissue and is an inherent physical property of human tissue that does not change with the collected parameters. 1 is the transverse relaxation time, an intrinsic tissue parameter that characterizes the decay characteristics of the tissue's inherent magnetization vector and is independent of external acquisition; e is the natural constant; TR is the repetition time, an equipment acquisition parameter set by the scanning protocol and the equipment, and is a core component of the acquisition parameter latent variable map; TE is the echo time, an equipment acquisition parameter used to quantify cross-center equipment acquisition specific differences.

[0064] By comparing the grayscale values, spatial structure, texture distribution, and lesion boundary features of the synthesized simulated image with the original input image on a voxel-by-voxel basis, a closed-loop voxel-fidelity constraint iterative optimization mechanism is constructed to continuously correct the feature mapping deviation of the two channels. Fixed technical parameters are set during the iterative optimization process. The iteration convergence threshold is set to a single voxel grayscale error of less than 3 HU, and the number of iterations can be fixed between 50 and 200. These parameters are derived from the industry standard for medical magnetic resonance imaging reconstruction accuracy and can adapt to the decoupling needs of most clinical cross-center imaging data. After iterative convergence, high-precision, distortion-free image information decoupling is achieved, ensuring that the two types of parameter maps after decomposition can completely reproduce all effective information of the original image.

[0065] Step 12: Construct a center-specific acquisition offset manifold based on the latent variable map of the acquisition parameters. Perform a normalization transformation on the manifold to obtain a standardized latent variable map of the acquisition parameters of the sample to be processed. Synthesize the tissue intrinsic parameter map and the standardized latent variable map using the magnetic resonance signal equation to obtain a virtual image, and compare it with the original image to obtain the residual. Use the residual to inversely compensate for the tissue intrinsic parameter map. The specific operations are as follows:

[0066] To further eliminate parameter distortion caused by equipment acquisition offset and fully adapt to cross-center image calibration of all mainstream 1.5T and 3.0T clinical MRI equipment, the following approach is first adopted: relying on the latent variable map of acquisition parameters corresponding to batch cross-center clinical samples, a surface fitting is performed in a high-dimensional parameter space to construct a continuous and smooth center-specific acquisition offset manifold. This manifold can completely cover the distribution patterns and offset characteristics of acquisition parameters of MRI equipment in different regions and medical institutions of different levels. The parameter space distance error of the manifold fitting is strictly controlled within 0.02, and the fitting accuracy meets the industry standard for normalization of medical imaging parameters, which can accurately characterize the acquisition-specific deviations of various equipment. The latent variable map of acquisition parameters of any clinical sample to be processed is mapped to the constructed offset manifold surface. Through global adaptive normalization transformation, the specific acquisition offset components unique to single-center equipment are filtered out, and a standardized acquisition latent variable map with unified dimension and normalized deviation is output. This achieves homogeneous and unified processing of acquisition parameters of all cross-center samples, eliminating the basic domain offset caused by cross-center equipment differences at the parameter level.

[0067] The initial intrinsic tissue parameter map output from step 11 and the normalized acquisition latent variable map are substituted into the magnetic resonance signal equation to complete the forward virtual image synthesis, generating a standardized virtual reconstructed image that completely eliminates device-specific interference. This virtual image is then compared with the patient's original real acquisition image in a global, voxel-by-voxel, and region-by-region fine-grained manner. The grayscale residual, spatial texture residual, and lesion boundary residual of the two images are accurately calculated. The resulting residual components can accurately correspond to the device acquisition artifacts and parameter distortion deviations mixed in the initial intrinsic tissue parameter map. Based on the spatial distribution position and value of the residuals, the parameter values ​​of the corresponding voxels in the intrinsic tissue parameter map are compensated and corrected point by point in reverse, eliminating the tissue parameter distortion problem caused by acquisition parameter offset one by one. Finally, a high-precision corrected intrinsic tissue parameter map that removes device acquisition interference and accurately represents the physiological and pathological characteristics of real human tissue is obtained, which provides a reliable and pure data base for subsequent lesion feature anchor point extraction, pseudo-annotation generation, and cross-center segmentation mapping.

[0068] In a preferred embodiment of the present invention, step 2 is further included: using the tissue parameters of the labeled lesion region and normal tissue in the tissue intrinsic parameter map, an anchor point dictionary containing fingerprints of lesion and normal tissue parameters is constructed. The anchor point dictionary is established based on the manifold topological continuity of the intrinsic parameter maps of different centers of tissue. The specific operation is as follows:

[0069] After decoupling and residual correction in step 1, the intrinsic parameter map of the tissue retains only the inherent physiological and pathological parameters of human tissues and does not have the problem of cross-center distribution shift, so it can be used as a stable feature extraction carrier. Using clinically accurate artificial lesion annotation as a constraint benchmark, the unique multidimensional parameter distribution features of lesion tissue and normal tissue are extracted from the intrinsic parameter maps of the tissues of massive cross-center samples. These stable and class-discriminative parameter distribution features are the intrinsic parameter fingerprints of the tissue. By screening, purifying, clustering and manifold optimization of the parameter fingerprints of the two types of tissues, feature anchors with cross-center universality and strong anti-interference ability are selected. All anchors and corresponding tissue semantic information are integrated to construct a standardized cross-center parameter fingerprint anchor dictionary. This dictionary can be freed from the constraints of device acquisition conditions and uniformly represent the inherent parameter differences between lesions and normal tissues in the global scene.

[0070] Step 2 further includes the following steps:

[0071] Step 21: On the intrinsic parameter map of the tissue, pixels with parameter gradients exceeding the threshold are removed based on lesion annotations, and candidate probe points in low-gradient stable regions are retained. The candidate probe points are clustered based on the intrinsic parameter vectors, and the cluster centers are taken as parameter probe points and assigned tissue semantic labels. The specific operations are as follows:

[0072] Within the corrected intrinsic tissue parameter map, relying on the precise lesion region contour marked by clinical manual annotation, the pixel range of the lesion region and the normal tissue region is divided. The gradient value of the multidimensional intrinsic parameters corresponding to each voxel is calculated pixel by pixel. The parameter gradient is used to characterize the degree of change of local tissue parameters. The parameter gradient threshold adjustment range of this technical solution can be set from 0.08 to 0.20. This threshold range is taken from the general standard for homogeneous region detection of tissue parameters in medical imaging and is suitable for the parameter distribution characteristics of most human soft tissues and lesion tissues. Pixels with gradient values ​​exceeding the preset threshold are removed. These pixels are mostly located in the transition area of ​​lesion boundary, tissue damage area or small residual artifact area. The parameter values ​​fluctuate greatly and have poor stability, and do not have the ability to represent tissues globally.

[0073] After removing unstable pixels, pixels with gradient values ​​below a threshold are considered to be from homogeneous and stable tissue regions. These pixels are uniformly marked as candidate probe points. For all candidate probe points, the multidimensional tissue intrinsic parameter vectors corresponding to each pixel are extracted as feature bases. An unsupervised clustering algorithm is used to spatially cluster all parameter vectors. The convergence condition for clustering iteration is set such that the change in intra-cluster parameter distance is less than 0.01, for example, the number of iterations is controlled between 40 and 120, which are conventional and controllable parameters for high-dimensional parameter clustering in medical images. After clustering, several stable clusters are obtained. The geometric center of each cluster is selected as the final parameter probe point. The cluster center can accurately represent the common parameter features of all pixels within the corresponding cluster, avoiding the random parameter fluctuation error of single-point pixels. Finally, based on the initially manually labeled regional attributes, each parameter probe point is uniformly assigned a fixed semantic label of lesion tissue or normal tissue, completing the accurate sampling and semantic labeling of stable feature probe points.

[0074] Step 22: Construct a point cloud from the intrinsic parameter vectors of the parameter detection points of cross-center lesions and normal tissues, construct a nearest neighbor graph and iteratively diffuse and smooth it until convergence to obtain a common parameter manifold. On the manifold, search for density peak points of detection points of each tissue class and propose them as anchor parameters. An anchor dictionary is constructed from the anchor parameters and their semantic labels. The specific operations are as follows:

[0075] We integrate semantically labeled lesion and normal tissue parameter probes from all cross-center clinical samples, and map the multidimensional intrinsic parameter vectors corresponding to each probe point to a high-dimensional parameter space to form a global parameter point cloud data covering multiple devices, multiple patients, and multiple lesion scenarios. Based on the global parameter point cloud, we construct a parameter space nearest neighbor association graph. The Euclidean distance of 0.15 in the parameter space can be used as the nearest neighbor determination benchmark. This distance threshold is adapted to the distribution density characteristics of multidimensional medical tissue parameters. We establish adjacency relationships for probe points in the parameter space whose distance is less than the benchmark value, and fully restore the spatial association rules between different tissue parameters. We perform iterative diffusion smoothing on the constructed nearest neighbor graph. The smoothing iteration number is set to 60 to 150 times. During the iteration process, we continuously correct the spatial distribution deviation of discrete parameter points and weaken the individual differences of parameters in single samples and single centers until the point cloud distribution tends to stabilize and converge. Finally, we fit a common parameter manifold that can characterize the distribution rules of global tissue parameters.

[0076] This manifold can eliminate individual differences and subtle parameter perturbations, preserving the most essential parameter distribution differences between lesion tissue and normal tissue. On the converged common parameter manifold surface, spatial density statistics are performed on the probe points of lesion tissue class and normal tissue class respectively, searching for local density peak points of parameter distribution in each category. Density peak points are the feature parameter points with the most concentrated distribution, the strongest stability, and the highest cross-center consistency in the corresponding tissue category, and these peak points are recommended as core anchor point parameters. The anchor point parameters corresponding to all categories are summarized, and the inherent semantic labels corresponding to each anchor point parameter are matched to form a complete cross-center parameter fingerprint anchor point dictionary, realizing the standardized construction of a general lesion feature library.

[0077] Step 23: Calculate the difference in average proximity between each anchor parameter in the anchor dictionary and the anchors for lesion and normal tissue types as the contrast score. Based on the contrast score, reorganize the anchor dictionary into first-level anchors with high contrast scores and second-level anchors with medium contrast scores. The second-level anchors participate in the response calculation for that pixel based on their proximity relationship with the dominant region of a first-level anchor of a certain category in the parameter space. The specific operation is as follows:

[0078] For each anchor parameter with defined semantics in the anchor dictionary, the average spatial proximity of the anchor parameter to all lesion-type anchors and the average spatial proximity to all normal tissue-type anchors are calculated. The difference between the two average proximity values ​​is the contrast score of the anchor parameter. The magnitude of the contrast score directly corresponds to the anchor parameter's ability to distinguish between lesion tissue and normal tissue; the higher the score, the better the feature differentiation effect between the two types of tissues. Based on the general quantile standard for medical image feature stratification, the top 30% of the global anchor contrast scores can be used as the high contrast judgment interval. Anchors in this interval are classified as first-level anchors, and the remaining anchors in the medium score interval are classified as second-level anchors. Invalid anchors with too low scores and no distinguishing value are removed, thus completing the hierarchical reorganization and optimization of the anchor dictionary.

[0079] The first-level anchor points possess extremely strong tissue category discrimination capabilities and can be directly used for accurate category determination of pixels in homogeneous regions. The second-level anchor points, when used alone, lack sufficient feature discrimination stability. Therefore, a subordinate response mechanism is set up. This type of anchor point does not participate in pixel discrimination calculation independently. Only when the distribution position of the pixel to be detected in the multi-dimensional parameter space falls within the dominant coverage area of ​​a certain category of the first-level anchor point, will the corresponding category's second-level anchor point synchronously participate in the feature response calculation of that pixel. Through the synergistic cooperation of the two-level anchor points, the deficiency of a single high-contrast anchor point in not fully covering the features of lesion boundaries and parameter transition areas is compensated for, thus improving the feature discrimination system for complex lesion regions.

[0080] In a preferred embodiment of the present invention, step 3 is further included: using a cross-center parameter fingerprint anchor point dictionary for retrieval to obtain an initial pseudo-mask; using the parameter gradient of the organization intrinsic parameter map as the driving force, the boundary of the initial pseudo-mask converges towards the parameter abrupt change position to obtain pseudo-labels. The specific operation is as follows:

[0081] Based on the clean tissue intrinsic parameter map after decoupling and correction, the lesion attribute response of all pixels in the image is calculated using cross-center universal two-level anchor point features. This initially delineates the distribution range of lesion regions and normal tissue regions, resulting in an initial pseudo-mask with low accuracy and rough boundaries. To eliminate problems such as boundary offset, regional redundancy, and edges not conforming to the boundaries of abrupt changes in real tissue parameters in the initial pseudo-mask, this step uses the spatial gradient change characteristics of tissue intrinsic parameters as the core evolutionary driving force. Relying on the pixel-level local parameter change law, the mask boundary is continuously iterated and updated, so that the boundary gradually conforms to the boundary position of the real lesion where tissue parameters change abruptly. After the boundary converges to a stable state, local neighborhood gradient extremum verification and boundary topology regularization are performed to correct misaligned pixels, smooth jagged edges, and retain reasonable spatial topology, ultimately generating a high-quality pseudo-label with accurate boundaries, conforming to the inherent characteristics of tissue, and adapting to the cross-center data distribution.

[0082] Step 3 further includes the following steps:

[0083] Step 31: Perform parametric distance matching on the tissue intrinsic parameter map using the first-level anchor points of the hierarchical anchor point dictionary, mark high-confidence candidate points, and for the remaining pixels in the pending state, activate the corresponding second-level anchor points to participate in the response calculation based on their proximity to the dominant region of a certain category's first-level anchor point in the parameter space, thereby generating the initial lesion response map. The specific operations are as follows:

[0084] First, the first-level anchors with stronger discriminative power from the hierarchical anchor dictionary are called. Using the spatial distance of multidimensional intrinsic parameters as the matching basis, the spatial distance between the parameter vector of each voxel in the image and the parameter vector of the first-level anchor of each category is calculated pixel by pixel. The parameter distance calculation adopts the Euclidean distance calculation method commonly used in medical high-dimensional feature matching. The high-confidence matching threshold range can be set to 0 to 0.12. This threshold range is suitable for the parameter distribution interval characteristics of clinical lesions and normal tissues. Pixels whose parameter distance is within this threshold range are judged as high-confidence candidate points. The pixel category is directly marked based on the semantic label of the matching anchor, realizing the rapid classification of high-determinism regions.

[0085] For undetermined pixels whose parameter distance exceeds the high confidence threshold and whose category cannot be directly determined, the landing area of ​​the pixel in the multi-dimensional parameter space is further determined to confirm the dominant coverage area of ​​the first-level anchor point to which the pixel belongs. Based on the dominant area of ​​the category to which the pixel belongs, the second-level anchor point of the corresponding category is selectively activated to participate in the feature response calculation of the pixel. Through the joint feature fitting of the two-level anchor points, the lesion belonging probability of the undetermined pixel is quantified and output. After traversing all pixels of the image to complete the global response calculation, the response probability value of each pixel is normalized to the numerical range of zero to one, forming an initial lesion response map with gray value corresponding to the lesion belonging confidence, which fully covers the feature response results of the high confidence area and the weak boundary undetermined area.

[0086] Step 32: On the intrinsic parameter map, calculate the gradient structure tensor of the multidimensional intrinsic parameter vectors. Based on the relative relationship between the maximum and minimum eigenvalues ​​of the gradient structure tensor, determine the local metric of the parameter transition band width and construct a spatial adaptive evolution velocity map. The specific operations are as follows:

[0087] On the intrinsic parameter map, a 3×3 local computation window is selected, and the multidimensional intrinsic parameter vectors of all pixels within the window are extracted to construct the gradient structure tensor at the corresponding voxel positions. The expression for the gradient structure tensor is: ;

[0088] In the formula, (x,y) represents the two-dimensional spatial coordinates of a voxel on the tissue intrinsic parameter map; The gradient vector of the multidimensional intrinsic parameters of the voxel is used to characterize the spatial rate of change of the intrinsic parameters of the tissue in the local region. It is the squared term of the parameter gradient in the x-direction, i.e. ; It is the squared term of the parameter gradient in the y-direction, i.e. .

[0089] and For gradient cross terms, satisfying It is used to describe the anisotropic characteristics of parameter gradient changes; for the gradient structure tensor Perform eigenvalue decomposition to obtain the largest eigenvalue. with minimum eigenvalue Based on the relative relationship between the two, the expression for the local transition band width measurement function is constructed as follows: ;

[0090] In the formula, w(x,y) is a measure of the width of the local parameter transition zone at voxel (x,y), and its value directly corresponds to the width of the parameter transition region between the lesion and normal tissue. , is the scale normalization coefficient, a fixed constant in clinical medical image gradient analysis, used to map the metric value to the adaptation range of the boundary evolution rate; The largest eigenvalue of the gradient structure tensor represents the intensity of the principal direction of local parameter changes and reflects the severity of abrupt changes in the intrinsic parameters of the tissue. The minimum eigenvalue of the gradient structure tensor represents the intensity of local parameter changes in the sub-directions, reflecting the uniformity of the tissue's intrinsic parameter distribution. After adaptively calculating the width of the local transition band pixel by pixel across the entire image using this metric function, the metric value is mapped to generate a spatially adaptive evolution rate map. Homogeneous tissue regions with narrow transition bands correspond to higher boundary evolution rates, while complex lesion boundary regions with wide transition bands correspond to lower boundary evolution rates. This achieves global adaptive control of the boundary evolution rate, avoiding the boundary overfitting or insufficient convergence problems caused by a fixed evolution rate.

[0091] Step 33: Using the closed surface intercepted by the given response threshold on the initial lesion response map as the boundary of the zero level set, the hierarchical anchor point dictionary is retrieved at the boundary points. The contrast score of the nearest neighbor anchor point is converted into a local driving force, which is then superimposed point-by-point with the spatial adaptive evolution velocity map to form a dual driving field. The driving boundary converges at the local maximum of the parameter gradient along the boundary evolution direction, resulting in pseudo-labels. The specific operations are as follows:

[0092] A fixed response threshold is pre-set, with a value range of 0.4 to 0.6, which is a general neutral threshold range for pseudo-mask extraction. Continuous regions with response values ​​greater than the threshold are extracted from the initial lesion response map, and their closed contours are extracted as the initial zero-level set boundary, completing the contour localization of the initial pseudo-mask. For each contour point on the zero-level set boundary, the anchor point with the closest parameter space distance is retrieved from the hierarchical anchor point dictionary. The pre-calculated contrast score of this anchor point is retrieved, and the contrast score is linearly mapped to a local boundary driving force value. The higher the contrast score of the boundary position, the stronger the local driving correction capability, adapting to the need for refined adjustment of complex lesion boundaries. The local driving force converted from the anchor point contrast score is superimposed pixel-by-pixel with the spatial adaptive evolution velocity map constructed in step 32, forming a global dual driving field, simultaneously considering the category discrimination constraint of anchor point semantic features and the spatial position constraint of local parameter gradients. Under the unified control of the dual driving fields, the initial zero level set boundary moves iteratively along the preset evolution direction. During the iteration process, the local maximum position of the parameter gradient on the boundary path is continuously searched. The parameter gradient maximum corresponds to the real tissue boundary position where the mutation of the tissue intrinsic parameters is most significant. The boundary eventually converges to the contour line composed of all local gradient maxima, completing the coarse optimization of the initial pseudo-label and obtaining the preliminary pseudo-label contour that fits the real boundary of the tissue parameters.

[0093] Step 34: For points on the converged pseudo-label boundary, check whether the magnitude of the parameter gradient in the intrinsic parameter map is the maximum value in the local neighborhood of the point. For boundary points that do not reach the local neighborhood maximum, adjust them along the normal direction to the gradient maximum position. Preserve the topology through surface interpolation to obtain regular pseudo-labels. The specific operation is as follows:

[0094] For each boundary pixel on the pseudo-annotated contour after iterative convergence, a 5x5 local neighborhood verification window is selected. This window size balances the integrity of local parameter features with noise immunity, conforming to the general standards for medical image boundary verification. The gradient magnitude of the tissue intrinsic parameters of the boundary pixel and all pixels in its neighborhood is calculated one by one to determine whether the gradient magnitude of the current boundary pixel is a maximum value in the local neighborhood. If the gradient magnitude of the boundary pixel reaches the neighborhood maximum, the boundary point is determined to fit the boundary of the true parameter abrupt change and no adjustment is needed; if the gradient magnitude of the boundary pixel does not reach the neighborhood maximum, the current boundary position is determined to be offset and position correction is required.

[0095] For boundary points with offsets, iterative adjustments with small step sizes are made along the local normal direction of the boundary. The step size for each adjustment can be set to 0.5 to 1 voxel unit, gradually moving to the pixel position corresponding to the maximum value of the gradient magnitude in the neighborhood, thus completing the position correction of all offset boundary points. After the boundary point adjustment is completed, a surface interpolation algorithm is used to perform continuous surface fitting on the discrete correction boundaries, completely preserving the spatial topological structure of the lesion contour and avoiding problems such as boundary breaks, topological distortion, and local voids caused by point-by-point adjustments. Finally, a regular pseudo-label with continuous contour, accurate boundaries, complete topology, and conforming to the inherent parameter distribution characteristics of human tissue is generated.

[0096] In a preferred embodiment of the present invention, step 4 is further included, which involves constructing a mapping relationship using pseudo-annotations, with the tissue intrinsic parameter map as input and the lesion segmentation result as output. The mapping relationship uses the tissue relaxation parameter as a decision feature. The specific operation is as follows:

[0097] Based on the pure tissue intrinsic parameter map and high-precision regular pseudo-annotations obtained in the previous steps, a lesion segmentation mapping relationship based entirely on the inherent parameter characteristics of human tissue is constructed, which is suitable for clinical segmentation scenarios of brain lesions and tumor magnetic resonance images with multiple centers, multiple devices, and small samples. Unlike the traditional deep learning model that relies on image texture, pixel grayscale and other features that are easily interfered with by the device to establish segmentation mapping, the mapping relationship constructed in this step is based entirely on the tissue intrinsic parameters corresponding to each pixel as the sole decision basis, avoiding false feature associations caused by scanner texture, imaging noise and cross-center scanning parameter differences.

[0098] After boundary correction and topological normalization, the normalized pseudo-labels accurately match the mutation features of tissue intrinsic parameters. This can serve as a reliable process for constructing supervised data constraint mapping relationships. By mining the inherent distribution differences between lesion tissue and normal tissue in the multidimensional intrinsic parameter space, parameter discrimination thresholds with cross-center universality are selected to form a structured parameter decision logic. Finally, a stable and generalizable exclusive mapping from tissue intrinsic parameter map to lesion segmentation results is established. This mapping is only associated with the inherent physiological and pathological properties of human tissue and is not bound to any device-collected features. It can fundamentally alleviate the problems of model cross-center drift and overfitting, and provide algorithmic support for direct segmentation inference without adaptive training at the target center. The entire process adopts general technical parameters for medical image parameter classification, and the entire mapping construction process can be stably reproduced.

[0099] Step 4 further includes the following steps:

[0100] Step 41: Overlay the regular pseudo-labels with the intrinsic parameter map of the tissue, extract the intrinsic parameter vectors of pixels pseudo-labeled as lesions and pixels pseudo-labeled as normal tissue, respectively forming lesion parameter sets and normal parameter sets. Perform density subdivision and recursive separation on the lesion parameter sets and normal parameter sets in the multidimensional intrinsic parameter space, traverse the intrinsic parameter dimensions and select the single-dimensional threshold for separating the lesion parameter sets and normal tissue sets, and generate a multidimensional threshold logical path tree. The specific operations are as follows:

[0101] The regularized pseudo-labels output from step 3 are precisely overlaid and matched with the preprocessed tissue intrinsic parameter map at the pixel level. Parameter samples are classified and extracted based on the pixel semantic labels of the pseudo-labels. The multidimensional intrinsic parameter vectors corresponding to all pixels pseudo-labeled as lesion areas are aggregated and integrated to form a lesion parameter set. Simultaneously, the multidimensional intrinsic parameter vectors corresponding to all pixels pseudo-labeled as normal tissue areas are aggregated and integrated to form a normal parameter set. Both parameter sets originate entirely from tissue intrinsic parameters without device interference, containing only the inherent pathological and physiological characteristics of human tissue and exhibiting cross-center consistency. Refined density subdivision processing is performed on both parameter sets in the multidimensional intrinsic parameter space. The parameter space subdivision step size can be set to 0.02, which is taken from the general accuracy standard for medical high-dimensional parameter space segmentation and can accurately capture minute differences in parameter distribution.

[0102] After density subdivision, the overlapping parameter regions are recursively separated and iterated. The convergence condition is set to the overlap ratio of the two parameter sets being less than 3% to ensure effective differentiation between the two parameter samples. All dimensions of the intrinsic parameters are traversed one by one, and a binary classification discrimination test is performed on each dimension parameter separately to select the optimal single-dimensional threshold that can effectively separate the lesion parameter set from the normal parameter set. The effective threshold discrimination logic and branch judgment relationship of all dimensions are integrated to build a multi-dimensional threshold logic path tree that covers all parameter dimensions and has multi-branch decision-making capabilities, thereby realizing structured decision modeling from parameter features to lesion categories.

[0103] Step 42: Verify the decision accuracy of each branch in the multidimensional threshold logical path tree using the intrinsic tissue parameter maps of other cross-center samples and known pseudo-labels. Prune branches with accuracy lower than the preset cross-center consistency threshold, merge equivalently replaceable continuous threshold conditions, and retain the mapping composed of intrinsic tissue parameter thresholds. The specific operations are as follows:

[0104] The constructed initial multidimensional threshold logic path tree contains a large number of dimensional discrimination branches. Some branches are only adapted to the parameter distribution characteristics of single-center data and do not have global universality. They need to be filtered and optimized through cross-center data validation. Independent clinical magnetic resonance imaging samples collected from multiple centers and multiple devices are selected as the validation set. The validation set covers data from mainstream 1.5T and 3.0T medical magnetic resonance imaging devices and is adapted to clinical scenarios with different scanning protocols and imaging parameters. Using the tissue intrinsic parameter map and corresponding regular pseudo-labels of the validation set, the lesion discrimination accuracy of each decision branch of the multidimensional threshold logic path tree is tested branch by branch. The cross-center consistency threshold can be set to 0.85, which is a general acceptable accuracy standard for cross-center medical image recognition tasks. The average decision accuracy of each branch in multi-center validation samples is calculated, and decision branches with an average accuracy below this preset threshold are pruned, eliminating specific discrimination logics that only fit single-center data and pose an overfitting risk. Simultaneously, the remaining valid branches are traversed, and adjacent threshold conditions with continuous parameter ranges, equivalent discrimination logic, and consistent output results are merged and integrated to simplify the redundant structure of the path tree and reduce the computational complexity of the algorithm. After this dual optimization of branch pruning and logic merging, the path tree retains only stable threshold discrimination logic that can adapt to multi-center data and relies on the intrinsic parameters of the tissue, ultimately forming a lesion segmentation mapping relationship that is purely based on the intrinsic parameters of the tissue and has strong cross-center generalization ability.

[0105] In a preferred embodiment of the present invention, step 5 is further included: acquiring a multimodal magnetic resonance image of the target center, obtaining the intrinsic tissue parameter map of the target center through decoupling, and obtaining the lesion segmentation result using the mapping relationship. The specific operation is as follows:

[0106] Based on the previously trained acquisition offset manifold, hierarchical parameter fingerprint anchor point dictionary, and multidimensional threshold logic path tree mapping relationship, standardized preprocessing and feature correction are performed on the newly acquired original multimodal magnetic resonance images of the target center. This completely removes the device-specific acquisition interference of the target center, resulting in a tissue intrinsic parameter map that retains only the inherent pathological features of human tissue. The corrected pure tissue intrinsic parameter map is then input into the solidified parameter discrimination mapping system. Pixel-level lesion classification is performed based on the cross-center universal discrimination logic of the tissue intrinsic parameters. Furthermore, the high-dimensional parameter distance and contrast features of the anchor point dictionary are combined to refine the boundary, ultimately outputting accurate lesion segmentation results that adapt to the data characteristics of the target center and eliminate the domain drift problem.

[0107] Step 5 further includes the following steps:

[0108] Step 51: Acquire multimodal magnetic resonance images of the target center. Normalize the decoupled acquisition parameter latent variable map onto the acquisition offset manifold to obtain a standardized acquisition latent variable map. Combine this map with the decoupled tissue intrinsic parameter map using the magnetic resonance signal equation to synthesize an image, which is then compared with the original image. The residuals are used to correct the tissue intrinsic parameter map. The specific operations are as follows:

[0109] First, acquire the original multimodal magnetic resonance image of the target center to be detected. Then, call the dual-branch decoupling structure built in step 1 to perform information decoupling operation on the target image, and output the initial tissue intrinsic parameter map and the acquisition parameter latent variable map corresponding to the target sample respectively. Map the decoupled acquisition parameter latent variable map to the global center-specific acquisition offset manifold surface pre-constructed in step 12. Based on the distribution law of the device acquisition parameters in all dimensions covered by the manifold, perform a global normalization transformation on the acquisition latent variable features of the target sample. The spatial distance error of the manifold normalization is strictly controlled within 0.02. This accuracy value is taken as the general standard for cross-center parameter normalization of medical images. Finally, a standardized acquisition latent variable map that eliminates the target device-specific offset features is obtained.

[0110] The normalized latent variable map and the tissue intrinsic parameter map obtained by synchronous decoupling are substituted into the magnetic resonance signal equation to complete the forward image reconstruction and generate a standardized synthetic virtual image without equipment interference. The virtual image is compared with the original acquired image of the target center on a voxel-by-voxel basis to accurately calculate the gray-level residual and structural residual. The residual judgment threshold adopts the 3HU accuracy standard commonly used in clinical image correction. Based on the spatial distribution and numerical magnitude of the residual, the initial tissue intrinsic parameter map is corrected by reverse voxel-by-voxel compensation. This thoroughly corrects the tissue parameter artifacts and numerical deviations caused by the equipment scanning parameters and imaging algorithm of the target center. Finally, a corrected tissue intrinsic parameter map with the same parameter distribution as the training data and only characterizing the inherent tissue features of the human body is obtained.

[0111] Step 52: The modified tissue intrinsic parameter map is used to determine the initial segmentation mask through a multi-dimensional threshold logical path tree. The boundary pixels of the initial segmentation mask are extracted, and the minimum parameter distance between their intrinsic parameter vector and the lesion anchor point and the normal tissue anchor point is retrieved in the hierarchical anchor point dictionary. The boundary pixel category is adjusted according to the minimum parameter distance and contrast score. The specific operations are as follows:

[0112] The tissue intrinsic parameter map corrected in step 51 is input into the multidimensional threshold logic path tree optimized and solidified in step 4. Following the threshold discrimination rules of each dimension of intrinsic parameters within the path tree, the entire tissue attribute is determined pixel-by-pixel, quickly distinguishing lesion tissue pixels from normal tissue pixels and generating an initial lesion segmentation mask. This initial segmentation mask can accurately identify the lesion subject in the homogeneous region of the image center. However, for the parameter transition region at the boundary between lesions and normal tissue, confusion of boundary pixel categories and local contour shifts are prone to occur. To optimize boundary segmentation accuracy, all boundary pixels of the initial segmentation mask are extracted one by one. Parameter retrieval is performed in the cross-center hierarchical parameter fingerprint anchor point dictionary. The multidimensional intrinsic parameter vector of the current boundary pixel is calculated, and the parameter space distance between it and all lesion-type and normal tissue-type anchor points in the dictionary is calculated. The minimum parameter distance between the corresponding two categories is then selected.

[0113] By combining the contrast score of the corresponding minimum distance anchor point, the confidence of the tissue category discrimination of the current boundary pixel is quantified. The minimum parameter distance is used as the basic criterion for category classification, and the contrast score is used as the discrimination weight correction condition. The final category attribute of the pixels in the boundary transition area is comprehensively adjusted. Through iterative correction of the boundary pixels of the whole domain, stray misclassified pixels in the initial mask are removed and the offset boundary contour is corrected, so that the segmentation boundary accurately fits the real abrupt change position of the tissue intrinsic parameters, and finally a high-precision and high-generalization target center lesion segmentation result is obtained.

[0114] In a preferred embodiment of the present invention, step 6 is further included: using the obtained lesion segmentation result as a shape prior, applying a shape consistency constraint with the shape prior to the decoupled output of the tissue intrinsic parameter map, in order to correct the tissue intrinsic parameter map. The specific operation is as follows:

[0115] After the aforementioned decoupling correction and segmentation inference process, the intrinsic tissue parameter map may still contain a small number of local artifact parameters caused by residual coupling of acquisition parameters and signal synthesis deviations. These artifact parameters do not match the physiological and pathological characteristics of real human tissue, causing local parameter distortion and slightly affecting the accuracy of the segmentation results. By using the high-precision lesion segmentation results output in step 5 as the constraint benchmark for the real tissue distribution, the rationality of the global parameters of the intrinsic tissue parameter map is verified in reverse. Artifact pixels that do not conform to the semantics of lesion segmentation are accurately identified and corrected, completing the secondary fine correction of the intrinsic tissue parameter map. At the same time, through artifact feature tracing analysis, the source of acquisition parameter leakage artifacts in the magnetic resonance signal synthesis process is located, and the parameter contribution weight of signal synthesis is adaptively adjusted. This suppresses the artifact residue problem in the subsequent image decoupling and parameter generation process from the source of the algorithm, forming a closed-loop mechanism of parameter correction and source optimization.

[0116] Step 6 further includes the following steps:

[0117] Step 61: Construct a tubular analysis band based on the boundary of the lesion segmentation mask. Calculate the parameter profiles of the intrinsic tissue parameters along the normal direction within the tubular analysis band. Compare the deviations of each parameter profile with the interpolated predicted profiles of adjacent reliable profiles. Mark the pixels corresponding to profiles whose deviations exceed the abnormal threshold set based on the local parameter variance within the tubular analysis band as artifact pixels. The specific operations are as follows:

[0118] Using the lesion segmentation mask boundary output in step 5 as the central reference, a tubular analysis band is constructed around the complete lesion outline. The single-sided extension width of the tubular analysis band is set to four to six voxel units. This width range is suitable for the conventional coverage range of the transition area of ​​the lesion boundary in clinical magnetic resonance imaging. It can completely encompass the parameter mutation range between the lesion and normal tissue, while avoiding invalid calculations in large homogeneous areas. Within the constructed tubular analysis band, the spatial normal direction of the tissue intrinsic parameter map is solved at each boundary position. The multidimensional intrinsic parameter change sequence of continuous pixels is extracted along the normal direction to form a parameter distribution profile at each position, which fully reflects the gradual change law of tissue parameters on both sides of the boundary.

[0119] Stable, distortion-free parameter profiles adjacent to the current profile are selected as benchmark samples. A linear interpolation algorithm is used to generate a predicted profile that conforms to the gradual change law of normal tissue parameters. The interpolation smoothing coefficient is fixed at 0.85, which is a common standard parameter for medical image profile fitting. The parameter distribution variance of local pixel groups within the tubular analysis zone is statistically analyzed. Based on the local variance value, a parameter anomaly threshold is dynamically set. The anomaly threshold is adaptively adjusted in the positive direction with the local variance, and the variance value range is controlled between 0.02 and 0.15 to adapt to the parameter fluctuation characteristics of different tissue regions. The actual extracted parameter profile and the interpolated predicted profile are compared point by point for deviation. Pixels corresponding to profiles with parameter deviation exceeding the dynamic anomaly threshold are uniformly marked as tissue parameter artifact pixels, completing the accurate screening and location of artifacts in the global boundary region.

[0120] Step 62: For the marked artifact pixels, search for samples with matching normal distances and expected tissue categories in the non-artifact regions, extract their intrinsic parameter vector centroids as restoration candidate values, project the restoration candidate values ​​onto the common parameter manifold to obtain manifold-regularized restoration values, and write the restoration values ​​into the artifact pixels to obtain the corrected tissue intrinsic parameter map. The specific operations are as follows:

[0121] For all labeled artifact pixels, sample retrieval is carried out in the non-artifact effective region of the current tissue intrinsic parameter map. Valid pixel samples with consistent normal spatial distance with artifact pixels and matching tissue categories corresponding to the segmentation results are selected first. The normal distance matching error is controlled within 0.1 voxel units to ensure the consistency of spatial tissue attributes between the retrieved samples and artifact locations. Summarize the retrieved valid samples of the same type, extract the multidimensional intrinsic parameter vectors corresponding to all samples, calculate the spatial centroid of the parameter vectors, and use the centroid parameter as the initial restoration candidate value for the artifact region to avoid random fluctuation errors of single-point sample parameters. Map the obtained restoration candidate values ​​to the cross-center common parameter manifold surface constructed in step 2. Utilize the cross-center general tissue parameter distribution constraint characteristics inherent in the common parameter manifold to perform regularization correction on the initial restoration candidate values, eliminate single-sample specific parameter biases, and obtain manifold regularized restoration values ​​that conform to the global tissue parameter distribution law. Write the regularized restoration values ​​pixel by pixel to the parameter positions of the original artifact pixels, replace the original distortion and abnormal parameters, and complete the parameter correction of all artifact regions. Finally, obtain the corrected tissue intrinsic parameter map with uniform global parameter distribution, no artifact distortion, and close to the characteristics of real human tissue.

[0122] Step 63: Difference is performed between the corrected tissue intrinsic parameter map and the uncorrected tissue intrinsic parameter map corresponding to the marked artifact pixel positions to obtain the artifact component map. Spatial cross-correlation identification is performed between the artifact component map and the obtained latent variable map of acquisition parameters. The components related to the artifact components in the latent variable map of acquisition parameters are extracted as leakage artifact sources. The contribution weight of the leakage artifact sources in the synthesis of the magnetic resonance signal equation is adjusted. The specific operation is as follows:

[0123] The corrected intrinsic tissue parameter map is compared with the original intrinsic tissue parameter map before artifact correction using a voxel-by-voxel difference operation. The difference in parameters at corresponding locations constitutes the artifact component map. This artifact component map can accurately and individually isolate all abnormal parameter components caused by acquisition parameter leakage, fully preserving the spatial distribution and numerical characteristics of artifacts. A global spatial cross-correlation analysis is then performed on the generated artifact component map and the acquisition parameter latent variable map obtained in step 1. The feature correlation coefficients of each region in the two types of images are calculated, with a correlation coefficient threshold set to 0.6. This threshold is a common standard for identifying the correlation of medical image features, and latent variable regions with correlation coefficients higher than the threshold are selected. These highly correlated acquisition parameter latent variable components are the leakage artifact sources that cause parameter artifacts, and the corresponding acquisition parameter components have an over-contribution problem in the signal synthesis process. Based on the correlation strength of the leakage artifact sources, the contribution weight of the corresponding acquisition parameter components in the magnetic resonance signal equation image synthesis process is adaptively reduced. The weight adjustment step size is controlled within the range of 0.03 to 0.08, ensuring a stable and controllable adjustment range in each iteration, avoiding parameter distortion caused by sudden weight changes. By continuously adjusting the weights, the ineffective leakage of latent variables of the acquired parameters into the intrinsic parameters of the tissue is suppressed, the probability of artifact generation in the subsequent cross-center image decoupling process is reduced, and the adaptive iterative optimization of the image decoupling and parameter generation modules is realized, thereby continuously improving the cross-center adaptation stability of the overall method.

[0124] Example 2

[0125] Please see Figure 4 Based on Example 1, this example provides a deep learning-based multimodal medical image lesion automatic segmentation system, including:

[0126] The intrinsic decoupling module is used to decouple cross-center multimodal magnetic resonance images into tissue intrinsic parameter maps and acquisition parameter latent variable maps, and achieves decoupling through voxel fidelity constraints between synthetic images and original images based on magnetic resonance signal equations;

[0127] The anchor point construction module is used to extract tissue intrinsic parameter fingerprints of lesions and normal tissues by utilizing lesion annotations on the tissue intrinsic parameter map, and to construct a cross-center parameter fingerprint anchor point dictionary.

[0128] The pseudo-label generation module is used to retrieve the initial pseudo-mask by using the cross-center parameter fingerprint anchor point dictionary. The initial pseudo-mask boundary converges to the parameter abrupt change position by using the parameter gradient of the organization intrinsic parameter map as the driving force to obtain the pseudo-label.

[0129] The mapping construction module is used to construct a mapping from the tissue intrinsic parameter map to the lesion segmentation result using the tissue intrinsic parameter map and the obtained pseudo-annotations. The mapping uses the tissue intrinsic parameters of each pixel in the tissue intrinsic parameter map as the decision basis.

[0130] The segmentation output module is used to acquire multimodal magnetic resonance images of the target center, obtain tissue intrinsic parameter maps after decoupling, and input the tissue intrinsic parameter maps into the mapping to obtain lesion segmentation results;

[0131] The artifact correction module is used to correct the obtained tissue intrinsic parameter map based on the lesion segmentation results, so as to suppress tissue intrinsic parameter artifacts that are inconsistent with the segmentation results.

[0132] The above description is merely a preferred embodiment of the present invention; however, 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 its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic segmentation of lesions in multimodal medical images based on deep learning, characterized in that, include: Step 1: Acquire cross-center unlabeled multimodal magnetic resonance images, perform dual-channel asymmetric encoding and decoding based on the magnetic resonance signal equation, decouple the images into latent variables of acquisition parameters and tissue intrinsic parameter maps, and obtain the decoupled tissue intrinsic parameter maps by combining the latent variables of acquisition parameters and tissue intrinsic parameter maps with the voxel fidelity constraint of the synthesized images and the original images. Step 2: Using the tissue parameters of the labeled lesion region and normal tissue in the tissue intrinsic parameter map, construct an anchor point dictionary containing the parameter fingerprints of lesions and normal tissue. The anchor point dictionary is established based on the manifold topological continuity of the intrinsic parameter maps of different centers. Step 3: Use the cross-center parameter fingerprint anchor point dictionary to search and obtain the initial pseudo mask. Use the parameter gradient of the organization intrinsic parameter map as the driving force to make the boundary of the initial pseudo mask converge to the parameter abrupt change position to obtain the pseudo label. Step 4: Use pseudo-labels to construct a mapping relationship with the tissue intrinsic parameter map as input and the lesion segmentation result as output. The mapping relationship uses the tissue relaxation parameter as the decision feature. Step 5: Obtain multimodal magnetic resonance images of the target center, obtain the intrinsic tissue parameter map of the target center through decoupling, and obtain the lesion segmentation result using the mapping relationship; Step 6: Using the obtained lesion segmentation results as a shape prior, apply shape consistency constraints with the shape prior to the decoupled output of the tissue intrinsic parameter map to correct the tissue intrinsic parameter map.

2. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 1, characterized in that, Step 1 includes: The cross-center multimodal magnetic resonance images are output as tissue intrinsic parameter maps and acquisition parameter latent variable maps through the tissue channel and acquisition channel, respectively. The images are synthesized based on the magnetic resonance signal equation and subjected to voxel fidelity constraints to achieve decoupling. A center-specific acquisition offset manifold is constructed based on the latent variable map of the acquisition parameters. The latent variable map of the acquisition parameters of the sample to be processed is normalized on the manifold to obtain a standardized latent variable map. The tissue intrinsic parameter map and the standardized latent variable map are synthesized into a virtual image by the magnetic resonance signal equation and compared with the original image to obtain the residual. The residual is used to inversely compensate the tissue intrinsic parameter map.

3. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 2, characterized in that, Step 2 includes: On the intrinsic parameter map of the tissue, pixels with parameter gradient exceeding the threshold are removed based on the lesion label, and candidate detection points in the low gradient stable region are retained. The candidate detection points are clustered based on the intrinsic parameter vector, and the cluster center is taken as the parameter detection point and assigned a tissue semantic label. The intrinsic parameter vectors of the parameter detection points of cross-center lesions and normal tissues are used to construct a point cloud. A nearest neighbor graph is constructed and iteratively diffused and smoothed until convergence to obtain a common parameter manifold. On the manifold, the density peak points of the detection points of each tissue class are searched and proposed as anchor parameters. An anchor dictionary is constructed from the anchor parameters and their semantic labels. The difference in average proximity of each anchor parameter in the anchor dictionary to the lesion-type anchor and the normal tissue-type anchor is calculated as the contrast score. Based on the contrast score, the anchor dictionary is reorganized into first-level anchors with high contrast scores and second-level anchors with medium contrast scores. The second-level anchors participate in the response calculation of the pixel based on the proximity relationship between the pixel and the dominant region of the first-level anchor of a certain category in the parameter space.

4. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 3, characterized in that, Step 3 includes: Using the first-level anchors of the hierarchical anchor dictionary, perform parametric distance matching on the tissue intrinsic parameter map, mark high-confidence candidate points, and for the remaining pixels in the pending state, activate the second-level anchors of the corresponding category to participate in response calculation based on their proximity to the dominant region of the first-level anchor of a certain category in the parameter space, and generate the initial lesion response map. On the intrinsic parameter map, the gradient structure tensor of the multidimensional intrinsic parameter vector is calculated. Based on the relative relationship between the maximum and minimum eigenvalues ​​of the gradient structure tensor, a local metric for the width of the parameter transition band is determined, and a spatial adaptive evolution velocity map is constructed.

5. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 4, characterized in that, Step 3 also includes: Using the closed surface intercepted by the given response threshold on the initial lesion response map as the boundary of the zero level set, the hierarchical anchor point dictionary is retrieved at the boundary points, and the contrast score of the nearest neighbor anchor point is converted into a local driving force. This is then superimposed point by point with the spatial adaptive evolution velocity map to form a dual driving field. The driving boundary converges at the local maximum of the parameter gradient along the boundary evolution direction, thus obtaining pseudo-labels. For points on the converged pseudo-label boundary, check whether the magnitude of the parameter gradient in the intrinsic parameter map is the maximum value in the local neighborhood of the point. For boundary points that do not reach the maximum value in the local neighborhood, adjust them along the normal direction to the position of the gradient maximum. Preserve the topology through surface interpolation to obtain regular pseudo-labels.

6. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 5, characterized in that, Step 4 includes: The regular pseudo-labels are superimposed with the intrinsic parameter map of the tissue. The intrinsic parameter vectors of pixels marked as lesions and pixels marked as normal tissue are extracted by pseudo-labels. These vectors form the lesion parameter set and the normal parameter set, respectively. The lesion parameter set and the normal parameter set are then subjected to density subdivision and recursive separation in the multidimensional intrinsic parameter space. The intrinsic parameter dimensions are traversed and a single-dimensional threshold for separating the lesion parameter set and the normal parameter set is selected to generate a multidimensional threshold logical path tree. The decision accuracy of each branch in the multidimensional threshold logic path tree is verified by using the intrinsic parameter maps of other cross-center samples and known pseudo-labels. Branches with accuracy lower than the preset cross-center consistency threshold are pruned, and equivalent replacement continuous threshold conditions are merged, retaining the mapping composed of intrinsic parameter thresholds.

7. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 6, characterized in that, Step 5 includes: Acquire multimodal magnetic resonance images of the target center, normalize the decoupled acquisition parameter latent variable map into a standardized acquisition latent variable map on the acquisition offset manifold, synthesize the image with the decoupled tissue intrinsic parameter map using the magnetic resonance signal equation, and compare it with the original image to correct the tissue intrinsic parameter map with residuals; The modified intrinsic parameter map of the tissue is used to determine the initial segmentation mask through a multi-dimensional threshold logic path tree. The boundary pixels of the initial segmentation mask are extracted, and the minimum parameter distance between the intrinsic parameter vector of the mask and the lesion anchor point and the normal tissue anchor point is retrieved from the hierarchical anchor point dictionary. The boundary pixel category is adjusted based on the minimum parameter distance and contrast score.

8. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 7, characterized in that, Step 6 includes: A tubular analysis zone is constructed based on the boundary of the lesion segmentation mask. Within the tubular analysis zone, the parameter profile of the intrinsic tissue parameter map along the normal direction is calculated. The deviation of each parameter profile is compared with the interpolated prediction profile of the adjacent reliable profile. Pixels corresponding to the profiles whose deviation exceeds the abnormal threshold set based on the local parameter variance within the tubular analysis zone are marked as artifact pixels. For the marked artifact pixels, search for samples with matching normal distance and expected tissue category in the non-artifact region, extract the centroid of its intrinsic parameter vector as restoration candidate value, project the restoration candidate value onto the common parameter manifold to obtain manifold regularized restoration value, write the restoration value into the artifact pixel to obtain the corrected tissue intrinsic parameter map.

9. The method for automatic segmentation of multimodal medical image lesions based on deep learning according to claim 8, characterized in that, Step 6 also includes: The intrinsic parameter map of the corrected tissue is compared with the intrinsic parameter map of the tissue before correction corresponding to the marked artifact pixel positions to obtain the artifact component map. The artifact component map is then spatially cross-correlation identified with the obtained latent variable map of the acquisition parameters. The components related to the artifact components in the latent variable map of the acquisition parameters are extracted as leakage artifact sources, and the contribution weight of the leakage artifact sources in the synthesis of the magnetic resonance signal equation is adjusted.

10. A deep learning-based automatic segmentation system for multimodal medical image lesions, applied to any one of the deep learning-based automatic segmentation methods for multimodal medical image lesions as described in claims 1-9, characterized in that, include: The intrinsic decoupling module is used to decouple cross-center multimodal magnetic resonance images into tissue intrinsic parameter maps and acquisition parameter latent variable maps, and achieves decoupling through voxel fidelity constraints between synthetic images and original images based on magnetic resonance signal equations; The anchor point construction module is used to extract tissue intrinsic parameter fingerprints of lesions and normal tissues by utilizing lesion annotations on the tissue intrinsic parameter map, and to construct a cross-center parameter fingerprint anchor point dictionary. The pseudo-label generation module is used to retrieve the initial pseudo-mask by using the cross-center parameter fingerprint anchor point dictionary. The initial pseudo-mask boundary converges to the parameter abrupt change position by using the parameter gradient of the organization intrinsic parameter map as the driving force to obtain the pseudo-label. The mapping construction module is used to construct a mapping from the tissue intrinsic parameter map to the lesion segmentation result using the tissue intrinsic parameter map and the obtained pseudo-annotations. The mapping uses the tissue intrinsic parameters of each pixel in the tissue intrinsic parameter map as the decision basis. The segmentation output module is used to acquire multimodal magnetic resonance images of the target center, obtain tissue intrinsic parameter maps after decoupling, and input the tissue intrinsic parameter maps into the mapping to obtain lesion segmentation results; The artifact correction module is used to correct the obtained tissue intrinsic parameter map based on the lesion segmentation results, so as to suppress tissue intrinsic parameter artifacts that are inconsistent with the segmentation results.

Images