Alzheimer's disease early warning method based on white matter lesion omics characteristics

Abstract

The invention discloses an Alzheimer's disease early warning method based on white matter lesion omics characteristics, and relates to the field of wisdom medicines.The method comprises the steps that magnetic resonance imaging data of a historical subject in the period from the mild cognitive impairment period to the period before diagnosis of Alzheimer's disease are obtained, and manual labeling of white matter and white matter lesion areas is carried out; a manual annotation data set is obtained; training a deep learning model for white matter lesion recognition based on the manual annotation data set; inputting to-be-identified magnetic resonance imaging data into the deep learning model, and extracting lesion features of the white matter; performing standardization and feature alignment on the extracted lesion features, and inputting the lesion features into a deep clustering model to form clustering results for different white matter lesion feature types; and an early risk assessment model is constructed based on the clustering result and the Alzheimer's disease transformation risk tag corresponding to the clustering result, and the Alzheimer's disease transformation risk level of the subject is output, so that the problems that multiple lesion features are difficult to quantify and details are difficult to identify are solved.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of smart medical treatment, and more particularly, the present application relates to an early warning method for Alzheimer's disease based on white matter lesion omics characteristics. BACKGROUND

[0002] Alzheimer's disease (AD) is one of the common types of neurodegenerative dementia. Its early identification and etiological diagnosis in the clinic have long relied on clinical symptoms, cerebrospinal fluid puncture, or high-cost, high-radiation imaging methods, which have problems such as high delay rate of diagnosis, poor acceptability, and low screening coverage. In recent years, domestic and foreign research has gradually focused on white matter lesions (White Matter Hyperintensities, WMH) as a potential imaging marker for early dementia, which shows extensive early performance characteristics in different types of dementia (including Alzheimer's disease, vascular dementia, subcortical dementia, etc.).

[0003] White matter lesions can affect the signal transduction speed of neural networks, thereby impairing the patient's memory, recognition, calculation, and reaction ability, and other multiple cognitive dimensions. Existing research has shown that white matter lesions distributed in the posterior circulation region are closely related to hippocampal atrophy, Aβ protein deposition, and memory decline, and other typical Alzheimer's disease pathological changes. Posterior circulation white matter lesions are considered the second largest pathogenic factor after Aβ, and may occur before Aβ deposition, and there is a potential interaction between the two in the pathological mechanism.

[0004] Although the magnetic resonance imaging (Magnetic Resonance Imaging, MRI) imaging technology and artificial intelligence assisted analysis have been continuously developed in recent years, there is still a significant analysis bottleneck in modeling the multi-dimensional characteristics of white matter lesions. In the team's previous research on automatic segmentation of white matter lesions, it was found that the lesions have significant heterogeneity in volume, distribution, signal intensity, boundary morphology, and accompanying lesions. Some studies have shown that specific distribution patterns such as posterior circulation concentrated white matter lesions are closely related to AD, while anterior circulation multiple and symmetric lesions are more common in hypertensive vascular dementia or amyloid angiopathy, but other important features such as boundary morphology and gray intensity cannot be quantified by artificial means, and the understanding of the structure-cause correlation mechanism is still weak.

[0005] With the rise of deep learning and other artificial intelligence technologies, a small number of studies have begun to explore the relationship between boundary morphology, signal intensity, and dementia evolution risk. For example, the signal intensity is proposed as an indicator of the severity of white matter lesions in related technologies, and attempts are made to use it for dementia subtype discrimination. However, how to establish an analysis system that can comprehensively learn the multiple image features of white matter and correlate the causes is still one of the core problems of current artificial intelligence medical image analysis.

[0006] Therefore, it is urgent to develop a comprehensive method integrating MRI image segmentation, multi-modal feature extraction, deep temporal modeling and clustering analysis to realize accurate identification of multiple features of white matter lesions, structural classification and transformation risk assessment. SUMMARY

[0007] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiments section. The summary section of the present application does not mean to attempt to limit the key features and essential technical features of the claimed technical solutions, nor to determine the protection scope of the claimed technical solutions.

[0008] In a first aspect, the present application proposes an early warning method for Alzheimer's disease based on white matter lesion omics characteristics, comprising: Obtaining magnetic resonance imaging data of historical subjects during the mild cognitive impairment period to the diagnosis of Alzheimer's disease, and performing artificial labeling of white matter and white matter lesion regions to obtain an artificial labeling data set; Training a deep learning model for white matter lesion recognition based on the artificial labeling data set, wherein the deep learning model uses convolution modules at the shallow layer to extract features from T1w images and FLAIR images, and integrates multi-modal information through a fusion module, the deep learning model uses lightweight parallel vision Mamba modules at the deep layer to model deep features and perform global perception on the FLAIR images, and the deep learning model designs a dual-task training structure for region segmentation and boundary detection to improve segmentation accuracy; Inputting the above-mentioned magnetic resonance imaging data to be identified into the above-mentioned deep learning model to extract lesion features of white matter; After standardizing and aligning the extracted lesion features, inputting them into a deep clustering model to form clustering results for different white matter lesion feature types; Based on the clustering results and their corresponding Alzheimer's disease transformation risk labels, an early risk assessment model is constructed, and the Alzheimer's disease transformation risk grade of the subject is output.

[0009] In a feasible implementation, the magnetic resonance imaging data images are obtained by a 3.0 Tesla scanning system, and 28-32 layer sequence images with a slice thickness of 5mm are used.

[0010] In a feasible implementation, the white matter and lesion labeling is independently labeled by at least two doctors with neuroimaging experience, and the segmentation results are subjected to statistical consistency test; if there is a difference, a third doctor reviews.

[0011] In an implementation, the deep learning model adopts a fusion module to fuse the multi-modal feature information of the T1w image and the FLAIR image in the channel dimension and the spatial dimension at the shallow layer of the encoder.

[0012] In an implementation, the lightweight parallel visual Mamba module is used to replace the convolutional neural network in the deep structure of the encoder and the decoder to achieve more efficient deep feature extraction. The visual Mamba module includes a VSS structure, and the VSS structure is used to model the complex spatial dependency and global context information in the image. The VSS structure includes a first branch for basic feature modeling and a second branch for complex feature modeling, the first branch includes a linear layer and an activation function, and the second branch includes a linear layer, a deep convolutional layer, an activation function, a two-dimensional selective scanning module, and a normalization layer.

[0013] In an implementation, the deep learning model adopts a double-task training structure, and the loss function of the deep learning model in the training process includes a weighted combination of the loss of the segmentation task and the boundary detection.

[0014] In an implementation, the deep clustering model includes a multi-layer perceptron and a Softmax probability mapping module, and the deep clustering model adopts an unsupervised training method to map the features to a clustering space.

[0015] In an implementation, the lesion features of the white matter of the brain include lesion volume, lesion gray signal intensity, lesion distribution pattern and regional relationship, and lesion boundary morphology features, and the lesion boundary morphology features include boundary sharpness, irregularity, gray contrast, and fractal dimension.

[0016] In an implementation, the risk assessment model adopts an LSTM network to encode the sequenced omics features and is used to output the Alzheimer's disease conversion risk level.

[0017] In summary, the above invention has significant beneficial technical effects in terms of early risk identification of Alzheimer's disease, white matter lesion feature extraction and characterization ability, model structure design, and disease progression prediction, etc. First, the invention innovatively uses magnetic resonance data of patients from Mild Cognitive Impairment (MCI) to Alzheimer's disease diagnosis period as the analysis object, and constructs a research sample with time sequence continuity. This data acquisition strategy across key stages of disease evolution is different from the traditional static case sample processing method, which can more truly reflect the evolution trend of white matter lesions and provide high-value time dimension information for subsequent model learning. Especially in MRI data acquisition, the invention uses a 3.0 Tesla (T) high-field scanning system and uniform slice parameters (5mm, 28-32 layers) to ensure image quality consistency and structural layer coverage, significantly improving the accuracy of feature segmentation and identification. Second, in terms of data labeling, the invention constructs a three-labeling mechanism of "two-doctors independent + consistency test + expert review", which effectively reduces the subjective differences of artificial labeling, improves the reliability and consistency of labeled data, and provides high-quality labels for supervised learning models, solving the reliability problem caused by traditional single labeling and having stronger clinical reproducibility and promotion value. In terms of model structure design, the invention combines multiple advanced deep neural network technologies to construct a white matter lesion segmentation model with lightweight and high expression ability. Among them, the encoder shallow layer simultaneously extracts features from T1w and FLAIR two medical images, and introduces a temporal fusion attention module (TFAM) for multi-modal feature fusion to enhance the collaborative expression ability between different MRI features, breaking through the bottleneck of traditional single-modal processing strategy in multi-channel image processing. In the deep structure of the encoder and decoder, a lightweight parallel visual Mamba module is used instead of the traditional convolutional neural network to realize more efficient deep feature extraction. The visual Mamba module includes a VSS (Visual State Space) structure, which models the basic semantic information and complex boundary features in a parallel double-branch manner, fully considering the diversity of white matter lesions in morphology, signal, and texture. In addition, the invention designs a multi-task training mechanism (MTL) to couple the regular segmentation task with the boundary detection task, significantly improving the recognition ability of the model for edge blur and boundary adhesion lesions, thereby solving the technical shortcoming of unclear white matter lesion boundary recognition and large error in the prior art.In feature extraction and modeling, the system extracts multi-dimensional white matter lesion features from the model output, including volume, signal intensity, spatial distribution pattern, boundary shape, and accompanying lesion information, covering multiple key indicators that cannot be systematically described by traditional manual annotation methods, improving the completeness and precision of lesion representation. Through standardization and feature alignment, the consistency and comparability of data between different dimensions are further improved. In lesion feature classification, the deep clustering model proposed in the application uses unsupervised training method and does not rely on artificial grouping standard. Through the combination of Multi-Layer Perceptron (MLP) and Softmax module, it can automatically find the implicit typing structure of lesions in structure and morphology in high-dimensional feature space, effectively making up for the shortcomings of existing methods in recognizing lesion heterogeneity. Finally, in risk prediction, the application introduces LSTM (Long Short-Term Memory) model to sequence model the features, so that the temporal dynamic changes of lesion features can be effectively modeled and quantified, supporting quantitative evaluation of the risk of individuals converting to Alzheimer's disease in the future. Compared with traditional static regression or classification methods, LSTM is more suitable for identifying nonlinear evolution patterns and mutation trends, making the risk assessment results more close to the real clinical process. In summary, the application solves the core bottlenecks in the prior art such as difficulty in quantifying multiple white matter lesion features, difficulty in identifying details, difficulty in associating causes, and difficulty in modeling temporal risk, significantly improving the accuracy, automation, and generalizability of early screening and cause identification of Alzheimer's disease, and has outstanding medical practicality, engineering feasibility, and social and economic value.

[0018] The Alzheimer's disease early warning method based on white matter lesion features proposed in the application, other advantages, objectives and features of the application will be partially embodied through the following description, and some will be understood by those skilled in the art through research and practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0019] Various other advantages and benefits will become clear to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered limitations thereon. Moreover, the same reference numbers in the drawings indicate the same components throughout the several drawings. In the drawings: Figure 1 A flowchart of an Alzheimer's disease early warning method based on white matter lesion features provided by an embodiment of the application is shown in the figure; Figure 2 A segmentation model based on multi-modal fusion provided by an embodiment of the application is shown in the figure; Figure 3A principle diagram of double task model training provided by an embodiment of the present application is provided. Figure 4 A model training loss function diagram provided by an embodiment of the present application is provided. Figure 5 A clustering model principle diagram provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0020] The terms "first", "second", "third", "fourth" and the like in the description, claims, and drawings of the present application (if any) are used for distinguishing between similar objects, and do not necessarily have to appear in a given order or succession. It is to be understood that the data used with these terms so distinguished, could be interchanged under appropriate circumstances, such that the embodiments described herein, can be practiced in other than the described order. Additionally, the terms "comprising", "including", "containing", and "having" and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, product, or apparatus that comprises, includes, contains, or has an element or a list of elements can include additional elements not expressly listed or inherent to such process, method, system, product, or apparatus. The following description will be made with reference to the accompanying drawings of the present application, which are provided to assist in understanding the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments.

[0021] Please refer to Figure 1 A flow principle diagram of an early warning method for Alzheimer's disease based on leukoaraiosis characteristics provided by an embodiment of the present application is provided, which can specifically include: S110, acquiring magnetic resonance imaging data of a historical subject during a mild cognitive impairment period to before an Alzheimer's disease diagnosis, and performing artificial labeling of brain white matter and leukoaraiosis regions to obtain an artificial labeling data set; S120, training a deep learning model for leukoaraiosis recognition based on the artificial labeling data set, the deep learning model using a convolution module at a shallow layer to perform feature extraction on a T1w image (T1-weighted image) and a FLAIR image (Fluid-Attenuated Inversion Recovery), and integrating multi-modal information through a fusion module, the deep learning model using a lightweight parallel vision Mamba module at a deep layer to perform deep feature modeling and global perception on the FLAIR image, and the deep learning model designing a double task training structure of region segmentation and boundary detection to improve segmentation accuracy; S130, inputting the to-be-recognized magnetic resonance imaging data into the deep learning model to extract leukoaraiosis features. S140, input the extracted lesion features into the deep clustering model after standardization and feature alignment, to form clustering results for different white matter lesion feature types; S150, based on the clustering results and their corresponding Alzheimer's disease conversion risk labels, an early risk assessment model is constructed, and the subject's Alzheimer's disease conversion risk level is output.

[0022] For example, first, in step S110, the system acquires MRI data collected from a large number of historical subjects in the key stage from mild cognitive impairment to the pre-diagnosis stage of Alzheimer's disease. To ensure the accuracy and annotatability of the data, these image data are derived from a 3.0T magnetic resonance system, using a standard 5mm slice thickness, covering 28 to 32 layers. Subsequently, two experienced neurology or neuroimaging professionals independently perform manual segmentation labeling on the white matter and white matter lesion regions. The system evaluates the segmentation consistency of the two physicians, and if there is a significant difference, a third expert reviews and forms the final labeling result. Through the above process, a high-quality manually annotated dataset is established, laying the foundation for subsequent model training.

[0023] In step S120, a white matter lesion recognition model is trained based on the manually annotated dataset. The model integrates multiple deep learning structure innovations, including: using TFAM to perform multi-modal fusion of T1w and FLAIR features of two medical images, and performing cross-channel and spatial dimension reinforcement fusion of different sequence image features; using a lightweight parallel visual Mamba module to replace the traditional convolutional neural network in the deep structure of the encoder and decoder, and the VSS structure of the Mamba module processes MRI image features in multiple levels, one branch focuses on shallow feature extraction, and the other branch integrates deep convolution and selective scanning modules to extract complex lesion boundary and texture features. At the same time, the model also constructs a boundary detection loss function, which focuses on optimizing the segmentation accuracy of the lesion boundary region; and introduces an MTL structure, which adds a boundary detection task based on the conventional segmentation task, thereby improving the recognition ability of the model for adhesion boundaries or fuzzy boundaries.

[0024] In step S130, the MRI image of the subject to be recognized is input into the above trained deep model, and the system automatically extracts multiple features related to white matter lesions, including lesion volume, signal intensity, spatial distribution, boundary morphology, and accompanying lesion information. The process is completed by forward inference of the model, and the feature set is output in a structured form.

[0025] Next, in step S140, the extracted multi-dimensional lesion features are standardized and time-series feature aligned to ensure consistent expression of features from different sources or dimensions in a unified semantic space. These features are then input into a deep clustering model composed of MLP and Softmax probability mapping modules, which clusters all subject samples in an unsupervised manner and outputs white matter lesion type labels or omics feature clusters. These clusters reflect the natural grouping differences of lesions in structure, morphology, or evolutionary trends.

[0026] Finally, in step S150, the system maps the clustering results to known Alzheimer's conversion labels to build an early risk assessment model. This assessment model models the feature sequence based on the LSTM structure, considering the development trend and morphological features of the lesion, and outputs the current subject's Alzheimer's conversion risk level as a clinical warning basis. This risk level can be refined into low, medium, and high risk levels, etc., to facilitate doctors to make further examination or intervention decisions.

[0027] In summary, the above invention has significant beneficial technical effects in terms of early risk identification of Alzheimer's disease, white matter lesion feature extraction and characterization ability, model structure design, and disease progression prediction, etc. First, the invention innovatively uses magnetic resonance data of patients from the mild cognitive impairment period to the confirmed period of Alzheimer's disease as the analysis object, constructing a research sample with time sequence continuity. This data acquisition strategy, which spans the key stages of disease evolution, differs from the traditional static case sample processing method, and can more truly reflect the evolution trend of white matter lesions, providing high-value time dimension information for subsequent model learning. Especially in MRI data acquisition, the invention uses a 3.0T high-field scanning system and uniform slice parameters (5mm, 28-32 layers) to ensure image quality consistency and structural layer coverage, significantly improving the accuracy of feature segmentation and identification. Second, in terms of data labeling, the invention constructs a three-fold labeling mechanism of "two doctors independent + consistency test + expert review", effectively reducing artificial subjective differences and improving the reliability and consistency of labeled data, providing high-quality labels for supervised learning models and solving the reliability problem caused by traditional single doctor labeling, with stronger clinical reproducibility and promotional value. In terms of model structure design, the invention combines multiple advanced deep neural network technologies to construct a white matter lesion segmentation model with lightweight and high expression ability. Among them, the encoder shallow layer simultaneously extracts features from T1w and FLAIR two medical images, and introduces TFAM for multi-modal feature fusion to enhance the collaborative expression ability between different MRI features, breaking through the bottleneck of traditional single-modal processing strategy in multi-channel image processing. In the deep structure of the encoder and decoder, the lightweight parallel visual Mamba module is used instead of the traditional convolutional neural network to realize more efficient deep feature extraction. The visual Mamba module includes the VSS structure, which models the basic semantic information and complex boundary features in a parallel double-branch manner, fully considering the diversity of white matter lesions in morphology, signal, and texture. In addition, the invention couples the regular segmentation task with the boundary detection task through the design of MTL, significantly improving the recognition ability of the model for edge blur and boundary adhesion lesions, thereby solving the technical shortcoming of unclear white matter lesion boundary recognition and large error in the prior art. In terms of feature extraction and modeling, the system extracts multi-dimensional white matter lesion omics features from the model output, including volume, signal intensity, spatial distribution pattern, boundary morphology, and accompanying lesion information, covering key indicators that are difficult to fully characterize by artificial labeling. Through standardization and feature alignment, the consistency and comparability of data between different dimensions are further improved.In terms of lesion feature classification, the deep clustering model proposed by the present application adopts an unsupervised training method and does not rely on artificial grouping standards. Through the combination of MLP and Softmax module, the model can automatically discover the implicit typing structure of lesions in structure and morphology in a high-dimensional feature space, effectively making up for the shortcomings of existing methods in recognizing lesion heterogeneity.

[0028] Finally, in terms of risk prediction, the present application introduces an LSTM model to sequence model the omics features, so that the temporal dynamic changes of lesion features can be effectively modeled and quantified, supporting quantitative evaluation of the risk of individuals in the future to develop into Alzheimer's disease. Compared with traditional static regression or classification methods, LSTM is more suitable for identifying nonlinear evolution patterns and mutation trends, making the risk assessment results more close to the real clinical process. In summary, the present application solves the core bottlenecks in the prior art such as difficulty in quantifying multiple features of white matter lesions, difficulty in identifying details, difficulty in correlating causes, and difficulty in modeling time series risks, significantly improving the accuracy, automation, and generalizability of early screening and cause identification of Alzheimer's disease, and has outstanding medical practicality, engineering feasibility, and social and economic value.

[0029] In a feasible implementation, the magnetic resonance imaging data image is obtained by a 3.0 scanning system, and a 28-32 layer sequence image with a 5mm slice thickness is used.

[0030] For example, the magnetic resonance imaging data image is obtained by a 3.0T scanning system, and a 28-32 layer sequence image with a 5mm slice thickness is used. This implementation is mainly used to ensure the imaging quality and spatial resolution of the white matter and its lesion area, and to meet the accuracy requirements of subsequent manual annotation and deep learning feature extraction.

[0031] Specifically, the 3.0T magnetic resonance scanning system is a high-field MRI device with higher signal-to-noise ratio (SNR), which can provide clearer and more detailed brain structure images. Compared with the conventional 1.5T system, the 3.0T device can obtain higher resolution images in the same scanning time, which is particularly important for identifying small lesions, fuzzy boundary areas, and gray transition areas between lesions and normal white matter.

[0032] In addition, the present embodiment adopts a scanning scheme with a 5mm slice thickness, which can balance the clarity and overall coverage efficiency of the image in clinical routine use, ensuring spatial resolution in the vertical direction while avoiding data storage pressure and model training computational burden caused by too many layers.

[0033] Meanwhile, to cover the three-dimensional anatomical structure of the white matter region of the brain, the system sets the number of scanning layers to 28-32 layers, which can cover most of the brain parenchyma region from the frontal lobe to the occipital lobe, and ensure that the lesion characteristics can be completely captured on different layers. This layer configuration is universal in actual clinical scanning, and also facilitates subsequent image registration and standard space mapping operations.

[0034] In summary, the embodiment effectively ensures the balance of image data in terms of resolution, spatial coverage range and structural clarity by using a 3.0T magnetic resonance system, setting a 5mm slice thickness and a 28-32 layer scanning scheme, thereby providing a high-quality data basis for the accurate identification, feature extraction and subsequent clustering analysis of white matter lesions.

[0035] In a feasible implementation, the above-mentioned white matter and lesion labeling is independently labeled by at least two doctors with experience in neuroimaging, and the segmentation results are subjected to statistical consistency test; if there is a difference, a third doctor reviews.

[0036] For example, the labeling of the above-mentioned white matter and lesion regions is independently completed by at least two professional doctors with experience in neuroimaging reading, to ensure the accuracy and medical professional consistency of the labeling results. The main purpose of this embodiment is to improve the reliability of artificial labeling data, thereby providing high-quality training samples for deep learning models and reducing the error transmission caused by subjective bias.

[0037] Specifically, the system first imports the MRI image into the medical image workstation and distributes it to two professional doctors with a background in neuroimaging. They separately segment the white matter region and the white matter lesion region layer by layer without referring to each other's labeling results. This process uses a standard brain region anatomy template as a reference to ensure that the labeling standards between different doctors are consistent and reproducible.

[0038] After completing the preliminary labeling, the system automatically performs statistical consistency test on the segmentation results of the two doctors. Common consistency test methods include quantitative indicators such as Dice coefficient, Jaccard index or voxel overlap rate, which are used to evaluate the spatial overlap degree between the two labeling results. If the consistency result reaches the preset threshold (such as Dice coefficient greater than 0.85), the segmentation result is considered consistent, and the averaged or fused result is taken as the final labeling.

[0039] If the labeling of the two doctors has obvious differences and the consistency index is lower than the set threshold, the system will automatically mark the case as a "divergent sample" and submit it to a third doctor with more experience for review. The third doctor will comprehensively consider the segmentation schemes of the first two doctors, combine his own clinical experience, re-interpret the areas with ambiguous lesion boundaries and unclear signal gray scale transitions, and give the final confirmed standard labeling result.

[0040] Through the "three medical annotations, consistency inspection and difference review" process mechanism, the embodiment significantly reduces the error caused by subjective judgment difference in the manual segmentation process, ensures that the annotation result of each training sample has sufficient reliability and objectivity in the medical profession, and provides reliable reference for subsequent model training, verification and reasoning.

[0041] In a feasible embodiment, as shown in Figure 2 Figure 2 An overall structure schematic diagram of a lesion segmentation model based on multi-modal fusion provided by the embodiment of the application is provided. The deep learning model adopts a fusion module for fusing multi-modal feature information of the T1w image and the FLAIR image in the channel dimension and the spatial dimension at the shallow layer of the encoder.

[0042] For example, the application adopts an encoder-decoder architecture and innovatively introduces a dual-modal input path to extract features of T1w and FLAIR respectively, and then fuse the information of the two imaging modalities to improve the feature expression capability.

[0043] Specifically, the front end of the encoder is designed as two parallel branches, which respectively perform shallow feature modeling on the T1w and FLAIR images, and each uses a convolution module to extract modal features. The two features are then integrated by the fusion module to complete the integration of multi-modal features, and are transmitted to the back end of the encoder for deep semantic modeling.

[0044] In the embodiment, the fusion module can optionally adopt the existing TFAM for fusing multi-modal features in the channel dimension and the spatial dimension. The fusion module realizes joint coding of multi-modal features by establishing the context relationship between sequences, thereby enhancing the semantic perception ability of the model to the lesion area. It should be noted that the fusion module is prior art, and its specific structure and implementation manner can refer to related public documents, which will not be described here.

[0045] To further improve the recognition accuracy of the model, the embodiment also introduces a lightweight parallel visual Mamba module to replace the traditional convolutional neural network structure for deep modeling of the encoder and the decoder. The module can capture deeper spatial dependency relationships and global context information in the image, which is helpful for fine segmentation of lesions.

[0046] In addition, in order to balance the modeling needs of region segmentation and boundary information, the model designs a training strategy based on dual-task training, which enhances the perception ability of the model to the edge of the adhesion lesion through the cooperative optimization of the region segmentation task and the boundary detection task, thereby improving the overall segmentation performance.

[0047] ​In summary, the embodiment realizes richer multi-modal expression by constructing a dual-modal input path and fusing T1w and FLAIR imaging features, while maintaining model lightweight, and provides effective technical support for precise segmentation and boundary recognition of white matter lesions by combining deep feature extraction structure and dual-task learning mechanism.

[0048] In a feasible implementation, the aforementioned lightweight parallel visual Mamba module is used to replace the convolutional neural network in the encoder and decoder deep structure to realize more efficient deep feature extraction. The aforementioned visual Mamba module includes a VSS structure, and the VSS structure is used to model complex spatial dependency and global context information in an image. The VSS structure includes a first branch for basic feature modeling and a second branch for complex feature modeling, the first branch includes a linear layer and an activation function, and the second branch includes a linear layer, a deep convolutional layer, an activation function, a two-dimensional selective scanning module, and a normalization layer.

[0049] For example, the VSS structure includes two structure-complementary branches, which are respectively used for modeling basic image features and enhancing extraction of complex image details, so as to effectively capture multi-scale and multi-level information of white matter and its lesion regions.

[0050] Specifically, the first branch in the structure is relatively simple and is composed of a linear layer and an activation function, and mainly functions to linearly transform and nonlinearly map the input magnetic resonance image features, so as to extract basic semantic information and global structural features in the image. The branch as a lightweight path helps to quickly establish an initial representation of a feature space, reduces computational complexity, and provides a stable input basis for deep feature processing.

[0051] The second branch focuses on modeling complex visual details, and its structure is more rich, including a linear layer, a deep convolutional layer, an activation function, a two-dimensional selective scanning module, and a normalization layer. The deep convolutional layer is used to extract high-order features in the image, such as lesion boundaries, gray transitions, and texture heterogeneity. The two-dimensional selective scanning module realizes intensive perception of the region of interest by directionally and regionally paying attention to information in a spatial range, thereby improving the expression ability of the model to fine-grained structures of lesions. The normalization layer is used to standardize the distribution of intermediate features, and improve the stability and generalization ability of model training.

[0052] The module adopts a "shallow layer + deep layer" parallel strategy, so that the model can capture global semantics and basic morphological features in the first branch, and deeply depict complex lesion features such as boundary ambiguity and signal difference details in the second branch. The outputs of the two branches are fused in the subsequent module to improve the multi-dimensional expression ability and segmentation accuracy of the model for white matter lesion regions.

[0053] In summary, the present embodiment balances the computational efficiency and modeling capability by integrating the visual Mamba module with complementary structure, providing a solid network foundation for accurate identification of white matter lesions in complex neuroimaging.

[0054] In a feasible implementation manner, as shown in Figure 3 , Figure 3 is a schematic diagram of a double-task model training principle provided by the embodiment of the present application. The deep learning model adopts a double-task training structure.

[0055] For example, the present embodiment designs an innovative double-task training framework for the common adhesion lesions in white matter lesions. This framework not only includes the conventional lesion segmentation task, but also introduces a boundary detection task to improve the recognition ability of the model for the boundary of adhesion lesions.

[0056] Specifically, the double-task training framework includes two parallel tasks: one is the lesion segmentation task realized by the encoder-decoder structure, and the other is the boundary detection module task. The output of the last convolution layer of the encoder is first transmitted to the boundary detection module, which successively passes through a convolution layer, a ReLU activation function, and another convolution layer, and finally generates a lesion boundary map.

[0057] In a feasible implementation manner, as shown in Figure 4 , Figure 4 is a schematic diagram of a model training loss function provided by the embodiment of the present application. The loss function of the deep learning model in the training process includes the loss weighted combination of the segmentation task and the boundary detection task.

[0058] For example, in the segmentation task, there is a serious class imbalance problem in the white matter lesion segmentation task, and the present application focuses on the boundary of the segmented region to improve the segmentation accuracy of the boundary region. The loss function is as follows, is the region loss, is the boundary loss, is the weight parameter balancing the two losses.

[0059] In the boundary detection task, the labeled data is processed by a boundary label generation function to obtain the corresponding boundary labels. To evaluate the accuracy of the boundary detection task, the generated boundary map is compared with the boundary labels, and the boundary loss is calculated using a binary cross-entropy loss function with logits.

[0060] The model's overall loss function consists of two parts: segmentation task loss and boundary detection task loss. Through this dual-task training strategy, the model not only improves the segmentation accuracy of lesion regions but also significantly enhances its ability to perceive lesion boundaries, thereby providing more comprehensive and accurate auxiliary information for clinical diagnosis.

[0061] In one feasible implementation, such as Figure 5 As shown, Figure 5 This is a schematic diagram illustrating the principle of a clustering model provided in an embodiment of this application. The aforementioned deep clustering model includes a multilayer perceptron and a Softmax probability mapping module. The deep clustering model employs an unsupervised training method to map features to the clustering space.

[0062] For example, the deep clustering model mentioned above includes an MLP and a Softmax probability mapping module. This clustering model uses an unsupervised training method to map the extracted white matter lesion features to a potential clustering space, thereby achieving automatic classification of lesion types and structural recognition.

[0063] Specifically, the model's input consists of pathomic features extracted and standardized by a preceding segmentation model. These features include information across multiple dimensions, such as lesion volume, signal intensity, boundary morphology, and spatial distribution patterns. First, the model performs nonlinear transformations and high-order abstract representations of the input features using an MLP structure. The MLP comprises several fully connected layers, activation functions (such as ReLU), and regularization techniques like Dropout, aiming to fully capture the complex relationships and underlying structures between various features and compress them into discriminative low-dimensional embeddings.

[0064] Subsequently, these feature representations are passed to the Softmax probability mapping module. This module maps the feature representation of each sample to a probability distribution vector, representing the probability of that sample belonging to different cluster centers. Unlike traditional K-means hard clustering, Softmax outputs a soft assignment method, allowing each lesion sample to have a non-zero probability across multiple categories. This preserves the ambiguity and transitional nature of lesion features, better reflecting the continuous and non-linear characteristics of actual brain pathology.

[0065] During the training process, the deep clustering model adopts an unsupervised learning strategy, i.e., without relying on artificial label annotation, but guiding the adaptive optimization of network parameters by minimizing a clustering objective function (such as KL divergence, entropy loss or center reconstruction error), so that the samples gradually move towards the most matching cluster center in the clustering space, and finally form a stable and distinguishable clustering structure.

[0066] Through the learning and clustering allocation of the deep clustering model, the system can divide the white matter lesions into multiple type clusters with significant feature differences, such as clear boundary-small volume type, signal enhancement-irregular diffusion type, etc. Each clustering cluster represents a type of lesion performance with similar morphology and potential risk level, providing a structured input basis for the subsequent Alzheimer's disease conversion risk assessment model.

[0067] In summary, the embodiment realizes the automatic clustering of high-dimensional white matter lesion features by constructing a deep clustering model composed of MLP and Softmax module, and adopting an unsupervised training method, which not only avoids the dependence on artificial labels, but also reveals the internal structural rules of lesions in the morphology and evolution level, providing a basis for disease typing, process prediction and individualized intervention.

[0068] In a feasible implementation, the lesion features of the white matter include lesion volume, lesion gray signal intensity, lesion distribution pattern and regional relationship, and lesion boundary morphology features, and the lesion boundary morphology features include boundary sharpness, irregularity, gray scale contrast and fractal dimension.

[0069] For example, the lesion features of the white matter include lesion volume, lesion gray signal intensity, lesion distribution pattern and regional relationship, and lesion boundary morphology features. Through systematic extraction and quantification of these features, the multi-dimensional performance of white matter lesions in the structure, signal and morphology level can be comprehensively described, providing rich input data for subsequent lesion clustering analysis and Alzheimer's disease risk assessment.

[0070] Specifically, the lesion volume is an important indicator for measuring the lesion range. By calculating the number of voxels contained in the lesion area in the MRI image, and combining the voxel size information of the image, the three-dimensional volume of the lesion area is obtained. This indicator can reflect the extension degree of the lesion in the brain tissue, and is closely related to the disease development stage.

[0071] The lesion gray signal intensity is used to evaluate the imaging brightness characteristics of the lesion area, which is usually represented as the average gray value or signal intensity value of all voxels in the region. To exclude the influence of individual differences in scanning parameters, the signal intensity is generally normalized based on the normal white matter region, so as to obtain the normalized lesion signal index. This feature helps to identify lesions of different properties (such as regional white matter lesion volume and corresponding gray matter cortex thickness), and has sensitivity in the early pathological evolution of Alzheimer's disease.

[0072] The lesion distribution pattern and regional relationship mainly focuses on the distribution rule of the lesion in the brain anatomical structure, including whether the lesion is symmetrical, whether it is concentrated in certain brain areas (such as frontal lobe, basal ganglia, precuneus, entorhinal cortex-hippocampus-posterior cingulate, etc.), and whether it is adjacent to certain anatomical structures (such as corpus callosum, ventricular system). This feature reflects the spatial heterogeneity of the lesion, which helps to distinguish the white matter changes caused by different pathological processes.

[0073] In terms of lesion boundary morphology characteristics, the embodiment quantitatively extracts the following dimensions: Boundary definition: By calculating the transition gradient of the gray value of the lesion boundary and the surrounding tissue, it is evaluated whether the lesion edge is clear and explicit. Clear boundary usually indicates that the lesion process is stable or isolated, while blurred boundary may indicate that the lesion is expanding or there is inflammatory exudation.

[0074] Boundary irregularity: The complexity of the lesion contour is measured by using shape indicators such as perimeter-area ratio and boundary curvature change rate. Irregular boundary is usually related to uneven growth, infiltrative expansion or multi-focal fusion of the lesion.

[0075] Gray contrast: It is calculated by the average gray difference between the lesion edge region and the background white matter, which is used to reflect the distinguishability and visual contrast of the lesion, and is an important influencing factor of the performance of segmentation and recognition model.

[0076] Fractal dimension: The structural complexity of the lesion boundary is described using fractal geometry method. The higher the fractal dimension, the rougher the boundary or the stronger the self-similarity. This indicator has unique advantages in capturing the evolution morphology of complex lesions.

[0077] In summary, the embodiment realizes quantitative, fine and structured description of the lesion by extracting the above-mentioned multiple dimensions of white matter lesion characteristics, which can provide stable and reliable high-value feature input for subsequent deep clustering analysis, type division and risk modeling, thereby significantly improving the recognition ability and explanation ability of the model for early warning of Alzheimer's disease.

[0078] In a feasible embodiment, the above-mentioned risk assessment model uses LSTM network to encode the sequenced omics features, and is used to output the risk grade of Alzheimer's disease conversion.

[0079] For example, the above risk assessment model uses an LSTM network to encode the sequenced white matter lesion omics features, and outputs the subject's Alzheimer's disease conversion risk level based on the extracted time series deep features. The core of this embodiment is to capture the evolution trend and potential risk law of the lesion features in the time dimension through time series modeling, thereby improving the sensitivity and prediction accuracy of risk assessment.

[0080] Specifically, the lesion omics features are extracted from the white matter lesion regions segmented from MRI images, covering volume changes, gray signal intensity changes, distribution pattern evolution, boundary shape variation, and other feature dimensions. To adapt these features to the input structure of the LSTM model, the system first performs time series processing on various features, i.e., arranging and constructing multi-dimensional sequence samples according to the image data of different time nodes of the subject, so that the model can perceive the longitudinal evolution information of the lesion features.

[0081] Subsequently, these sequenced features are input into the LSTM network. LSTM, as a recurrent neural network structure specifically designed for processing time series data, can remember the hidden state of the previous time step during encoding and update the information with the input features of the current time step, thereby effectively modeling the time series dependency of the features. In the context of white matter lesions, LSTM can not only capture the trend of lesion volume or signal intensity changes, but also identify the evolution law of boundary complexity, heterogeneity, and other morphological indicators at different stages.

[0082] After the LSTM network encodes the input omics sequence, the final output hidden state vector will serve as a deep representation of the subject's overall lesion process. The system inputs this representation vector into the subsequent discriminant layer (such as a fully connected layer + Softmax activation function), thereby outputting the corresponding Alzheimer's disease conversion risk level. This risk level can be divided into low, medium, and high risk levels according to model settings, or presented as a probability value representing the individual's risk of converting to Alzheimer's disease in the future.

[0083] Compared to traditional static feature assessment methods, this embodiment significantly improves the ability to characterize the progression pattern of white matter lesions, especially in identifying non-linear trends, sudden changes, or hidden risk accumulation. By introducing an LSTM network to dynamically model the lesion sequence, individualized prediction of Alzheimer's disease conversion risk can be achieved, providing strong technical support for early intervention and precise prevention.

[0084] To sum up, the embodiment realizes the time sequence modeling and risk prediction of the white matter lesion omics characteristics by constructing the LSTM-based deep risk assessment model, has high medical applicability and engineering popularization value, and is especially suitable for clinical intelligent auxiliary scenes such as early screening, follow-up evaluation and disease conversion early warning of Alzheimer's disease.

[0085] The above examples are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An early warning method for Alzheimer's disease based on the morphological characteristics of brain white matter lesions, characterized in that, include: We acquired magnetic resonance imaging data of historical subjects from the period of mild cognitive impairment to before the diagnosis of Alzheimer's disease, and manually labeled the white matter and white matter lesion areas to obtain a manually labeled dataset. A deep learning model for identifying white matter lesions is trained based on the artificially labeled dataset. The deep learning model uses a convolutional module to extract features from T1w and FLAIR images in the shallow layer and integrates multimodal information through a fusion module. In the deep layer, the deep learning model uses a lightweight parallel vision Mamba module to perform deep feature modeling and global perception on the FLAIR images. The deep learning model is designed with a dual-task training structure of region segmentation and boundary detection to improve segmentation accuracy. The magnetic resonance imaging data to be identified is input into the deep learning model to extract lesion features of brain white matter; The extracted lesion features are standardized and aligned before being input into a deep clustering model to generate clustering results for different types of white matter lesion features. An early risk assessment model is constructed based on the clustering results and their corresponding Alzheimer's disease conversion risk labels, and the Alzheimer's disease conversion risk level of the subjects is output.

2. The method according to claim 1, characterized in that, The magnetic resonance imaging data images were obtained using a 3.0 Tesla scanning system and consisted of 28-32 layer sequences with a slice thickness of 5 mm.

3. The method according to claim 1, characterized in that, The white matter and lesion annotations were independently annotated by at least two physicians with neuroimaging experience, and the segmentation results passed the statistical consistency test; if there were discrepancies, they were reviewed by a third physician.

4. The method according to claim 1, characterized in that, The deep learning model employs a fusion module to fuse the multimodal feature information of the T1w image and the FLAIR image in the channel and spatial dimensions at a shallow layer of the encoder.

5. The method according to claim 1, characterized in that, The lightweight parallel vision Mamba module is used to replace the convolutional neural networks in the deep structure of the encoder and decoder to achieve more efficient deep feature extraction. The visual Mamba module includes a VSS structure, which is used to model complex spatial dependencies and global context information in images. The VSS structure includes a first branch for basic feature modeling and a second branch for complex feature modeling. The first branch includes a linear layer and an activation function, and the second branch includes a linear layer, a deep convolutional layer, an activation function, a two-dimensional selective scanning module, and a normalization layer.

6. The method according to claim 1, characterized in that, The deep learning model adopts a dual-task training structure, and the loss function of the deep learning model during the training process includes a weighted combination of the losses of the segmentation task and the boundary detection task.

7. The method according to claim 1, characterized in that, The deep clustering model includes a multilayer perceptron and a Softmax probability mapping module. The deep clustering model uses an unsupervised training method to map features to the clustering space.

8. The method according to claim 1, characterized in that, The lesion characteristics of the brain white matter include lesion volume, lesion gray signal intensity, lesion distribution pattern and regional relationship, and lesion boundary morphological characteristics, including boundary clarity, irregularity, gray contrast and fractal dimension.

9. The method according to claim 1, characterized in that, The risk assessment model uses an LSTM network to encode sequential omics features and outputs the risk level of Alzheimer's disease conversion.

Images