Alzheimer's disease low-resolution MRI diagnosis method based on multi-source domain adaptation

Abstract

This invention belongs to the field of medical imaging diagnostic technology, specifically relating to a low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation. The method includes constructing and training an unsupervised multi-source domain adaptive model for Alzheimer's disease, and using the trained model to achieve low-resolution MRI diagnosis. The unsupervised multi-source domain adaptive model includes a cross-resolution dual-path general feature encoder, a domain-specific attention feature extractor, and a shared diagnostic classifier. The proposed method only requires input of a standard clinical thick-slice low-resolution MRI during the inference phase to achieve diagnostic performance similar to multimodal methods, significantly reducing examination costs and operational barriers, and providing a feasible path for the large-scale popularization of early AD screening.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging diagnostic technology, specifically relating to a low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation. Background Technology

[0002] Alzheimer's disease (AD) is often referred to as a "blind spot of modern medicine." As one of the world's leading neurodegenerative diseases, it not only seriously threatens the health of the elderly but is also the leading cause of dementia. Its typical pathological features include β-amyloid (Aβ) plaque deposition and neurofibrillary tangles. These abnormal accumulations gradually destroy synaptic function, leading to neuronal loss and ultimately widespread cerebral cortical atrophy. To date, the pathogenesis of AD remains incompletely understood. Drugs developed based on mainstream pathological hypotheses typically only alleviate symptoms and cannot effectively reverse or halt disease progression. Therefore, achieving early and accurate diagnosis has become a key breakthrough in slowing disease progression and reducing the social burden.

[0003] Neuroimaging techniques provide an irreplaceable means for the in vivo characterization of Alzheimer's disease (AD), enabling highly sensitive detection of key pathological changes such as amyloid deposition, tau protein aggregation, neurodegenerative changes, and brain atrophy through multimodal biomarkers. Among these, amyloid positron emission tomography (Aβ-PET) is considered the molecular gold standard for in vivo visualization and quantification of fibrillary amyloid load, providing direct evidence of AD-related pathological processes. However, its widespread clinical application is limited by high costs, radiation exposure, and dependence on radiopharmaceutical infrastructure, restricting its adoption in large-scale screening and typically limiting its use to highly screened patient populations. In contrast, magnetic resonance imaging (MRI) offers advantages such as being non-invasive, radiation-free, and widely available. High-resolution (HR) thin-slice MRI can clearly capture the fine anatomical details of brain structures and is highly sensitive to subtle atrophy caused by early neurodegenerative diseases. However, its long acquisition time makes it unsuitable for high-throughput clinical scenarios. Therefore, rapid scanning with low-resolution (LR) thick-slice MRI is primarily used in routine clinical practice. Despite its high efficiency, LR MRI suffers from significant partial volume effects and reduced signal-to-noise ratio. These shortcomings mask early changes in brain structure volume and morphology, resulting in limited sensitivity in the early detection of Alzheimer's disease (AD).

[0004] This creates a diagnostic dilemma: on the one hand, the most precise biomarkers are difficult to obtain routinely; on the other hand, clinically available imaging tools lack sufficient sensitivity for early identification. Therefore, the current major challenge lies in developing novel analytical methods that can integrate the pathological information from PET with the detailed anatomical knowledge from HR MRI, and rely solely on clinically available LR MRI data.

[0005] In recent years, deep learning has further propelled the development of neuroimaging analysis, enabling the extraction of subtle, multi-scale image features beyond the capabilities of the human eye. While multimodal fusion models based on MRI and PET demonstrate excellent diagnostic performance, their reliance on simultaneous multimodal data acquisition severely limits their feasibility for clinical application. What we need is not multimodal fusion during the inference phase, but rather the effective transfer of multimodal knowledge during model training, allowing models based solely on conventional LR MRI to absorb molecular pathological information from PET and anatomical details from HR MRI.

[0006] However, achieving such cross-modal and cross-resolution transfer faces significant challenges: PET, HR MRI, and LR MRI originate from different data distributions and differ in imaging modality, spatial resolution, contrast, acquisition protocols, and the dimensions of biological information they carry. Overcoming these differences requires constructing a framework capable of coordinating feature representations across different domains without relying on a large number of paired or labeled samples. While domain-adaptive methods offer a potential path, traditional approaches often focus on single-dimensional distribution alignment (e.g., addressing only modal or resolution differences), making it difficult to address the multidimensional heterogeneity issues present in AD images: bridging the biological semantic gap between PET molecular information and MRI anatomical information, eliminating structural distortion caused by resolution differences between HR and LR MRI, and overcoming image contrast shifts caused by inconsistencies in equipment and protocols across different medical institutions. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation. An unsupervised multi-source domain adaptive model for Alzheimer's disease is constructed and trained, and the trained model is used to achieve low-resolution MRI diagnosis. The unsupervised multi-source domain adaptive model for Alzheimer's disease includes a cross-resolution dual-path general feature encoder, a domain-specific attention feature extractor, and a shared diagnostic classifier.

[0008] The training process of an unsupervised multi-source domain adaptive model for Alzheimer's disease includes:

[0009] S1. Construct a standard dataset for knowledge transfer, comprising: a first source domain consisting of Aβ-PET images; a second source domain consisting of high-resolution MRI images; and a target domain consisting of low-resolution MRI images.

[0010] S2. Construct dual-resolution input pairs:

[0011] For the Aβ-PET image in the first source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair in the first source domain.

[0012] For the high-resolution MRI of the second source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair of the second source domain;

[0013] For low-resolution MRI of the target domain, interpolation is performed to generate corresponding high-resolution images, and the two are paired to form a dual-resolution input pair for the target domain;

[0014] S3. For each domain's dual-resolution input pair, pass it through a cross-resolution dual-path general feature encoder to obtain perceptual features; the cross-resolution dual-path general feature encoder includes a high-resolution feature coding branch and a low-resolution feature coding branch, the two feature coding branches are in parallel structure, their network structures are the same but their weight parameters are independent of each other; wherein, the high-resolution image is processed by the high-resolution feature coding branch to generate a high-resolution feature pyramid; the low-resolution image is processed by the low-resolution feature coding branch to generate a low-resolution feature pyramid; the perceptual features are composed of the above high-resolution and low-resolution feature pyramids.

[0015] S4. Input the perceptual features of the two source domains into the domain-specific attention feature extractor of the corresponding domain to obtain the refined features; the refined features include high-resolution refined features and low-resolution refined features;

[0016] S5. Input the perceptual features of the target domain into the domain-specific attention feature extractors of the two source domains respectively: obtain the structural features through the processing of the domain-specific attention feature extractor corresponding to the second source domain; obtain the pseudo-molecule features through the processing of the domain-specific attention feature extractor corresponding to the first source domain.

[0017] S6. After flattening the structural features, pseudo-molecular features, and refined features from the two source domains, they are input into the shared diagnostic classifier, which outputs four prediction results: anatomical high-resolution prediction, anatomical low-resolution prediction, high-resolution molecular prediction, and low-resolution molecular prediction.

[0018] S7. Based on the four prediction results, calculate the multi-objective joint loss, optimize the model parameters through the backpropagation algorithm, and iterate the training until the model converges; the multi-objective joint loss includes supervised classification loss, multi-source cross-domain distribution alignment loss, and multi-branch prediction consistency constraint loss.

[0019] The beneficial effects of this invention are:

[0020] First, the proposed method only requires input of routine clinical thick-slice low-resolution MRI during the inference stage to achieve diagnostic performance similar to multimodal methods, significantly reducing examination costs and operational barriers, and providing a feasible path for the large-scale popularization of early AD screening.

[0021] Secondly, this method integrates multi-source heterogeneous information during training, while relying solely on a single conventional low-resolution MRI in practical applications. This achieves high-precision screening that is radiation-free, low-cost, and easily accessible, effectively overcoming the bottleneck that makes it difficult to widely implement Aβ-PET and high-resolution MRI in clinical practice.

[0022] Furthermore, by introducing an unsupervised domain adaptation mechanism, the model can significantly alleviate the differences in data distribution among different medical institutions and different scanning devices, and demonstrates stable diagnostic performance and good generalization ability in cross-center validation.

[0023] From an interpretability perspective, gradient-weighted class activation mapping shows that the model continuously focuses on brain regions typically associated with AD pathology, such as the hippocampus and parahippocampal gyrus, during the decision-making process. The visualization results are highly consistent with clinical pathological cognition, providing a reliable spatial basis for diagnosis.

[0024] Finally, the model employs end-to-end full three-dimensional volumetric image analysis, avoiding the spatial context fragmentation caused by two-dimensional slicing, enabling more complete modeling of structural and functional changes across the entire brain and improving the sensitivity to the identification of diffuse and subtle pathological patterns. Attached Figure Description

[0025] Figure 1 This is a flowchart of a low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation, according to the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention provides a low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation. By inputting a single routinely acquired low-resolution magnetic resonance imaging (MRI) image from a clinical case, the corresponding AD diagnosis result can be directly output. The core of this method lies in the unsupervised multi-source domain adaptation technique used during model training. This technique effectively transfers detailed anatomical information from high-resolution MRI and pathological features from amyloid positron emission tomography (Aβ-PET) scans to the analysis model for low-resolution MRI, thus requiring only a single low-resolution MRI modality as input during the inference phase. This design significantly reduces dependence on expensive imaging equipment, radioactive tracers, and high-resolution data, while enhancing the model's adaptability and robustness under different acquisition conditions. It provides a feasible path to achieve low-cost, highly accessible, and highly reliable early screening for AD.

[0028] Please see Figure 1 A low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation is proposed. The method involves constructing and training an unsupervised multi-source domain adaptive model for Alzheimer's disease, and then using the trained model to achieve low-resolution MRI diagnosis. The unsupervised multi-source domain adaptive model includes a cross-resolution dual-path general feature encoder, a domain-specific attention feature extractor, and a shared diagnostic classifier.

[0029] The training process of an unsupervised multi-source domain adaptive model for Alzheimer's disease includes:

[0030] S1. Construct three types of standard datasets for knowledge transfer, including: a first source domain standard dataset consisting of Aβ-PET images; a second source domain standard dataset consisting of high-resolution MRI; and a target domain standard dataset consisting of low-resolution MRI.

[0031] Preferably, structural magnetic resonance imaging data are acquired using a 3.0T MRI scanner; molecular imaging data are obtained through amyloid positron emission tomography (Aβ-PET); ultimately forming three domain datasets. Each domain dataset contains Alzheimer's disease (AD) patient samples and healthy control (HC) samples from different subjects, and it is not required that the subjects in each domain dataset correspond one-to-one.

[0032] During the data acquisition and preprocessing phase, standardized processes are employed to ensure data consistency and compatibility with deep learning models. Preprocessing includes the following steps:

[0033] First, all medical images were converted from the Digital Imaging and Communications in Medicine (DICOM) format to the National Informatics for Neuroimaging (NIFTI) format, and all personally identifiable information was removed during the conversion process to achieve data anonymization.

[0034] Secondly, spatial resampling is performed on each modality image according to the model input requirements. The high-resolution MRI and Aβ-PET images in the source domain are uniformly adjusted to a voxel size of 128×128×96, and the low-resolution MRI in the target domain is resampled to a voxel size of 128×128×24.

[0035] Finally, the resampled images are intensity normalized using Z-score normalization to convert the intensity values ​​of each modality into a distribution with a mean of 0 and a standard deviation of 1. This eliminates differences caused by different scanning protocols and enhances the stability and generalization performance of the model training. Subsequently, the processed image array is converted into a tensor format adapted for PyTorch framework training to provide a unified and optimized data input.

[0036] S2. Construct dual-resolution input pairs.

[0037] Preferably, the core function of the cross-resolution dual-path universal feature encoder lies in resolving feature misalignment and direct transfer obstacles caused by resolution mismatch by processing the heterogeneous spatial inputs in the source and target domains due to resolution differences. Therefore, dual-resolution input pairs need to be constructed for samples in both the source and target domains. Specifically:

[0038] For the Aβ-PET image in the first source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair in the first source domain.

[0039] For the high-resolution MRI of the second source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair of the second source domain;

[0040] For low-resolution MRI of the target domain, an interpolation method is used to generate corresponding high-resolution images, and the two are paired to form a dual-resolution input pair for the target domain.

[0041] This strategy achieves dual-stream feature alignment by constructing high- and low-resolution input pairs for different domains. In the high-resolution (HR) stream, the model aligns the true anatomical details of the source domain with the upsampled and enhanced structural information of the target domain; in the low-resolution (LR) stream, it aligns the robust features simulated by downsampling with the low-resolution data actually acquired in clinical settings. This design effectively mitigates the domain distribution shift caused by differences in slice thickness and other factors, improving the model's adaptability across images of different resolutions.

[0042] S3. For each domain's dual-resolution input pair, pass it through a cross-resolution dual-path general feature encoder to obtain perceptual features; the perceptual features include a high-resolution feature pyramid and a low-resolution feature pyramid.

[0043] Preferably, to maximize the preservation of inherent information from images at different resolutions, the cross-resolution dual-path universal feature encoder includes a high-resolution feature coding branch and a low-resolution feature coding branch. The two feature coding branches are connected in parallel, with identical network structures but independent weight parameters. The two feature coding branches process high-resolution and low-resolution inputs respectively, avoiding compromises between detail extraction and coarse pattern recognition by a single encoder.

[0044] To enhance feature representation capabilities, the feature encoding branch uses the ResNet3D network as its backbone and combines it with a three-dimensional feature pyramid to achieve multi-scale feature fusion. This includes: extracting features L2, L3, and L4 from the C2, C3, and C4 stages of the ResNet3D network, aligning the channel dimensions through 1×1×1 convolutions to obtain features L'2, L'3, and L'4; upsampling feature L'4 through trilinear interpolation and adding it element-wise with feature L'3 to obtain feature P3; upsampling feature P3 again and adding it element-wise with feature L'2 to obtain feature P2; finally, the feature pyramid F composed of features P2, P3, and L'4 is output, fusing high-resolution details and deep semantic information.

[0045] S4. Input the perceptual features of the two source domains into the domain-specific attention feature extractor of the corresponding domain to obtain the refined features; the refined features include high-resolution refined features and low-resolution refined features.

[0046] S5. Input the perceptual features of the target domain into the domain-specific attention feature extractors of the two source domains respectively: obtain the structural features through the processing of the domain-specific attention feature extractor corresponding to the second source domain; obtain the pseudo-molecule features through the processing of the domain-specific attention feature extractor corresponding to the first source domain.

[0047] Preferably, the domain-specific attention feature extractor is mainly responsible for refining the general multi-scale features output by the cross-resolution dual-path general feature encoder, capturing the unique biological features of AD pathology based on different modal imaging, and avoiding the general processor from confusing or suppressing modality-specific features.

[0048] This invention sets up independent domain-specific processing paths for feature flows of different modalities and resolutions. Specifically, the domain-specific attention feature extractors for both source domains include both high-resolution and low-resolution feature extractors, and all four feature extractors have identical structures. Each feature extractor consists of a cascaded array of three convolutional layers and one modality-aware collaborative attention (MS-CSAM3D) module, with residual connections between the input of the feature extractor and the output of the MS-CSAM3D module.

[0049] In any feature extractor, the input features are first processed through three convolutional layers to extract the primary feature F1. Subsequently, the MS-CSAM3D module is introduced to perform modality-adaptive enhancement on the primary feature F1. This module designs a dual-channel collaborative calibration mechanism specifically for the imaging characteristics of both Aβ-PET and high-resolution MRI source domains in this invention.

[0050] Modality-aware channel attention submodule: This submodule introduces modality prior embedding vectors. For the feature extractor corresponding to the first source domain (Aβ-PET), the embedding vector focuses on activating feature channels strongly correlated with Aβ deposition load; for the feature extractor corresponding to the second source domain (high-resolution MRI), it focuses on activating feature channels associated with brain structural atrophy. Channel attention weights. The calculation method is as follows:

[0051]

[0052] GAP stands for Global Average Pooling. For element-wise multiplication, For the Sigmoid function, For ReLU, , These are the weights for the fully connected layer. The learnable parameter vectors are defined for each of the two source domains and jointly optimized with the network, requiring no manual design or pre-computation. This mechanism enables feature extractors of different modalities to adaptively enhance channels relevant to their own modality diagnosis while suppressing responses of irrelevant channels, thereby avoiding interference between molecular and structural features.

[0053] Spatial Attention Submodule: To accurately locate irregular lesion morphologies in Alzheimer's disease-related brain regions (such as irregular boundaries of hippocampal atrophy and scattered amyloid plaques), this submodule employs a 3D deformable roll. Spatial Attention Weights The calculation is as follows:

[0054]

[0055] in, For the feature map after channel attention calibration, DeformConv3 represents a 3D deformable convolution. Deformable convolution adjusts the spatial position of each sampling point through a learnable offset field, enabling the receptive field to fit the actual geometry of the lesion area, thereby enhancing the feature response to irregular pathological patterns.

[0056] The final output of the enhanced features is as follows:

[0057]

[0058] in, For element-wise multiplication, residual connection This approach mitigates gradient vanishing while preserving underlying details. By fusing element-wise channel weights and spatial weights, it achieves dual focus on modally important channels and irregular regions of disease.

[0059] In this embodiment of the invention, within the labeled source domain, the two feature extractors in the domain-specific attention feature extractor corresponding to the first source domain are optimized for the high-resolution feature pyramid and the low-resolution feature pyramid obtained from the Aβ-PET image, respectively, to generate high-resolution refined features. and low-resolution refinement features ;in, Corresponding to Aβ load characteristics, Corresponding molecular features. The two feature extractors in the domain-specific attention feature extractor corresponding to the second source domain are optimized for the high-resolution feature pyramid and the low-resolution feature pyramid obtained from high-resolution MRI, respectively, to generate high-resolution refined features. and low-resolution refinement features ;in, Corresponding structural shrinkage characteristics Corresponding organizational characteristics.

[0060] Specifically, instead of setting up a dedicated domain-specific attention feature extractor for the target domain, a dual-source domain knowledge transfer mechanism is used to collaboratively generate structural and molecular dual-modal features. Specifically, the perceptual features of the target domain are input into the domain-specific attention feature extractors corresponding to the two source domains: the perceptual features of the target domain are processed by the domain-specific attention feature extractor of the second source domain to obtain structural features consistent with the high-resolution MRI distribution of the second source domain. Simultaneously, by using a domain-specific attention feature extractor of the first source domain, the perceptual features of the target domain are processed, and the corresponding pseudo-molecular features are inferred solely from structural MRI. This design effectively bridges the semantic gap between the multimodal source domain and the single-modal target domain.

[0061] It is important to emphasize that although the low-resolution MRI of the target domain and the Aβ-PET image of the first source domain differ fundamentally in their imaging principles, the training mechanism subsequently employed in this invention ensures the effectiveness of the generated "pseudo-molecular features." Specifically:

[0062] Distribution alignment constraint: Through multi-source cross-domain distribution alignment loss, the model forces the distribution of pseudo-molecular features generated in the target domain in the latent space to be consistent with the true Aβ-PET molecular features in the first source domain. This constraint helps guide the model to effectively learn the mapping rules of PET molecular features from MRI structural information.

[0063] Pre-training and Co-training: First, the domain-specific attention feature extractor of the first source domain was pre-trained on labeled Aβ-PET data, learning and mastering the encoding patterns of relevant molecular features. During joint training, although the parameters of the domain-specific attention feature extractor of the first source domain were updated slowly, they provided stable latent space "anchors" for the target domain features. Simultaneously, by applying a consistency loss and combining it with backpropagation of the diagnostic classification gradient of the target domain, the "pseudo-molecular features" generated by the model were forced to contribute to the final AD / NC classification, thus ensuring that the "pseudo-molecular features" possess genuine diagnostic discriminative power. Therefore, the obtained "pseudo-molecular features" are effective approximations based on structure-pathology association knowledge transfer, rather than meaningless noise.

[0064] S6. After flattening the structural features, pseudomolecule features, and refined features from the two source domains, they are input together into a shared diagnostic classifier. The shared diagnostic classifier consists of four independent binary fully connected layers and outputs four prediction results; the four prediction results include anatomical high-resolution prediction. Anatomical low-resolution prediction High-resolution molecular prediction and low-resolution molecular prediction .

[0065] S7. Based on the four prediction results, calculate the multi-objective joint loss, optimize the model parameters through the backpropagation algorithm, and iterate the training until the model converges.

[0066] Preferably, the present invention designs a multi-objective joint optimization objective through three core modules: feature semantic preservation and discriminative learning objective, multi-source cross-domain distribution alignment objective, and multi-branch prediction consistency constraint objective.

[0067] First, to address the goals of feature semantic preservation and discriminative learning, this invention designs a supervised classification loss. The core function of this loss term is to ensure high diagnostic classification accuracy on labeled source domain data, providing a reliable foundation for cross-domain knowledge transfer. Its purpose is to guide the source domain feature extractor to learn core semantic features strongly correlated with Alzheimer's disease diagnosis, while effectively suppressing interference from redundant noise features. The constraints of this loss term cover all parameters of the domain-specific attention feature extractors and the shared diagnostic classifier from both source domains. Functionally, it employs multi-class cross-entropy loss, specifically expressed as:

[0068]

[0069] in, This represents the total number of labeled samples in the source domain. This represents the total number of categories in the diagnostic classification task. Indicates the first One-hot encoding of the true labels of the source domain samples Indicates a shared diagnostic classifier for the first... The source domain sample at the th th Predicted probabilities on a class.

[0070] Secondly, for the goal of adaptive cross-domain distribution alignment in multiple source domains, a multi-source cross-domain distribution alignment loss is designed. This loss term is the core of unsupervised knowledge transfer. Its role is to simultaneously align the feature distributions of the two source domains and the target domain in the latent feature space, thereby mitigating the triple domain shift caused by modal heterogeneity, resolution differences, and multi-center acquisition protocols. This constraint applies to the domain-specific attention feature extractor parameters of the first and second source domains and the front-end feature encoding module parameters of the target domain. The functional form adopts the multi-kernel maximum mean difference (MK-MMD) loss adapted for multi-source domain feature alignment. The specific expression consists of two parts: the alignment loss between the first source domain and the target domain, and the alignment loss between the second source domain and the target domain, as shown below:

[0071]

[0072] This represents the distribution alignment loss between the first source domain and the target domain. Its function is to constrain the distribution of pseudo-molecular features generated by the target domain to be consistent with the distribution of molecular features in the first source domain, thereby realizing cross-modal transfer of molecular pathology knowledge from the first source domain to the target domain. The expression is as follows:

[0073]

[0074] In the formula, Indicates the number of samples in the first source domain. This represents the high-resolution refined feature of the i-th sample in the first source domain. This indicates the number of unlabeled samples in the target domain. This represents the low-resolution refined feature of the i-th sample in the first source domain. This represents the high-resolution pseudo-molecule feature of the j-th sample in the target domain. This represents the low-resolution pseudo-molecule feature of the j-th sample in the target domain.

[0075] This represents the distribution alignment loss between the second source domain and the target domain. Its function is to ensure that the structural features generated in the target domain are consistent with the distribution of structural features in the second source domain, thereby enabling the transfer of structurally shrunken knowledge across resolution and center-of-view scenarios. The expression is:

[0076]

[0077] In the formula, Indicates the number of samples in the second source domain. Represents the regenerated nucleus Hilbert space In the feature mapping, a multi-kernel mapping method combining Gaussian kernel and multinomial kernel is used to improve the model's ability to align features with different distributions. This represents the high-resolution refined feature of the i-th sample in the second source domain. This represents the high-resolution structural features of the j-th sample in the target domain. This represents the low-resolution refined feature of the i-th sample in the second source domain. This represents the low-resolution structural features of the j-th sample in the target domain.

[0078] Subsequently, a multi-branch prediction consistency constraint loss is designed for the multi-branch prediction consistency constraint objective. The core function of this loss term is to minimize the difference in prediction results between branches at different resolutions, thereby enhancing the model's robustness in resolution degradation scenarios and further promoting the transfer of discriminative knowledge from the source domain to the target domain. Its constraints cover the bi-branch parameters corresponding to the domain-specific attention feature extractors of the first and second source domains, as well as the parameters of the shared diagnostic classifier. The specific expression is as follows:

[0079]

[0080] in It is an L1 norm. This represents the high-resolution anatomical prediction result corresponding to the j-th sample in the target domain. This represents the low-resolution anatomical prediction result corresponding to the j-th sample in the target domain. This represents the high-resolution molecular prediction result corresponding to the j-th sample in the target domain. This represents the low-resolution molecular prediction result corresponding to the j-th sample in the target domain.

[0081] This invention employs a dynamic weight allocation mechanism to balance the contributions of each loss component, achieving multi-objective collaborative optimization and resolving the priority conflict between source domain discriminative feature learning and domain adaptation in the early stages of training.

[0082] Preferably, the total loss is calculated as follows:

[0083]

[0084] in, This is a dynamic weighting factor, whose value is updated dynamically with each training round. The specific update formula is as follows:

[0085]

[0086] in, This represents the upper limit of the weighting factor (preferably ranging from 0.5 to 2.0, which can be adjusted based on the distribution of the dataset). For the current training round, This refers to the total number of training rounds.

[0087] Unsupervised multi-source domain adaptive training strategy is the core of knowledge transfer. Its goal is to effectively transfer diagnostic knowledge from labeled multi-source domains (such as high-resolution MRI and Aβ-PET) to unlabeled low-resolution MRI target domains, thereby significantly mitigating domain distribution shifts caused by resolution differences, modal differences and inconsistent scanning protocols.

[0088] Preferably, to clarify the interpretability of the model's decisions and its potential neurobiological basis, this invention employs gradient-weighted class activation mapping (Grad-CAM) to visually interpret the model. This method does not require modification of the trained model structure and can directly reveal the key spatial features that the network focuses on during the decision-making process.

[0089] For a given 3D MRI input, let the activation value of the k-th feature map in the last convolutional layer at spatial location (x, y, z) be... For target category c (AD or HC), its corresponding Grad-CAM heatmap is generated using the following weighted combination method:

[0090]

[0091]

[0092] in, Z represents the model's prediction score for category c, and Z represents the spatial volume of the feature map. represents the weight coefficient of feature map k with respect to category c; the ReLU function is used to filter out negative contributions, retaining only regions that have a positive effect on category determination; It is a three-dimensional heat map, which can visualize the spatial distribution of brain regions of interest during model diagnosis, and provide support for the interpretation of model diagnostic logic.

[0093] To verify the synergistic effect of each technical module, an ablation experiment was designed and compared with a baseline method in this embodiment of the invention.

[0094] The ablation experiment setup is as follows:

[0095] 1. HR-only: Only the high-resolution MRI source domain and related branches are retained, while the Aβ-PET source domain and its branches are removed;

[0096] 2. w / o L1: Retain the dual-source domain of high-resolution MRI and Aβ-PET, but remove the multi-branch prediction consistency constraint loss;

[0097] 3. The complete model proposed in this invention (ADMAN): simultaneously uses dual-source domains (high-resolution MRI + Aβ-PET) and applies multi-branch prediction consistency constraint loss.

[0098] Table 1. Results of the ablation pre-experiment

[0099]

[0100] Experimental results show that, from "HR only" to "w / o L1", introducing only the Aβ-PET source domain (without adding multi-branch prediction consistency constraint loss) brings significant performance improvements: accuracy increases by 5.17 percentage points (from 80.98% to 86.15%), and F1-score increases by 4.98 percentage points (from 81.52% to 86.50%). This demonstrates that the "pseudo-molecular features" provided by the Aβ-PET source domain contain pathological information that high-resolution MRI cannot obtain, and the dual-source domain design achieves functional complementarity rather than simple data superposition.

[0101] Further observation of the changes from "w / o L1" to the complete model "ADMAN" reveals that after adding the multi-branch prediction consistency constraint loss, the accuracy improved by 2.64 percentage points (86.15% → 88.79%), and the sensitivity improved by 2.90 percentage points (87.80% → 90.70%). This indicates that the loss effectively promotes the fusion of dual-source domain knowledge by constraining the prediction consistency of high- and low-resolution branches, enabling "pseudo-molecular features" and "structural features" to mutually verify each other at the decision-making level, resulting in a significant synergistic enhancement effect.

[0102] The complete model (ADMAN) proposed in this invention was compared with several classic unsupervised domain adaptation (UDA) methods based solely on high-resolution MRI source domains, including DANN, DSAN, and DeepCoral. Preliminary experimental results are shown in the table below.

[0103] Table 2

[0104]

[0105] Experimental results show that the traditional single-source domain unsupervised domain adaptation (UDA) method can only handle distribution differences (such as alignment at different resolutions) within a single modality and cannot utilize the molecular pathological information contained in Aβ-PET. Therefore, its highest accuracy is only 68.93%. In contrast, the accuracy of the complete model ADMAN proposed in this invention reaches 88.79%, which is nearly 20 percentage points higher.

[0106] This significant improvement stems from a crucial methodological breakthrough in this invention: traditional UDA is typically limited to alignment within the same modality, while this method innovatively introduces the Aβ-ET source domain and combines it with multi-branch prediction consistency constraint loss, achieving cross-modal knowledge transfer from molecular images to structural images for the first time. This allows low-resolution MRI to acquire both detailed anatomical structural information and crucial molecular pathological features during the inference stage, thereby significantly enhancing diagnostic performance. Furthermore, to verify the effectiveness of the MS-CSAM3D module proposed in this invention, comparative experiments were conducted within a unified multi-source domain adaptive framework. In the experiments, a baseline model (labeled "no attention") was set up, whose domain-specific feature extractor contained only three convolutional layers and residual connections, without introducing any attention mechanism; the model of this invention embedded the MS-CSAM3D module at the same structural location. Both sets of experiments used dual-source domains (Aβ-PET and high-resolution MRI) as input and employed the L1 consistency loss function, with all other experimental settings remaining identical. The experimental results are shown in Table 3.

[0107] Table 3

[0108]

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation, characterized in that, An unsupervised multi-source domain adaptive model for Alzheimer's disease was constructed and trained, and the trained model was used to achieve low-resolution MRI diagnosis. The unsupervised multi-source domain adaptive model for Alzheimer's disease includes a cross-resolution dual-path general feature encoder, a domain-specific attention feature extractor, and a shared diagnostic classifier. The training process of the unsupervised multi-source domain adaptive model for Alzheimer's disease includes: S1. Construct a standard dataset for knowledge transfer, comprising: a first source domain consisting of Aβ-PET images; a second source domain consisting of high-resolution MRI images; and a target domain consisting of low-resolution MRI images. S2. Construct dual-resolution input pairs: For the Aβ-PET image in the first source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair in the first source domain. For the high-resolution MRI of the second source domain, it is downsampled to generate a corresponding low-resolution image, and the two are paired to form a dual-resolution input pair of the second source domain; For low-resolution MRI of the target domain, interpolation is performed to generate corresponding high-resolution images, and the two are paired to form a dual-resolution input pair for the target domain; S3. For each domain's dual-resolution input pair, pass it through a cross-resolution dual-path general feature encoder to obtain perceptual features; the cross-resolution dual-path general feature encoder includes a high-resolution feature coding branch and a low-resolution feature coding branch, the two feature coding branches are in parallel structure, their network structures are the same but their weight parameters are independent of each other; wherein, the high-resolution image is processed by the high-resolution feature coding branch to generate a high-resolution feature pyramid; the low-resolution image is processed by the low-resolution feature coding branch to generate a low-resolution feature pyramid; the perceptual features are composed of the above-mentioned high-resolution and low-resolution feature pyramids. S4. Input the perceptual features of the two source domains into the domain-specific attention feature extractor of the corresponding domain to obtain the refined features; the refined features include high-resolution refined features and low-resolution refined features; S5. Input the perceptual features of the target domain into the domain-specific attention feature extractors of the two source domains respectively: obtain the structural features through the processing of the domain-specific attention feature extractor corresponding to the second source domain; obtain the pseudo-molecule features through the processing of the domain-specific attention feature extractor corresponding to the first source domain. S6. After flattening the structural features, pseudo-molecular features, and refined features from the two source domains, they are input into the shared diagnostic classifier, which outputs four prediction results: anatomical high-resolution prediction, anatomical low-resolution prediction, high-resolution molecular prediction, and low-resolution molecular prediction. S7. Based on the four prediction results, calculate the multi-objective joint loss, optimize the model parameters through the backpropagation algorithm, and iterate the training until the model converges; the multi-objective joint loss includes supervised classification loss, multi-source cross-domain distribution alignment loss, and multi-branch prediction consistency constraint loss.

2. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 1, characterized in that, The feature encoding branch uses the ResNet3D network as its backbone and combines it with a three-dimensional feature pyramid to achieve multi-scale feature fusion. This includes: extracting features L2, L3, and L4 from the C2, C3, and C4 stages of the ResNet3D network, aligning the channel dimensions through 1×1×1 convolutions to obtain features L'2, L'3, and L'4; upsampling feature L'4 through trilinear interpolation and adding it element-wise with feature L'3 to obtain feature P3; upsampling feature P3 again and adding it element-wise with feature L'2 to obtain feature P2; finally, the output is a feature pyramid composed of features P2, P3, and L'4.

3. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 1, characterized in that, Both source domain-specific attention feature extractors include two feature extractors: one high-resolution and one low-resolution. The four feature extractors have the same structure.

4. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 3, characterized in that, The feature extractor consists of three convolutional layers and one modality-aware collaborative attention module cascaded together, with a residual connection between the input of the feature extractor and the output of the modality-aware collaborative attention module.

5. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 4, characterized in that, The feature extractor's processing steps include: The input features are processed through three convolutional layers to obtain primary features. ; primary features The channel attention weights are obtained through the modality-aware channel attention submodule. , represented as: ， In the formula, GAP represents global average pooling. Represents element-wise product. This represents the Sigmoid function. Represents the ReLU function. , Indicates the weights of the fully connected layer; This represents the modal prior embedding vector, which is a learnable parameter; Channel attention weights With primary features After multiplication, the input is given to the spatial attention submodule to obtain the spatial attention weights. , represented as: ， In the formula, DeformConv3 represents three-dimensional deformable convolution. Represents the Sigmoid function; The final output of the feature extractor .

6. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 1, characterized in that, The joint loss of multiple objectives is expressed as: ， ， In the formula, Represents dynamic weighting factors. Indicates the loss of the supervised classification. This represents the multi-source cross-domain distribution alignment loss. This represents the distribution alignment loss between the first source domain and the target domain. This represents the distribution alignment loss between the second source domain and the target domain. This represents the loss due to the consistency constraint in multi-branch predictions.

7. The low-resolution MRI diagnostic method for Alzheimer's disease based on multi-source domain adaptation according to claim 6, characterized in that, Supervision and classification of losses The expression used to guide the domain-specific attention feature extractor to learn core semantic features strongly correlated with Alzheimer's disease diagnosis and suppress interference from redundant noise features is: ， Distribution Alignment Loss The expression used to constrain the distribution of pseudo-molecular features generated in the target domain to be consistent with the distribution of molecular features in the first source domain, thereby enabling cross-modal transfer of molecular pathology knowledge from the first source domain to the target domain, is as follows: ， Distribution Alignment Loss To ensure that the structural features generated in the target domain are consistent with the distribution of structural features in the second source domain, and to achieve the transfer of structural shrinkage knowledge across resolution and center-of-view scenarios, the expression is: ， Multi-branch prediction consistency constraint loss To minimize the difference in prediction results between branches of different resolutions and facilitate the transfer of discriminative knowledge from the source domain to the target domain, the expression is: ， in, This represents the total number of labeled samples in the source domain. This represents the total number of categories in the diagnostic classification task. Indicates the first One-hot encoding of the true labels of the source domain samples Indicates a shared diagnostic classifier for the first... The source domain sample at the th th Predicted probabilities on class Indicates the number of samples in the first source domain. This represents the high-resolution refined feature of the i-th sample in the first source domain. This indicates the number of unlabeled samples in the target domain. This represents the low-resolution refined feature of the i-th sample in the first source domain. This represents the high-resolution pseudo-molecule feature of the j-th sample in the target domain. This represents the low-resolution pseudo-molecule feature of the j-th sample in the target domain. Indicates the number of samples in the second source domain. Represents the regenerated nucleus Hilbert space Feature mapping in; This represents the high-resolution refined feature of the i-th sample in the second source domain. This represents the high-resolution structural features of the j-th sample in the target domain. This represents the low-resolution refined feature of the i-th sample in the second source domain. Represents the low-resolution structural features of the j-th sample in the target domain; It is an L1 norm. This represents the high-resolution anatomical prediction result corresponding to the j-th sample in the target domain. This represents the low-resolution anatomical prediction result corresponding to the j-th sample in the target domain. This represents the high-resolution molecular prediction result corresponding to the j-th sample in the target domain. This represents the low-resolution molecular prediction result corresponding to the j-th sample in the target domain.

Images