An alzheimer's disease early diagnosis method and device based on a diffusion model

Abstract

The application discloses an Alzheimer's disease early diagnosis method and device based on a diffusion model, and the method comprises the following steps: acquiring and preprocessing a plurality of brain MRI images of a diagnosis object at different periods; obtaining MRI compression features corresponding to each brain MRI image by using an encoder in a trained auto-encoder network; taking the MRI compression features as a conditional variable, and performing prediction by using a trained diffusion model to obtain MRI prediction features; acquiring a brain MRI image of the diagnosis object obtained by performing MRI scanning again, obtaining MRI real features by using a trained encoder, taking an error between the MRI real features and the MRI prediction features as an abnormal score, and determining that the diagnosis object is positive for Alzheimer's disease if the abnormal score exceeds a predetermined threshold. The method has the advantages of high calculation efficiency, strong interpretability and the like by combining the generation capability of the diffusion model and the time series modeling advantage, and can realize early diagnosis of Alzheimer's disease only by using healthy population data.

Description

Technical Field

[0001] The present invention relates to the field of computer image processing, and in particular to a method and device for early diagnosis of Alzheimer's disease based on a diffusion model. Background Art

[0002] Alzheimer's disease (AD) is a common neurodegenerative disorder with clinical manifestations such as memory impairment, decreased executive function, and personality changes, which severely impact patients' quality of life. Studies have shown that subtle changes in brain structure begin to occur in the early stages of the disease. Magnetic resonance imaging (MRI), as a non-invasive imaging method, can clearly demonstrate phenomena such as brain atrophy and ventricular enlargement. Consequently, MRI-based early diagnosis strategies have become a research hotspot. The rise of artificial intelligence, particularly deep learning methods, has provided new directions for medical image modeling, with widespread applications in tasks such as brain region identification, image enhancement, and anomaly detection. In recent years, diffusion models have also demonstrated great potential in medical image analysis.

[0003] In the process of implementing the present invention, the inventors discovered that the prior art has at least the following problems:

[0004] First, existing methods mostly model static MRI images and lack the ability to model the temporal evolution of continuous MRI scans, making it difficult to identify potential abnormalities in structural change trends (Yang, Y., Fu, H., Aviles-Rivero, AI, et al. 2023. DiffMIC: Dual-guidance diffusion network formedical image classification. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 95-105.).

[0005] Second, although some studies have attempted to adopt unsupervised strategies, they still rely heavily on pseudo-annotations, structural segmentation maps, or expert rules as priors during the training phase, which limits the versatility of the model in real unlabeled scenarios (Siddiquee, MMR, Shah, J., Wu, T., et al. 2024. Brainomaly: Unsupervised neurologic disease detection utilizing unannotated t1-weighted brain MR images. InProceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision, 7573-7582.).

[0006] Third, most current diffusion models are directly modeled in high-dimensional image space, which has high computational complexity and slow inference speed, making it difficult to deploy in actual scenarios with limited clinical resources (Rombach, R., Blattmann, A., Lorenz, D., et al. 2022. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition, 10684-10695.). Summary of the Invention

[0007] The purpose of the present invention is to address the shortcomings of the existing technology and provide a method and device for early diagnosis of Alzheimer's disease based on a diffusion model. By combining latent space compression, time series dynamic prediction and unsupervised anomaly detection, the advantages of the diffusion model in image generation and structural modeling are fully utilized to achieve efficient, explainable and low-cost implementation of early diagnosis of Alzheimer's disease.

[0008] According to a first aspect of an embodiment of the present application, a method for early diagnosis of Alzheimer's disease based on a diffusion model is provided, comprising:

[0009] Acquire several brain MRI images of the diagnosed subject at different times and perform preprocessing;

[0010] Based on the preprocessed brain MRI images, the encoder in the trained autoencoder network is used to obtain the MRI compression features corresponding to each brain MRI image;

[0011] The MRI compression features extracted from several brain MRI images of the diagnosed subjects at different times are used as conditional variables, and the trained diffusion model is used for prediction to obtain the MRI prediction features.

[0012] A brain MRI image of the diagnosed subject is obtained by another MRI scan. The trained encoder obtains the MRI true features, compares the MRI true features with the MRI predicted features, and uses the error between the two as an abnormality score. If the abnormality score exceeds a predetermined threshold, it is determined to be positive for Alzheimer's disease.

[0013] Furthermore, the preprocessing process of the brain MRI image includes:

[0014] Image space normalization: aligning all brain MRI images to the same spatial template;

[0015] Image cropping: cropping the image boundaries to unify all brain MRI images to the same size;

[0016] Intensity normalization: The grayscale distribution of the brain tissue area in each image is counted, and the grayscale of the entire image is remapped using a linear transformation function with a mean of 0 and a standard deviation of 1.

[0017] Furthermore, the autoencoder network includes an encoder and a decoder, wherein the encoder is used to compress the MRI brain image layer by layer to obtain MRI compression features, and the decoder is used to reconstruct the image according to the MRI compression features;

[0018] The trained autoencoder network is trained on a preprocessed MRI image training set. During the training process, the loss function includes a reconstruction loss and a KL penalty term, wherein the reconstruction loss is calculated by comparing the error between the original image and the decoder reconstructed image, and the KL penalty term is used to constrain the distribution of the encoder output features.

[0019] Furthermore, the feature vector output by the encoder is formatted through a spatial alignment and quantization strategy to obtain the MRI compression feature.

[0020] Furthermore, the diffusion model introduces a cross-attention module into the U-Net network, and the cross-attention module dynamically weights the features output by each upsampling layer or downsampling layer in the U-Net network.

[0021] Furthermore, during the training process of the diffusion model, Gaussian noise is added to the original data multiple times, the difference between the noise predicted by the model and the actually added Gaussian noise is calculated, and the difference is used as the noise loss function of the diffusion model to further train and optimize the diffusion model.

[0022] Furthermore, it also includes:

[0023] Several time-series brain MRI image combinations are constructed for the diagnostic object, and an abnormality score is calculated for each combination. The evolution trajectory of the brain structure is evaluated by integrating the trends of each abnormality score to assist doctors in making further diagnosis.

[0024] According to a second aspect of an embodiment of the present application, a computer program product is provided, comprising a computer program / instruction, which implements the method described in the first aspect when executed by a processor.

[0025] According to a third aspect of the embodiments of the present application, there is provided an electronic device, including:

[0026] one or more processors;

[0027] a memory for storing one or more programs;

[0028] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in the first aspect.

[0029] According to a third aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which computer instructions are stored. When the instructions are executed by a processor, the steps of the method described in the first aspect are implemented.

[0030] The technical solution of the present invention is summarized as follows:

[0031] This paper provides a diffusion-based approach to early diagnosis of Alzheimer's disease, centered around a comprehensive spatiotemporal feature analysis and anomaly detection system. This approach utilizes an innovative neural network architecture to accurately model the dynamic changes in brain structure. Specifically, it employs an end-to-end computational framework: First, a deep convolutional autoencoder is used to extract features and reduce the dimensionality of three-dimensional brain MRI images, mapping the high-dimensional image data into a low-dimensional latent space. A conditional diffusion model with time series modeling capabilities is then designed. This model uses historical scan data as conditional input via a specific network connection, and utilizes an improved attention mechanism to fuse and transfer spatiotemporal features, ultimately outputting predictions of brain states at future time points.

[0032] At the implementation level, the solution establishes a complete computational process: starting with standardized preprocessing of raw image data, through a complete chain of feature encoding, temporal diffusion modeling, anomaly detection, and visualization output. The feature encoding module is constructed using a three-dimensional convolutional neural network, preserving spatial structural information through specific inter-layer connections. The conditional diffusion module uses a carefully designed network topology to fuse temporal conditional information with the current state at multiple scales. The anomaly detection module calculates prediction deviations based on a specific distance metric algorithm and employs an adaptive threshold strategy for judgment. Through the coordinated coordination of data and control flows between modules, early identification and quantitative assessment of brain degenerative changes are achieved.

[0033] The beneficial effect of the present invention is that, by adopting the time series dynamic modeling technology, it is possible to more accurately capture the characteristics of the gradual changes in brain structure and greatly improve the sensitivity and specificity of early diagnosis. The innovative unsupervised learning paradigm only requires data from healthy people to train the model, which effectively solves the problem of scarce medical annotated data. The combination of latent space calculation and conditional diffusion prediction significantly reduces the demand for computing resources while ensuring diagnostic accuracy. The introduction of the cross-attention mechanism enables the model to have the ability to fuse spatiotemporal features, and the generated prediction results have better interpretability, which can provide clinicians with intuitive lesion development analysis. This method not only achieves early diagnosis, but also can dynamically monitor the evolution of the disease, providing a scientific basis for the formulation of personalized diagnosis and treatment plans, and has important clinical application value.

[0034] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0036] Figure 1 The figure is a flow chart showing a method for early diagnosis of Alzheimer's disease based on a diffusion model according to an exemplary embodiment.

[0037] Figure 2 is a structural diagram of an autoencoder network according to an exemplary embodiment.

[0038] Figure 3 is a structural diagram of a diffusion model according to an exemplary embodiment.

[0039] Figure 4 The present invention is a block diagram of an early diagnosis device for Alzheimer's disease based on a diffusion model according to an exemplary embodiment.

[0040] Figure 5The figure is a schematic diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0041] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with this application.

[0042] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0043] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0044] The core technology of this invention is a time-series dynamic modeling and unsupervised anomaly detection technology based on a diffusion model, which enables accurate early diagnosis of Alzheimer's disease. By constructing a conditional diffusion model to capture the gradual changes in brain structure and combining it with latent space computational optimization technology, the sensitivity and specificity of early diagnosis are significantly improved.

[0045] like Figure 1 As shown, the present invention proposes an early diagnosis method for Alzheimer's disease based on a diffusion model, comprising the following steps:

[0046] (1) Obtain several brain MRI images of the diagnosed subject at different times and perform preprocessing;

[0047] Specifically, two or more MRI images of the diagnostic subject at different times may be selected. In this embodiment, three MRI scan images of the diagnostic subject are preferably selected. A wide range of data sets meet this requirement, and the three images provide sufficient basis for predicting the fourth image, which can perform better than a prediction network composed of fewer MRI images. The preprocessing process may include the following sub-steps:

[0048] (1.1) Image spatial normalization: After the brain MRI images are acquired, all MRI images must first be spatially normalized, that is, aligned to the same spatial template. To this end, the SyN algorithm (Avants, BB, Epstein, CL, Grossman, M., and Gee, JC 2008. Symmetricdiffeomorphic image registration with cross-correlation: Evaluating automatedlabeling of elderly and neurodegenerative brain. Medical Image Analysis 12(1), 26-41.), which combines rigid and nonlinear registration, is used to map the original MRI images to a standard anatomical template (such as MNI space). This process ensures the consistency of the spatial position of the brain images, especially the stable positioning of key areas such as the temporal lobe and hippocampus in the three-dimensional coordinate system, which allows the subsequent feature compression model to focus on subtle changes in physiological structure rather than image displacement.

[0049] (1.2) Image Cropping: MRI images may have margin gaps or structural variations. To unify the input size, we further automatically crop the image boundaries and resize all images to a 160×160×160 three-dimensional tensor based on the diffusion model input size. The cropping region is determined by the distribution of grayscale information in the image, ensuring that the retained region fully covers key structures such as the left and right hemispheres and the hippocampus. In practice, center cropping is typically performed to ensure that the required information is within the cropped region. This process not only improves the input efficiency of the model but also significantly reduces the redundant computation caused by excess background during neural network training.

[0050] (1.3) Intensity Normalization: After spatial and structural alignment, the uniformity of image signal intensity becomes a key factor influencing the stability of model learning. Therefore, images need to be intensity normalized. By statistically analyzing the grayscale distribution of brain tissue regions in each image, a linear transformation function with a mean of 0 and a standard deviation of 1 is used to remap the grayscale of the entire image, thereby eliminating signal drift caused by differences in equipment and acquisition parameters between individuals and time series. The standardized image has a uniform contrast scale, which is more conducive to the neural network extracting common structural features across diagnostic subjects.

[0051] According to the above method, MRI images are highly unified in multiple dimensions such as spatial position, structural boundary, image resolution and signal intensity, which is convenient for subsequent input into the model.

[0052] (2) Based on the preprocessed brain MRI images, the encoder in the trained autoencoder network is used to obtain the MRI compression features corresponding to each brain MRI image;

[0053] like Figure 2 As shown, the autoencoder network consists of two sub-networks: an encoder and a decoder. The encoder, with a 3D convolutional layer at its core, combines strided convolution with a downsampling mechanism to compress the input MRI image layer by layer. Through continuous spatial compression, the original high-dimensional image is ultimately mapped to a low-dimensional 16×16×16 latent tensor. This tensor not only preserves the spatial topology of the entire brain but also aggregates morphological information of key neural regions, providing a compact and semantically rich feature representation for subsequent modeling. The decoder, in contrast, employs a symmetric 3D deconvolutional architecture, progressively restoring the spatial resolution of the latent features through upsampling layers. Each upsampling module consists of a transposed convolutional layer, a batch normalization layer, and an activation function, layer by layer restoring the 16×16×16 latent tensor to the original 160×160×160 voxel structure. In this process, the decoder not only reconstructs the overall shape of the image but also restores cortical details and boundary structures, minimizing information loss during compression. The final output layer uses a sigmoid activation function to normalize the reconstructed image intensity range, ensuring consistent contrast with the original MRI image.

[0054] The encoder is represented as , the decoder is represented as , which acts on the original image and noise between, so that:

[0055]

[0056] Right now,

[0057] in, is the feature vector output by the encoder, The reconstructed data output by the decoder.

[0058] Before using the encoder to compress the features of the MRI image of the diagnostic object, the network needs to be trained, and the training set is processed by the preprocessing process of step (1). In order to make the features output by the encoder have good distribution consistency and reconstruction ability, the KL divergence is introduced as a regularization term in the training process to guide the encoding space to approach the standard normal distribution. The specific training objective function consists of reconstruction loss and KL penalty term, in which the reconstruction part is calculated by comparing the mean square error between the original image and the decoded reconstructed image, and the KL penalty term is used to constrain the distribution form of the encoder output features, thereby improving the expression ability and generalization performance of the model during the compression process. Overall loss function Expressed as:

[0059]

[0060] in, is the weight used to balance the reconstruction loss and KL penalty term, represents the operation of finding the KL divergence of two distributions, and represents the probability distribution of feature vector z given image x, The mean is 0 and the standard deviation is Gaussian distribution.

[0061] After the autoencoder network is trained using the overall loss function, in the inference phase, only the encoder part is retained to participate in the calculation, and the decoder does not participate in the prediction, further improving the computational efficiency.

[0062] After encoding, the encoder can further normalize the features using spatial alignment and quantization strategies to obtain compressed MRI features. This ensures that different images are completely consistent in feature dimensions and scales, providing a good basis for comparison. These compressed MRI features can be stored as independent representation vectors and passed as input to subsequent conditional diffusion models for time series modeling and prediction.

[0063] (3) The MRI compression features extracted from several brain MRI images of the diagnosed subjects at different times are used as conditional variables, and the trained diffusion model is used for prediction to obtain MRI prediction features;

[0064] After acquiring compressed features from multiple MRIs of the same patient at different time periods, the present invention uses a conditional diffusion model incorporating a cross-attention mechanism to model the structural evolution of the time series and predict the brain MRI feature representation of the patient at the next time point (i.e., the fourth scan). The entire prediction process is based on a latent space feature generation mechanism to ensure that the generated results are structurally biologically plausible and temporally continuous.

[0065] The diffusion model adopts the DDPM (Denoising Diffusion Probabilistic Model) structure to achieve conditional guidance of the prediction process, such as Figure 3As shown, the present invention introduces a cross-attention module after each upsampling / downsampling layer of the diffusion model's core network, the U-Net. This module dynamically weights features at each layer, calculating the attention weight between the current prediction state and historical conditions, integrating the structural information of the time series and effectively capturing the characteristic trends of the brain over time. The attention mechanism is calculated based on a query-key-value triple mapping, and the generation strategy is dynamically adjusted at each diffusion step, making the prediction results more consistent with the evolution of the historical sequence.

[0066] The attention mechanism is calculated as follows:

[0067]

[0068] Where Q represents the query (from the predicted state), K and V are the key and value (generated from the historical condition input).

[0069] For the diffusion model, the MRI compression features extracted from n consecutive MRI scans of the same diagnostic object are used as time series input in the format of ,in Indicates the The potential representation after the scan compression is , is the number of MRI compression features. This sequence constitutes the conditional variable c of the diffusion model, which is used to provide a guiding signal in the evolution direction during the generation process, that is, to provide control noise. All sequence inputs are embedded into the context encoding space of the diffusion model through a specific linear mapping module to adapt to the cross-attention module of the subsequent network.

[0070] The first step in training the diffusion model is to add noise to the sample. During the training phase, multi-step Gaussian noise perturbations are applied to the target features in the latent space:

[0071]

[0072] in, is the diffusion model noise loss function, is the Gaussian noise added to the sample, is the model’s predicted output for the feature, To add noise to the original features The value after the step, is the number of diffusion steps.

[0073] The neural network learns the inverse transformation process of each step of the noise and models the generative path from noise to the original structure. The training goal is to minimize the noise prediction error so that the prediction representation of the target structure can be gradually restored from the random noise in the inference stage. .

[0074] (4) Obtaining a brain MRI image of the diagnosed subject obtained by another MRI scan, obtaining the MRI true features from the trained encoder, comparing the MRI true features with the MRI predicted features, and using the error between the two as an abnormality score. If the abnormality score exceeds a predetermined threshold, it is determined to be positive for Alzheimer's disease;

[0075] After generating the fourth MRI features for the patient, the generated features are compared with the actual MRI compression features to assess whether the patient has significant structural abnormalities at the current time point and, based on this, determine whether the patient is positive for Alzheimer's disease. This process uses an unsupervised anomaly detection strategy, avoiding reliance on explicit pathological annotations during the training phase and improving the model's adaptability in real clinical scenarios. Specifically, it includes the following sub-steps:

[0076] (4.1) Calculation of the error between prediction and true features

[0077] The prediction features generated by the diffusion model The features actually compressed by the encoder Perform element-by-element comparison and use the mean square error (MSE) metric as the anomaly scoring function. The MSE metric is calculated as follows:

[0078]

[0079] in, is the total dimension of the feature vector, Indicates the The true value of the eigenvector, and Indicates the The predicted value of the feature vector, Represents the eigenvector The predicted values ​​in each dimension.

[0080] This error metric not only reflects the model's ability to fit the structural trends of the time series but also serves as a quantitative indicator for evaluating the degree of deviation between the current state and the healthy model. Larger error values ​​indicate that the generative model is unable to fit the current structure under the healthy prior, indicating a higher probability of anomaly.

[0081] (4.2) Threshold determination and positive diagnosis strategy: The present invention sets abnormal threshold parameters , calculate the prediction error for each diagnosis object, when the error value is greater than When the error is below a threshold, the system automatically marks it as abnormal, indicating a positive diagnosis for Alzheimer's disease. When the error is below a threshold, the structural changes are considered consistent with a healthy pattern and are classified as negative. This threshold is determined through performance optimization on the validation set and can be adaptively adjusted based on the distribution of different populations. This strategy automatically identifies early signs of brain structural variation without relying on manual annotation, enabling efficient and sensitive unsupervised diagnosis.

[0082] (5) Sequence review and dynamic trend analysis: To further improve the stability and interpretability of diagnosis, multiple time series combinations can be constructed for the same diagnostic object, and prediction and error evaluation operations can be performed on them separately. By integrating the trends of multiple groups of abnormal scores, doctors can be assisted in judging the long-term evolution trajectory of brain structure. In addition, combined with the visualization of MRI reconstruction results, it helps clinicians accurately locate the lesion area and provide further intervention recommendations.

[0083] In one embodiment, the present invention is implemented on a machine with an NVIDIA A100 graphics card with 32GB GPU memory for the ADNI dataset. The Adam optimizer is used to train the autoencoder and the conditional diffusion model, respectively, with the initial learning rate set to The total number of training rounds is set to 200. The autoencoder is trained unsupervised using only the data of healthy controls as input. The compressed latent feature dimension is 16×16×16, and the latent space KL penalty weight is set to The diffusion model adopts the classic DDPM architecture, with the number of noise steps T=1000 and the network size of 115.1M parameters. It embeds the cross-attention mechanism to strengthen the temporal conditional dependency.

[0084] During the testing phase, each subject's first three MRI scans were compressed and used as conditional input to predict the latent feature representation of their next MRI image. The resulting encoding was compared with the actual scan image, and the mean squared error (MSE) was calculated as an anomaly score. When the MSE value exceeded a preset threshold of λ=1.2, the sample was considered likely to be positive for Alzheimer's disease.

[0085] In a 5-fold cross-validation analysis on the ADNI dataset, the model achieved a classification accuracy of 0.807 and a specificity of 0.855 without using pathological labels, demonstrating high sensitivity and generalization. Visualization results showed that the model's predicted abnormalities were typically concentrated in typical Alzheimer's disease manifestations, such as ventriculomegaly and hippocampal atrophy, validating the method's potential for practical application in detecting structural brain variations.

[0086] f-AnoGAN (Schlegl T, Seeböck P, Waldstein SM, et al. f-AnoGAN: Fastunsupervised anomalydetection with generative adversarial networks. Medicalimage analysis, 2019, 54: 30-44.), ALOOC (M. Sabokrou, M. Khalooei, M. Fathy, and E. Adeli, “Adversarially learned one-class classifier for noveltydetection,” in Proceedings of the IEEE Conference on Computer Vision andPattern Recognition, 2018, pp. 3379–3388.), ALAD (Zenati H, Romain M, Foo CS, et al. Adversarially learned anomaly detection, 2018 IEEE Internationalconference on data mining (ICDM). IEEE, 2018: 727-736.),ADDFORMER (R. Kushol,A. Masoumzadeh, D. Huo, S. Kalra and Y. -H. Yang, "Addformer: Alzheimer's Disease Detection from Structural MRI Using Fusion Transformer," 2022 IEEE19th International Symposium on Biomedical Imaging (ISBI), Kolkata, India, 2022, pp. 1-5) and this method were respectively applied to the ADNI dataset for experiments. The results are shown in Table 1, which compares the accuracy and sensitivity of the experiments with those of related methods.

[0087] Table 1

[0088]

[0089] Corresponding to the aforementioned embodiment of the method for early diagnosis of Alzheimer's disease based on a diffusion model, the present application also provides an embodiment of an apparatus for early diagnosis of Alzheimer's disease based on a diffusion model.

[0090] Figure 4 FIG1 is a block diagram of an early diagnosis device for Alzheimer's disease based on a diffusion model according to an exemplary embodiment. Figure 4 , the apparatus may include:

[0091] An acquisition module 21 is used to acquire and pre-process a number of brain MRI images of the diagnosed subject at different times;

[0092] A feature compression module 22 is configured to obtain MRI compression features corresponding to each brain MRI image based on the preprocessed brain MRI image using an encoder in the trained autoencoder network;

[0093] The prediction module 23 is configured to use the MRI compression features extracted from a plurality of brain MRI images of the diagnosis subject at different time periods as conditional variables and perform prediction using the trained diffusion model to obtain MRI prediction features;

[0094] The comparison module 24 is used to obtain a brain MRI image of the diagnosed subject obtained by another MRI scan, and the trained encoder obtains the MRI true features, compares the MRI true features with the MRI predicted features, and uses the error between the two as an abnormality score. If the abnormality score exceeds a predetermined threshold, it is determined to be positive for Alzheimer's disease.

[0095] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0096] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application scheme. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0097] Accordingly, the present application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-mentioned diffusion model-based early diagnosis method for Alzheimer's disease.

[0098] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method for early diagnosis of Alzheimer's disease based on the diffusion model. Figure 5 As shown in FIG, a hardware structure diagram of an early diagnosis device for Alzheimer's disease based on a diffusion model provided by an embodiment of the present invention is provided in any device with data processing capability, except Figure 5 In addition to the processor, memory, and network interface shown, any device with data processing capabilities in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0099] Accordingly, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-mentioned method for diagnosing early Alzheimer's disease based on a diffusion model. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0100] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the contents disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art that are not disclosed in this application.

Claims

1. A method for early diagnosis of Alzheimer's disease based on a diffusion model, characterized in that: include: Acquire several brain MRI images of the diagnosed subject at different times and perform preprocessing; Based on the preprocessed brain MRI images, the encoder in the trained autoencoder network is used to obtain the MRI compression features corresponding to each brain MRI image; The MRI compression features extracted from several brain MRI images of the diagnosed object at different times are used as conditional variables, and the trained diffusion model is used for prediction to obtain the MRI prediction features. For the diffusion model, the MRI compression features extracted from n MRI scans of the same diagnosed object are used as time series input. , generate prediction features ; A brain MRI image of the diagnosed subject is obtained by another MRI scan. The trained encoder obtains the MRI true features, compares the MRI true features with the MRI predicted features, and uses the error between the two as an abnormality score. If the abnormality score exceeds a predetermined threshold, it is determined to be positive for Alzheimer's disease.

2. The method according to claim 1, characterized in that The preprocessing process of the brain MRI image includes: Image space normalization: aligning all brain MRI images to the same spatial template; Image cropping: cropping the image boundaries to unify all brain MRI images to the same size; Intensity normalization: The grayscale distribution of the brain tissue area in each image is counted, and the grayscale of the entire image is remapped using a linear transformation function with a mean of 0 and a standard deviation of 1.

3. The method according to claim 1, characterized in that The autoencoder network includes an encoder and a decoder, wherein the encoder is used to compress the MRI brain image layer by layer to obtain MRI compression features, and the decoder is used to reconstruct the image according to the MRI compression features; The trained autoencoder network is trained on a preprocessed MRI image training set. During the training process, the loss function includes a reconstruction loss and a KL penalty term, wherein the reconstruction loss is calculated by comparing the error between the original image and the decoder reconstructed image, and the KL penalty term is used to constrain the distribution of the encoder output features.

4. The method according to claim 1, wherein The feature vector output by the encoder is formatted through spatial alignment and quantization strategy to obtain the MRI compression feature.

5. The method according to claim 1, wherein The diffusion model introduces a cross-attention module into the U-Net network, which dynamically weights the features output by each upsampling layer or downsampling layer in the U-Net network.

6. The method according to claim 1, characterized in that During the training process of the diffusion model, Gaussian noise is added to the original data multiple times, the difference between the noise predicted by the model and the actually added Gaussian noise is calculated, and the difference is used as the noise loss function of the diffusion model to further train and optimize the diffusion model.

7. The method according to claim 1, characterized in that Also includes: Several time-series brain MRI image combinations are constructed for the diagnostic object, and an abnormality score is calculated for each combination. The evolution trajectory of the brain structure is evaluated by integrating the trends of each abnormality score to assist doctors in making further diagnosis.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

9. An electronic device, characterized in that: include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images