A predictive system for the probability of mild cognitive impairment converting to Alzheimer's disease

By using PET image standardization preprocessing and deep transfer characterization techniques, combined with ECA channel attention enhancement, GeM adaptive global feature aggregation, and PCA dimensionality reduction, the problems of inconsistent image preprocessing and high-dimensional feature redundancy in MCI-to-AD prediction are solved, achieving individualized and stable AD conversion probability prediction.

CN122575748BActive Publication Date: 2026-11-17CHANGCHUN UNIV OF SCI & TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611064634.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-17
Publication Date
2026-11-17
Estimated Expiration
2046-07-17

AI Technical Summary

Technical Problem

Existing methods for predicting the probability of conversion from MCI to AD have problems such as inconsistent PET image preprocessing, redundancy of high-dimensional deep features, insufficient utilization of AD phenotypic probability, and insufficient stability of individual conversion prediction probability output.

Method used

By employing PET image standardization preprocessing, residual convolutional transfer backbone, ECA channel attention enhancement, GeM adaptive global feature aggregation, and PCA dimensionality reduction techniques, combined with AD phenotypic probability fusion, a probability prediction system for mild cognitive impairment (MCI) transforming into Alzheimer's disease (AD) is constructed, outputting the predicted probability of an individual with MCI transforming into AD in the future.

Benefits of technology

It improves the consistency of PET image input, reduces interference in model training, enhances the ability to express disease-related image features, reduces the risk of high-dimensional feature redundancy and small sample overfitting, outputs continuous and individualized AD conversion prediction probabilities, and provides auxiliary reference for individual MCI conversion trends.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122575748B_ABST
    Figure CN122575748B_ABST
Patent Text Reader

Abstract

This invention relates to a system for predicting the probability of mild cognitive impairment (MCI) progressing to Alzheimer's disease (AD). It falls under the technical fields of intelligent medical image analysis and auxiliary diagnosis of neurodegenerative diseases. The system addresses existing problems in predicting the probability of MCI progressing to AD, including insufficient standardization of PET image preprocessing, redundancy of high-dimensional deep features, inadequate utilization of AD phenotypic probabilities, and insufficient stability of individual conversion prediction probabilities. Based on standardized PET image preprocessing, a transfer learning model is used to extract stable deep image representations. Through ECA channel attention enhancement, GeM adaptive global feature aggregation, PCA dimensionality reduction, and AD phenotypic probability fusion, the system outputs the predicted probability of an MCI individual progressing to AD. This provides an auxiliary reference for judging the conversion trend and follow-up management of MCI individuals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image analysis and auxiliary diagnosis technology for neurodegenerative diseases, specifically involving a system for predicting the probability of mild cognitive impairment turning into Alzheimer's disease. Background Technology

[0002] Alzheimer's disease (AD) is a common neurodegenerative disease, primarily characterized by memory loss, cognitive decline, and impairment of daily living abilities. Mild Cognitive Impairment (MCI) is an important transitional stage between normal aging and AD, but clinical outcomes vary significantly among individuals with MCI. Some individuals with MCI remain stable during follow-up, termed stable mild cognitive impairment (sMCI), while others progress to AD, termed progressive mild cognitive impairment (pMCI). Therefore, predicting the likelihood of MCI individuals developing AD based on imaging characteristics at baseline is a crucial issue in early AD identification and MCI follow-up management.

[0003] PET imaging can reflect changes in brain region metabolism, tracer uptake, or pathological burden distribution, serving as crucial imaging evidence in the auxiliary diagnosis of Alzheimer's disease (AD) and the prediction of MCI conversion probability. PET imaging can reveal whether individuals with MCI have developed imaging patterns approaching AD, thus providing information for predicting their future conversion trend. However, PET images often vary among different subjects in terms of spatial location, voxel scale, overall uptake level, and extracerebral background. Without a standardized preprocessing procedure, models can easily learn noise information irrelevant to the disease, thereby reducing the reliability of prediction results.

[0004] Existing methods for predicting AD conversion using MCI still have the following shortcomings: Firstly, some methods lack uniformity in PET image preprocessing, failing to adequately eliminate spatial and intensity distribution differences among different subjects. Secondly, deep learning models extract high-dimensional image features with high dimensionality, while the sample size for MCI conversion prediction is usually limited; directly inputting high-dimensional features into the prediction model easily leads to feature redundancy and overfitting. Furthermore, existing methods often focus on the final classification label, failing to fully utilize the AD phenotypic probabilities output by the AD and MCI binary classification models, resulting in insufficient representation of the degree to which an individual's image pattern is close to AD. Some methods primarily output pMCI and sMCI category judgments, lacking continuous conversion prediction probability outputs, making it difficult to intuitively reflect the differences in conversion probability between individuals with different MCI levels.

[0005] Therefore, existing MCI conversion AD probability prediction has problems such as insufficient standardization of PET image preprocessing, redundancy of high-dimensional deep features, insufficient utilization of AD phenotypic probability, and insufficient stability of individual conversion prediction probability output. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a system for predicting the probability of mild cognitive impairment (MCI) progressing to Alzheimer's disease (AD). Based on standardized preprocessing of PET images, it utilizes a transfer learning model to extract stable depth image representations. Through ECA channel attention enhancement, GeM adaptive global feature aggregation, PCA dimensionality reduction, and AD phenotypic probability fusion, it outputs the predicted probability of an MCI individual progressing to AD. Simultaneously, by comparing the predicted probability distributions of sMCI and pMCI, the model's ability to distinguish the probabilities of MCI individuals at different progression stages is evaluated, providing auxiliary reference for judging the progression trend and follow-up management of MCI individuals.

[0007] The system specifically includes: Data Acquisition and Indexing Module: Acquires PET image data and establishes a traceable data index based on the subject as the basic unit; PET image preprocessing module: Standardizes PET images, calculates SUVR of cerebellar gray matter reference area, extracts brain tissue mask and normalizes intensity to obtain a unified model input image; Transfer Representation Extraction Module: Construct a deep transfer representation model based on residual convolutional transfer backbone, ECA channel attention enhancement module and GeM adaptive global feature aggregation module. After receiving input images, complete AD and MCI binary classification training, and extract patient-level deep image features and AD phenotype probabilities. Dimensionality reduction and fusion module: Reads the patient-level migration feature table of MCI, identifies sMCI and pMCI labels, performs PCA dimensionality reduction on patient-level depth image features, and fuses them with AD phenotype probabilities; Probability Prediction Module: Constructs a probability prediction model for MCI to AD conversion, outputs the predicted probability of each MCI individual converting to AD in the future, and generates two sets of predicted probability distribution results: sMCI and pMCI.

[0008] Furthermore, after acquiring PET image data, it is organized according to diagnostic category, subject number, visit information and image record number, and a traceable data index is established with the subject as the basic unit.

[0009] Furthermore, the cerebellar gray matter reference area SUVR calculation specifically involves: generating a binary mask for the cerebellar gray matter reference area using a brain region template; subsequently, resampling the cerebellar gray matter reference area mask to a standardized PET image space, ensuring consistency with the PET image in matrix size, voxel resolution, and spatial coordinates; then, using the average uptake value of the cerebellar gray matter region as a reference value, converting the original PET signal into an SUVR image with relative quantitative significance. .

[0010] Furthermore, the extraction of brain tissue masks specifically involves: Tissue probability segmentation was performed on SUVR images in standard space to obtain probability maps of gray matter, white matter, cerebrospinal fluid, and other tissues; a brain tissue mask was constructed using the gray matter, white matter, and cerebrospinal fluid probability maps. The brain tissue mask was applied to the SUVR image. Subsequently, SUVR images of extracted brain tissue were obtained. : .

[0011] Furthermore, the data processing flow in the transfer representation extraction module is as follows: First, the two-dimensional PET image is input into the residual convolution transfer backbone to extract deep image features layer by layer, from low-level grayscale distribution to high-level disease semantics; second, after the high-level semantic features are output, an ECA channel attention enhancement module is added to adaptively weight the importance of different channels and enhance the response of key channels related to AD and MCI classification; third, the high-level feature map is aggregated through the GeM adaptive global feature aggregation module to highlight local high-response brain regions while retaining global distribution information, thereby obtaining patient-level deep image features; finally, the aggregated patient-level deep image features are input into the classification head to output the AD and MCI binary classification results and AD phenotype probabilities.

[0012] Furthermore, the classification head includes a flattening layer, a one-dimensional batch normalization layer, a random deactivation layer, a fully connected layer, and a Softmax probability output layer.

[0013] Furthermore, PCA dimensionality reduction is performed on the patient-level depth image features, and then fused with the AD phenotypic probability. The resulting fused feature vector is expressed as follows: ; in, Represents the fused feature vector. to Indicates the front dimension after PCA dimensionality reduction Principal component features, This represents the probability of the AD phenotype. This represents the number of principal components retained. Let the eigenvalues ​​obtained from the PCA on the training set be sorted in descending order as follows: The preset cumulative variance contribution rate threshold is ,but It can be represented as:

[0014] in, Indicates the first The eigenvalues ​​corresponding to each principal component This indicates the number of candidate principal components used in calculating the cumulative variance contribution rate. This represents the original dimension of the patient-level depth image features in the training set, or the maximum number of principal components that can be decomposed.

[0015] Furthermore, in the probability prediction module, the fused feature vector Z obtained after dimensionality reduction by principal component analysis and fusion with the AD phenotypic probability is used as a fixed input. Based on this, a Logistic regression model with L2 regularization is used as a lightweight MCI to AD probability prediction model. Leave-one-out cross-validation is used to train and validate the MCI to AD probability prediction model. Specifically, the training set fused feature vector, the test sample fused feature vector, and the corresponding sMCI and pMCI labels of the training set are received. The MCI to AD probability prediction model is fitted only using the training set fused feature vector and its labels. After the model is trained, the fused feature vector of the current test sample is input into the model, and the predicted probability of the MCI individual being converted to AD is output.

[0016] The system's intended effects are as follows: First, by standardizing PET image spatial processing, calculating the standardized uptake value ratio (SUVR) using cerebellar gray matter as a reference region, extracting brain tissue masks, and normalizing intensity, the system improves the consistency of PET image input from different subjects and reduces the interference of extra-brain background and overall uptake differences on model training. Second, the system extracts multi-level deep features from PET images through residual convolution transfer backbone extraction and introduces an ECA channel attention enhancement module after high-level semantic features, enabling the model to adaptively enhance the responses of key channels related to AD and MCI classification and suppress redundant or weakly correlated features. Third, the system replaces the fixed average aggregation method with a GeM adaptive global feature aggregation module, which enhances the contribution of locally high-response brain regions while preserving global image distribution information, thereby enhancing the model's ability to express the differences between AD and MCI images. Finally, this system integrates PCA principal component features with AD phenotypic probabilities and constructs an MCI-to-AD probability prediction model through leave-one-out cross-validation. This reduces the risk of high-dimensional feature redundancy and small-sample overfitting, outputs continuous and individualized AD conversion prediction probabilities, and demonstrates the model's ability to distinguish the probabilities of MCI individuals in different conversion states through the analysis of the sMCI and pMCI prediction probability distributions. This provides an auxiliary reference for judging the conversion trend of MCI individuals and follow-up management. Attached Figure Description

[0017] Figure 1 This is a flowchart of the PET image standardization preprocessing process in an embodiment of the present invention; Figure 2 This is a diagram of the deep transfer representation model architecture based on the residual convolutional transfer backbone, the ECA channel attention enhancement module, and the GeM adaptive global feature aggregation module in this embodiment of the invention. Figure 3 This is a structural diagram of the ECA channel attention enhancement module in an embodiment of the present invention; Figure 4 This is a structural diagram of the GeM adaptive global feature aggregation module in an embodiment of the present invention; Figure 5 This is a flowchart of the MCI to AD probability prediction model in an embodiment of the present invention; Figure 6 This is a probability distribution diagram of the conversion of individuals with sMCI and pMCI to AD in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0019] Example 1 Existing methods for predicting the conversion probability of mild cognitive impairment (MCI) to Alzheimer's disease (AD) suffer from several problems, including inconsistent PET image input, redundancy of high-dimensional deep features, insufficient utilization of AD phenotypic probabilities, and insufficient stability of individual conversion prediction outputs. To address these issues, this implementation proposes a system for predicting the conversion probability of mild cognitive impairment to Alzheimer's disease. The system's workflow is as follows: Figure 5 As shown, the system obtains the predicted probability of an MCI individual transforming into AD through PET image standardization preprocessing, deep transfer characterization extraction, channel attention enhancement, adaptive global feature aggregation, PCA dimensionality reduction, and AD phenotypic probability fusion.

[0020] The system includes the following units: Data Acquisition and Indexing Module: Acquires PET image data and establishes a traceable data index based on the subject as the basic unit; PET image preprocessing module: Standardizes PET images, calculates SUVR of cerebellar gray matter reference area, extracts brain tissue mask and normalizes intensity to obtain a unified model input image; Transfer Representation Extraction Module: Construct a deep transfer representation model based on residual convolutional transfer backbone, ECA channel attention enhancement module and GeM adaptive global feature aggregation module, complete AD and MCI binary classification training, and extract patient-level deep image features and AD phenotype probabilities; Dimensionality reduction and fusion module: Reads the patient-level migration feature table of MCI, identifies sMCI and pMCI labels, performs PCA dimensionality reduction on patient-level depth image features, and fuses them with AD phenotype probabilities; Probability Prediction Module: Constructs a probability prediction model for MCI to AD conversion, outputs the predicted probability of each MCI individual converting to AD in the future, and generates two sets of predicted probability distribution results: sMCI and pMCI.

[0021] Example 2 This embodiment further defines Implementation 1 and provides a further description of the data acquisition and indexing module.

[0022] PET imaging data were acquired for AD and MCI analysis. Raw images were first organized according to diagnostic category, subject number, visit information, and image record number, and a traceable data index was established with the subject as the basic unit. For PET images included in the study, each case was examined for issues such as image orientation, brain tissue coverage integrity, spatial alignment, abnormal voxels, obvious artifacts, and image truncation to avoid abnormal images affecting subsequent quantitative calculations and model training.

[0023] Through the above data processing, it can be ensured that the imaging data, diagnostic labels, visit information, patient-level migration features and subsequent prediction probabilities of each subject can be matched one by one, providing a foundation for subsequent deep feature extraction, probability fusion and MCI to AD probability prediction.

[0024] Example 3 This embodiment further defines Implementation 1 and provides a further description of the PET image preprocessing module.

[0025] Standardized preprocessing workflow for PET images, such as Figure 1 As shown. Regarding image format and spatial standardization, PET images were uniformly standardized into the NIfTI format to meet the input requirements for subsequent image processing and deep learning modeling. For PET images entering the analysis, their matrix size, voxel resolution, and spatial orientation were further checked to ensure consistency, so that the preprocessed images were uniformly placed in a standard space. This step reduces the impact of differences in scanning orientation, image size, and spatial location between different subjects on model learning.

[0026] To improve the comparability of PET signal intensity among different subjects, this implementation uses cerebellar gray matter as a reference region to calculate whole-brain SUVR images. First, a binary mask of the cerebellar gray matter reference region is generated using a brain region template. Then, the cerebellar gray matter reference region mask is resampled to a standardized PET image space, ensuring consistency with the PET image in matrix size, voxel resolution, and spatial coordinates. Next, using the average uptake value of the cerebellar gray matter region as a reference value, the original PET signal is converted into an SUVR image with relative quantitative significance. Let the PET image in the standard space be... ,in Indicates the voxel location; the cerebellar gray matter reference area is... The average intake value for the reference area is:

[0027] in, This represents the average uptake value of the cerebellar gray matter reference area. This indicates the number of voxels within the reference region. Further, the formula for calculating whole-brain SUVR images is:

[0028] in, Voxel representation The SUVR value at the location is calculated. This method can reduce the impact of differences in overall intake levels among different subjects on subsequent model training. Non-finite values ​​such as NaN and Inf generated during the calculation process are cleaned up to obtain the cleaned whole-brain SUVR image.

[0029] To reduce the interference of the extracerebral background region on the model's feature learning, further brain tissue masking is performed. Specifically, after completing the spatial standardization of PET images and calculating SUVR images, the resulting SUVR images are located in a unified standard space, meaning that images from different subjects have consistent matrix size, voxel resolution, spatial orientation, and spatial coordinates. Subsequently, gray matter, white matter, and cerebrospinal fluid tissue probability maps are generated or resampled within the same standard space, ensuring that the tissue probability maps and SUVR images maintain a correspondence at voxel locations. Therefore, for any voxel location... Both methods can simultaneously obtain the SUVR value and the probability values ​​of gray matter, white matter, and cerebrospinal fluid tissue at that location. Let the probability maps of gray matter, white matter, and cerebrospinal fluid be... , , The threshold is Then the brain tissue mask can be represented as:

[0030] in, Voxel representation Does it belong to a brain tissue region? This indicates an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. This represents the tissue probability threshold. After applying this brain tissue mask to the SUVR image, the resulting SUVR image after brain tissue extraction is obtained:

[0031] in, This represents the SUVR value after brain tissue extraction. This step retains only effective voxels within the brain, removing bone tissue, soft tissue, air, and background areas outside the brain. After brain tissue extraction, the SUVR images within the brain are intensity normalized to ensure consistent image intensity ranges across different subjects. This standardized preprocessing of PET images provides stable and standardized image input for subsequent deep transfer characterization extraction.

[0032] Example 4 This embodiment further defines Implementation 1 and provides a further description of the migration characterization extraction module.

[0033] After completing PET image preprocessing and 2D image input construction, to further automatically learn depth discriminative features related to AD and MCI classification from the images, this implementation constructs a depth transfer representation model based on residual convolutional transfer backbone, such as... Figure 2 As shown, PET images are single-channel grayscale medical images, and their grayscale intensity reflects the distribution of standardized uptake values. Therefore, the input layer of the transfer backbone is adjusted to a single-channel input structure, enabling the model to directly receive two-dimensional PET images.

[0034] The residual convolutional transfer learning backbone learns features from shallow to deep layers through shallow convolutions and multi-level residual feature extraction units. Shallow convolutions primarily extract basic grayscale distribution, local edges, brain tissue contours, and preliminary structural information from PET images. Mid-level residual units further learn local structural patterns and metabolic distribution differences in brain regions. High-level residual units extract abstract semantic features related to differences in tracer uptake, changes in pathological load, or disease states. Residual connections can alleviate gradient vanishing and feature degradation problems as the network deepens, which is beneficial for the stable extraction of deep discriminative features from medical images.

[0035] Considering that key responses in PET images related to the differences between AD and MCI may be concentrated in specific brain regions or specific high-response channels, direct global pooling may weaken local discriminative information. Therefore, an ECA channel attention enhancement module is introduced after the high-level semantic feature output, with the structure as follows: Figure 3 As shown. Let the high-level features output by the residual convolutional transfer backbone be... The channel attention enhancement process can then be represented as:

[0036] in, Indicates the first The original high-level characteristics of each channel, Indicates the first Features enhanced by each channel Indicates the first Attention weights for each channel, Indicates global average pooling. Represents one-dimensional convolution. This represents the Sigmoid activation function. In this way, the model can enhance the key channel responses relevant to AD and MCI classification, while suppressing redundant or weakly correlated channel information.

[0037] After channel attention enhancement, a GeM adaptive global feature aggregation module is introduced to compress high-level feature maps into fixed-length patient-level depth image features. Compared with fixed average pooling, adaptive global feature aggregation can highlight the contribution of locally high-response brain regions to the final feature expression while preserving global image distribution information.

[0038] In one specific embodiment of this implementation, adaptive global feature aggregation can employ GeM pooling. The structure diagram of the GeM adaptive global feature aggregation module is shown below. Figure 4 As shown. For the first The aggregation result of the channels can be expressed as:

[0039] in, Indicates the first The feature values ​​after aggregating the channels Represents the set of spatial locations of feature maps. Indicates the number of spatial locations. Indicates the first The spatial location of each channel Enhanced eigenvalues ​​at the location, This represents adjustable or trainable pooling parameters. When At this time, this aggregation method is equivalent to global average pooling. As the size increases, this aggregation method gradually approaches global max pooling. Therefore, GeM can adaptively balance global distribution information and local high-response information, enabling the model to both retain the overall uptake distribution information of PET images and enhance the contribution of local high-response brain regions in the final feature expression.

[0040] Furthermore, This only represents the feature values ​​of a single channel after GeM aggregation and is not equivalent to the patient-level depth image features of the final input classification head. For images containing... CGeM performs spatial aggregation on each channel's high-level feature map to obtain... C Each channel-level feature value is then used. Subsequently, all channel-level feature values ​​are arranged according to the channel dimension to form a fixed-length depth image feature vector.

[0041] in, This represents the patient-level depth image feature vector formed after GeM aggregation. In this embodiment, if the number of high-level output channels of the residual convolutional transfer backbone is 512, then a 512-dimensional patient-level depth image feature vector can be obtained after GeM aggregation. This feature vector integrates image response information from different high-level channels and is used to characterize the deep image patterns in the subject's PET images related to AD and MCI classification.

[0042] Subsequently, the patient-level depth image feature vectors Input a classification head to perform binary classification prediction for AD and MCI. The classification head may include a flattening layer, a one-dimensional batch normalization layer, a random deactivation layer, a fully connected layer, and a Softmax probability output layer. After mapping by the classification head, the original classification scores for AD and MCI output by the classification head can be expressed as:

[0043] in, This represents the depth image feature vector after flattening, batch normalization, and random deactivation. This represents the weight matrix of the fully connected layer. This represents the bias vector. This represents the original classification scores for the MCI and AD classes output by the classification header. This represents the original classification score of the MCI class. This represents the original classification score for class A. Then, the Softmax function is used to convert the two original classification scores into predicted class probabilities:

[0044] in, This indicates the input of a PET image. This indicates the predicted probability that the subject belongs to category k, which can serve as important input information for subsequent MCI conversion to AD probability prediction.

[0045] The deep transfer representation model is specifically as follows: The model's architecture consists of a "residual transfer backbone - channel attention enhancement - adaptive global aggregation - classification output." First, 2D PET images are input into the residual convolutional transfer backbone, extracting deep image features layer by layer from low-level grayscale distribution to high-level disease semantics. Second, a channel attention enhancement module is added after the high-level semantic features are output, adaptively weighting the importance of different channels to enhance the responses of key channels related to AD and MCI classification. Third, an adaptive global feature aggregation module aggregates the high-level feature maps, highlighting locally high-response brain regions while preserving global distribution information. Finally, the aggregated patient-level deep image features are input into the classification head, outputting the binary classification results for AD and MCI, as well as the AD phenotype probability.

[0046] In one specific embodiment of this implementation, the residual convolutional transfer backbone can adopt a ResNet18 structure, the channel attention enhancement module can adopt an Efficient Channel Attention (ECA) module, and the adaptive global feature aggregation module can adopt a Generalized Mean Pooling (GeM) module. The above specific model structure is merely one embodiment of the present invention and does not constitute a limitation on the scope of protection of the present invention.

[0047] Example 5 This embodiment further defines Implementation 1 and provides a further introduction to the dimensionality reduction and fusion module.

[0048] After extracting deep transfer characteristics, a conversion probability prediction model for MCI individuals was further constructed. MCI individuals can be divided into sMCI and pMCI based on follow-up outcomes. sMCI represents MCI individuals who did not convert to AD during the follow-up period, while pMCI represents MCI individuals who progressed to or converted to AD during the follow-up period. Therefore, the essence of predicting the conversion probability of MCI to AD is to predict the probability that an MCI individual will belong to pMCI in the future, based on baseline PET image characteristics.

[0049] In this implementation, patient-level transfer features extracted from the previous stage deep transfer representation model are used as input. The input features mainly include two parts: first, patient-level deep image features, which reflect deep image patterns in PET images related to AD and MCI classification; and second, AD phenotype probability, which represents the degree to which an individual's image features are close to the AD phenotype.

[0050] Because depth image features have high dimensionality, while the sample size for MCI-converted prediction is relatively limited, directly inputting all high-dimensional features into the prediction model can easily introduce redundant information and increase the risk of overfitting. Therefore, this implementation uses principal component analysis (PCA) to reduce the dimensionality of depth image features, obtaining low-dimensional principal component feature vectors.

[0051] The aforementioned patient-level depth image features and AD phenotype probabilities are generated from the output of the deep transfer representation model. Specifically, preprocessed PET images are input into the trained deep transfer representation model. Patient-level depth image features are extracted through the residual convolutional transfer backbone, the ECA channel attention enhancement module, and the GeM adaptive global feature aggregation module, and the AD phenotype probabilities are output through the classification head. Subsequently, using the subject ID as an index, the patient-level depth image features, AD phenotype probabilities, diagnostic categories, and follow-up outcome information are summarized to form the MCI patient-level transfer feature table.

[0052] When constructing the MCI to AD probability prediction model, patient-level depth imaging features and AD phenotype probabilities of MCI individuals are read from the patient-level migration feature table, and sMCI and pMCI labels are identified based on follow-up outcomes. In this embodiment, leave-one-out cross-validation is used as the training and validation framework for the MCI to AD probability prediction model. For each round of cross-validation, one MCI individual is used as the test sample, and the remaining MCI individuals are used as training samples.

[0053] Simultaneously, the patient-level depth image features and AD phenotype probabilities corresponding to the training and test samples are assigned to the training and test sets respectively using the same partitioning method. In each fold, only the patient-level depth image features from the training set are used for feature standardization and PCA fitting. The patient-level depth image features of the test samples are mapped according to the standardization parameters and principal component space determined by the training set to obtain the low-dimensional principal component feature vector corresponding to the test sample. The AD phenotype probability does not participate in PCA dimensionality reduction but is used as a probabilistic feature reflecting the degree to which an individual's image pattern resembles AD. After obtaining the low-dimensional principal component features, it is horizontally concatenated with the AD phenotype probability to form the fused feature vector corresponding to the training set and the test sample. This processing method can avoid the participation of test sample information in the feature transformation process, thereby reducing the risk of data leakage.

[0054] After obtaining the low-dimensional principal component features, they are horizontally concatenated with the AD phenotypic probabilities to form the final fused feature vector. The fused feature vector can be represented as:

[0055] in, Represents the fused feature vector. to Indicates the front dimension after PCA dimensionality reduction Principal component features, This represents the probability of the AD phenotype. This represents the number of principal components retained in each round of leave-one-out cross-validation. K The number of principal components is determined solely by the PCA model fitted to the patient-level depth image features in the training set for the current round; test samples and AD phenotype probabilities are not involved in determining the number of principal components. Specifically, in the current round, the patient-level depth image features in the training set are standardized and fitted with PCA, and the resulting feature values ​​are arranged in descending order. The preset cumulative variance contribution rate threshold is ,but K It can be represented as:

[0056] in, Indicates the first The eigenvalues ​​corresponding to each principal component This represents the original dimension of the depth image features in the training set, or the number of the largest principal components that can be decomposed. This indicates the preset cumulative variance contribution rate threshold. This represents the number of candidate principal components used in calculating the cumulative variance contribution rate. This method reduces feature dimensionality while preserving key information from depth image features, thereby mitigating redundant features and the risk of overfitting with small samples.

[0057] This fusion strategy utilizes two types of information simultaneously: PCA principal component features preserve the main structural information in the depth image representation, while AD phenotypic probabilities provide a direct estimate of the degree of AD-like image features in an individual. Combining these two approaches helps to characterize the conversion probability of an MCI individual from both the depth feature space and the AD phenotypic probability space.

[0058] Example 6 This embodiment further defines implementation 1 and provides a further description of the probability prediction module.

[0059] After fusing the PCA principal component features with the AD phenotypic probabilities, an MCI-to-AD probability prediction model is constructed. Considering the relatively limited sample size for MCI conversion prediction, retraining a deep convolutional network could easily lead to overfitting. Therefore, this implementation does not retrain the deep convolutional model. Instead, it uses the fused feature vector Z obtained from the previous stage's dimensionality reduction via principal component analysis and fusion with the AD phenotypic probabilities as a fixed input, and constructs a lightweight MCI-to-AD probability prediction model based on this.

[0060] Specifically, this implementation uses leave-one-out cross-validation to train and validate the MCI-to-AD probability prediction model. In each round of cross-validation, the dimensionality reduction and fusion module generates corresponding fusion feature vectors based on the training set and test samples, respectively. These fusion feature vectors are the Z obtained by fusing PCA principal component features with AD phenotypic probabilities. Subsequently, the probability prediction module receives the training set fusion feature vector, the test sample fusion feature vector, and the corresponding sMCI and pMCI labels from the current round, and fits a lightweight probability prediction model with regularization constraints using only the training set fusion feature vector and its labels. After model training is complete, the fusion feature vector of the current test sample is input into the model, and the predicted probability of the MCI individual converting to AD is output. By rotating the test samples in leave-one-out cross-validation, each MCI individual can obtain a conversion probability result generated under independent testing conditions.

[0061] In one specific embodiment of this implementation, the probabilistic prediction model can be a Logistic Regression model with L2 regularization. Let the fused feature vector be... The predicted probability that an MCI individual belongs to the pMCI category can be expressed as:

[0062] in, Represents the intercept term. Represents the model coefficients. Indicates the pMCI category, This represents the predicted probability that an MCI individual will transform into AD in the future. Through the leave-one-out prediction process described above, each MCI individual can obtain an independent predicted transformation probability.

[0063] Furthermore, based on the predicted probabilities of all MCI individuals obtained through leave-one-out cross-validation, the MCI individuals are divided into sMCI and pMCI groups according to their actual transformation states, and two sets of predicted probability distributions are generated, as follows: Figure 6 As shown, if the overall predicted probability of the pMCI group is higher than that of the sMCI group, it indicates that the continuous probability output by the model can reflect the differences between individuals with different conversion states at the individual level. In this way, the model can not only provide pMCI / sMCI category judgments, but also provide continuous and individualized AD conversion prediction probabilities.

[0064] In summary, this system constructs an AD conversion probability prediction workflow for individuals with MCI by employing PET image standardization preprocessing, residual transfer representation extraction, ECA channel attention enhancement, GeM adaptive global feature aggregation, PCA dimensionality reduction, and AD phenotypic probability fusion. This method can improve the representational power of disease-related deep image features while ensuring the consistency of PET image input, reduce the risk of high-dimensional feature redundancy and small-sample overfitting, and output continuous and interpretable MCI-to-AD conversion prediction probabilities.

Claims

1. A system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease, characterized in that, The system includes: Data Acquisition and Indexing Module: Acquires PET image data and establishes a traceable data index based on the subject as the basic unit; PET image preprocessing module: Standardizes PET images, calculates SUVR of cerebellar gray matter reference area, extracts brain tissue mask and normalizes intensity to obtain a unified model input image; Transfer Representation Extraction Module: Construct a deep transfer representation model based on residual convolutional transfer backbone, ECA channel attention enhancement module and GeM adaptive global feature aggregation module. After receiving input images, complete AD and MCI binary classification training, and extract patient-level deep image features and AD phenotype probabilities. Dimensionality reduction and fusion module: Reads the patient-level migration feature table of MCI, identifies sMCI and pMCI labels, performs PCA dimensionality reduction on patient-level depth image features, and fuses them with AD phenotype probabilities; Probability Prediction Module: Constructs a probability prediction model for MCI to AD conversion, outputs the predicted probability of each MCI individual converting to AD in the future, and generates two sets of predicted probability distribution results: sMCI and pMCI.

2. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 1, characterized in that, After acquiring PET image data, it is organized according to diagnostic category, subject number, visit information and image record number, and a traceable data index is established with the subject as the basic unit.

3. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 2, characterized in that, The cerebellar gray matter reference region SUVR calculation is specifically as follows: a binary mask of the cerebellar gray matter reference region is generated using a brain region template; subsequently, the cerebellar gray matter reference region mask is resampled to a normalized PET image space to ensure consistency with the PET image in matrix size, voxel resolution, and spatial coordinates; then, using the average uptake value of the cerebellar gray matter region as a reference value, the original PET signal is converted into an SUVR image with relative quantitative significance. .

4. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 3, characterized in that, Brain tissue mask extraction specifically involves: Tissue probability segmentation was performed on SUVR images in standard space to obtain probability maps of gray matter, white matter, cerebrospinal fluid, and other tissues; a brain tissue mask was constructed using the gray matter, white matter, and cerebrospinal fluid probability maps. The brain tissue mask was applied to the SUVR image. Subsequently, SUVR images of extracted brain tissue were obtained. : .

5. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 4, characterized in that, The data processing flow in the transfer characterization extraction module is as follows: First, the two-dimensional PET image is input into the residual convolution transfer backbone, and deep image features from low-level gray-scale distribution to high-level disease semantics are extracted layer by layer. Secondly, an ECA channel attention enhancement module is added after the high-level semantic feature output to adaptively weight the importance of different channels and enhance the response of key channels related to AD and MCI classification. Thirdly, the high-level feature map is aggregated through the GeM adaptive global feature aggregation module to highlight local high-response brain regions while retaining global distribution information, thereby obtaining patient-level deep image features. Finally, the aggregated patient-level deep image features are input into the classification head to output the AD and MCI binary classification results and AD phenotype probabilities.

6. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 5, characterized in that, The classification head includes a flattening layer, a one-dimensional batch normalization layer, a random deactivation layer, a fully connected layer, and a Softmax probability output layer.

7. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 6, characterized in that, After performing PCA dimensionality reduction on patient-level depth image features and fusing them with AD phenotypic probabilities, the resulting fused feature vector is represented as follows: ; in, Represents the fused feature vector. to Indicates the front dimension after PCA dimensionality reduction Principal component features, This represents the probability of the AD phenotype. This represents the number of principal components retained. Let the eigenvalues ​​obtained from the PCA on the training set be sorted in descending order as follows: The preset cumulative variance contribution rate threshold is ,but It can be represented as: in, Indicates the first The eigenvalues ​​corresponding to each principal component This indicates the number of candidate principal components used in calculating the cumulative variance contribution rate. This represents the original dimension of the patient-level depth image features in the training set, or the maximum number of principal components that can be decomposed.

8. The system for predicting the probability of mild cognitive impairment converting to Alzheimer's disease according to claim 7, characterized in that, In the probability prediction module, the fused feature vector Z, obtained by principal component analysis dimensionality reduction and fusion with the AD phenotypic probability, is used as a fixed input. Based on this, a lightweight MCI to AD probability prediction model is adopted as a Logistic regression model with L2 regularization. Leave-one-out cross-validation is used to train and validate the MCI to AD probability prediction model. Specifically, the module receives the fused feature vector of the training set, the fused feature vector of the test sample, and the corresponding sMCI and pMCI labels of the training set for the current round. The module fits the MCI to AD probability prediction model using only the fused feature vector of the training set and its labels. After the model is trained, the fused feature vector of the current test sample is input into the model, and the predicted probability of the MCI individual being converted to AD is output.

Citation Information

Patent Citations

  • Alzheimer's disease classification and prediction system based on multi-task learning

    CN111488914A

  • Long-term individual AD risk prediction model and prediction method based on multi-scale deep learning

    CN119943388A