Personalized federated learning method for auxiliary diagnosis of Alzheimer's disease

Abstract

The present invention relates to a personalized federated learning method for auxiliary diagnosis of Alzheimer's disease, characterized in that it includes the following steps: composing and preprocessing an image dataset, obtaining a dataset for auxiliary diagnosis, establishing a federated GBDT framework for federated training, and after completing the federated training using the federated GBDT framework, each client weaves the resulting GBDT model into a local deep forest model. In the auxiliary diagnosis method for Alzheimer's disease, the present invention uses federated learning and deep forests to achieve accurate AD classification tasks, which is used to address privacy challenges in the healthcare field without compromising privacy security. The present invention overcomes the problem of information isolation by calculating high-order information distributed among different clients without exchanging their private data, and the present invention enables distributed clients to collaboratively train shared models without sharing their training data.

Description

Technical Field

[0001] The present invention relates to a personalized federated learning method for auxiliary diagnosis of Alzheimer's disease, and belongs to the field of computer image processing. Background Art

[0002] Alzheimer's disease is a persistent neurological disorder that often begins in old age or pre-elderly age. The disease progresses slowly, worsening over time with an insidious onset and rapid progression, accompanied by structural and functional changes in the brain. Symptoms typically include memory loss, impaired visual-spatial abilities, abstract thinking and calculation, personality and behavioral changes, and irreversible progressive worsening until complete disability. Mild cognitive impairment (MCI) is the intermediate stage between normal function and Alzheimer's disease. This stage is when Alzheimer's disease symptoms begin to manifest and treatment with medications and other interventions are most appropriate. Over half of patients at this stage will develop dementia within five years. However, minor symptoms such as memory loss at this stage may be mistaken for normal aging and ignored, leading to missed treatment opportunities. Once MCI progresses to full-blown Alzheimer's disease, there are no effective medications or treatments to delay or cure it.

[0003] Since there is currently no drug that can effectively treat Alzheimer's disease, most treatments for the disease are based on the different stages of the patient's disease course, and the corresponding drugs are used to delay the progression of the disease to a certain extent. Therefore, it is particularly important to be able to accurately judge the stage of the patient's disease course. Traditional screening methods for mild cognitive impairment include neuroimaging examinations, neuropsychological assessments, biomarker tests, etc. However, these methods all require the intervention of professional doctors, and the time and economic costs are high. Some examination methods are even invasive to the human body, which objectively leads to the failure of most Alzheimer's patients to be diagnosed early. Therefore, the present invention uses voice features to judge the patient's cognitive status, effectively lowering the threshold for early screening of mild cognitive impairment, and has scientific value, social value and economic value.

[0004] To date, neuroimaging is the best non-invasive technique for detecting abnormalities in the human brain. Deep learning has revolutionized image processing, solving challenging problems such as image colorization, classification, segmentation, and detection. For example, a deep neural network (DNN) is a multi-layered stack that allows models to become more effective at learning complex features and performing more computationally intensive tasks, surpassing the classical machine learning paradigm for machine perception tasks involving unstructured data. While deep neural networks have achieved impressive performance, existing deep learning methods for training deep architectures from scratch have limitations, including the requirement for large amounts of labeled data, high computational complexity, and a lack of theoretical explanations. For these reasons, researchers are searching for an alternative paradigm, and deep ensemble learning, such as GrowNet, S-DNN, DSN, and Deep Forest, has gained attention. It combines the advantages of DNNs and ensemble learning, resulting in models with improved generalization performance. However, another challenge in the healthcare field is the well-known issue of privacy. Traditional deep learning generally consists of two stages: collecting data from different participants, pre-processing the data, and then inputting the data into the overall model for training. This inevitably poses the risk of privacy leakage.

[0005] Therefore, existing technologies require a new image processing algorithm to solve the problem of privacy leakage, and a personalized federated learning method for auxiliary diagnosis of Alzheimer's disease to improve the accuracy of early screening for Alzheimer's disease. Summary of the Invention

[0006] The purpose of this invention is to provide a new image processing algorithm to solve the privacy leakage problem faced by existing image processing technologies, and to provide a personalized federated learning method for auxiliary diagnosis of Alzheimer's disease to improve the accuracy of early screening for Alzheimer's disease.

[0007] To achieve the above objectives, the technical solution of the present invention is to provide a personalized federated learning method for auxiliary diagnosis of Alzheimer's disease, which is characterized by comprising the following steps:

[0008] Step 1: Obtain the subject's 1.5T T1-weighted structural magnetic resonance (sMRI) image data to form an image dataset;

[0009] Step 2: Preprocess the image dataset;

[0010] Step 3: Generate a dataset for auxiliary diagnosis based on the preprocessed image dataset, including the following steps:

[0011] The image dataset processed by the preprocessing unit is fed into the 3DUnet-CBAM model for training. After the hippocampus is extracted by the 3DUnet-CBAM model, the voxels of the image are labeled in a neuroanatomical manner using IBASPM, and the volume is further automatically calculated to ultimately generate a dataset for auxiliary diagnosis.

[0012] Step 4: Establish a federated GBDT framework for federated training:

[0013] The federated GBDT framework enables distributed clients to collaboratively train a global model without sharing their local datasets generated in step 3. The federated GBDT framework has two participants, a central server and a client. Each client protects its security by preventing any other party from accessing its local private data. The client's local model selects a GBDT model for local training. In each training iteration, the central server broadcasts the GBDT model to each client, and then each client trains its own GBDT model using its local dataset and uploads the gradient to the central server. Finally, the server aggregates the gradient to train the global model.

[0014] Step 5. After completing the federated training using the federated GBDT framework, each client weaves the resulting GBDT model into the local deep forest model and uses GBDT instead of the decision tree to build a personalized model. The personalized model takes the client's private data as input, and then uses multi-granularity scanning to extract and process the original features of the input to generate a feature vector, which is further sent to the cascade forest to complete the classification task.

[0015] Preferably, in step 1, the data obtained are all from the Alzheimer's Disease Neuroimaging Initiative ADNI dataset, and the selected data are 1.5T T1-weighted structural magnetic resonance sMRI image data acquired using fast gradient echo imaging technology, and the data are in NIFTI format.

[0016] Preferably, all image data in the image dataset are divided into three categories, namely, from subjects with normal cognition, patients with Alzheimer's disease, and patients with mild cognitive impairment.

[0017] Preferably, in step 2, the preprocessing of the image dataset includes skull stripping, bias field correction, standardization, and cropping to a uniform size.

[0018] Preferably, in the constructed personalized model, each cascade level contains four GBDTs, each level outputs a class distribution vector, the four vectors at the same level are connected with the output of the multi-granularity scanner, and the result becomes the input vector of the next level. The final classification result is the maximum value of the average class vector output from the last level.

[0019] In an assisted diagnosis method for Alzheimer's disease, this paper uses federated learning and deep forests to achieve accurate AD classification, addressing privacy challenges in healthcare without compromising privacy security. This paper overcomes the issue of information isolation by computing high-level information distributed across different clients without exchanging their private data. Furthermore, this paper enables distributed clients to collaboratively train a shared model without sharing their training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is an overview of the personalized federated deep forest framework for auxiliary diagnosis of Alzheimer's disease described in the present invention;

[0021] Figure 2 Schematic diagram of the training process of the federated GBDT framework described in the present invention;

[0022] Figure 3 It is a schematic diagram of the personalized model structure of the present invention;

[0023] Figure 4 This is a flowchart of the data set preprocessing described in the present invention. DETAILED DESCRIPTION

[0024] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0025] The present invention discloses an auxiliary diagnosis method for Alzheimer's disease based on personalized federated learning and deep forest. The model architecture used in this method is as follows: Figure 1 As shown ( Figure 1 This paper presents a framework for achieving accurate AD classification without compromising privacy through federated learning and deep forests. This framework overcomes the information isolation issue by computing high-level information distributed across clients without exchanging their private data. The framework includes the following steps:

[0026] Step 1: Obtain an image dataset of the subject. The data used in this paper are all from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The selected data are 1.5T T1-weighted structural magnetic resonance imaging (sMRI) images acquired using rapid gradient echo imaging (MPRAGE) technology, and the data are in NIFTI format. All image data in the image dataset can be divided into three categories: from subjects with normal cognition, patients with Alzheimer's disease, and patients with mild cognitive impairment.

[0027] Step 2: Preprocess the image dataset of the subjects collected in step 1. This process includes four steps: skull stripping, bias field correction, standardization, and cropping to a uniform size. The sMRI images used as training data and test data in the image dataset are preprocessed using the above four steps. The overall process is as follows: Figure 4 shown.

[0028] In this embodiment, step 2 specifically includes the following steps:

[0029] Step 201: The subject's skull is peeled off. This is because the skull is relatively far away from the hippocampus and contains irrelevant voxel information of the hippocampus.

[0030] Step 202: Use N4BiasFieldCorrection in the SimpleITK software package to correct the bias field. This is to address the brightness differences that may occur during the scanning process due to factors such as the scanning position and the scanning itself. If not corrected, the magnetic field distribution of the MRI machine will be uneven.

[0031] Step 203: Register the images processed in steps 101 and 102 to the standard brain template space MNI to unify the coordinate space of all images.

[0032] Step 204: Crop the sMRI images into a uniform size of 196×271×181.

[0033] Step 3: After preprocessing, the image dataset is fed into the 3DUnet-CBAM model for training. This model can also focus on the hippocampus's boundary region, further improving hippocampal segmentation accuracy. After extracting the hippocampus, IBASPM is used to neuroanatomically label the image voxels and automatically calculate the volume, ultimately generating a dataset for auxiliary diagnosis. This dataset serves as the local data for each client device in steps 4 and 5 for model training and prediction.

[0034] In the auxiliary diagnosis method for Alzheimer's disease, the present invention uses federated learning as the basic framework to address privacy challenges in the healthcare field. It enables distributed clients to collaboratively train shared models without sharing their training data. Specifically, the model parameters are calculated locally by each client device and exchanged with a central server, which aggregates local models for a global view. It is worth noting that the present invention uses a deep forest model to provide personalized services, making the model more adaptable to each client. The above method is specifically implemented by the following steps:

[0035] Step 4: Build a federated GBDT framework that enables distributed clients to collaboratively train a global model without sharing their local training data. This federated GBDT framework has two participants: a central server and clients. Each client protects its local private data by preventing any other party from accessing it. The client's local model selects a GBDT model for local training. In each training iteration, the central server broadcasts the GBDT model to each client. Each client then trains its own GBDT model using its local data and uploads the gradients to the central server. Finally, the server aggregates the gradients to train the global model.

[0036] The client of the present invention selects the GBDT model for local training. Figure 2 As shown in the figure, for each iteration of federated training, each client first initializes its local GBDT model and then feeds private data into the GBDT model. In the GBDT model, each decision tree is constructed by fitting the negative gradient of the previous decision tree. Therefore, the complete GBDT model is composed of a chain of decision trees. When the client completes local training, it calculates the first-order and second-order gradient statistics of the loss function. These two gradients are then uploaded to the server for aggregation. Finally, the central server collects feedback from all clients and updates the global GBDT model. The server then broadcasts the aggregated gradients to each client. Therefore, in the next iteration, these gradients become the building blocks of the new local GBDT model. This process is repeated until the global model converges and satisfactory federated learning results are achieved.

[0037] Gradient boosted decision tree (GBDT) is an integrated model for training decision tree sequences. Formally, given a loss function and a dataset with n samples and d features The model uses K addition functions to predict the output It can be obtained by the following formula:

[0038]

[0039] In it: is the space of regression trees, ω q(x) represents the score of leaf node q, f(x) represents one of the trees, q represents the structure of each tree, and maps an example to the corresponding leaf index, q: T is the number of leaves in the tree, ω represents the leaf weight, ω i represents the score on the i-th leaf; X i represents the i-th sample; Represents the prediction result after iteration; f k(X i ) represents the k-th tree model. For a given sample, we will use the decision rule in the tree (given by q) to classify it into leaves and calculate the final prediction result by summing up the scores in the corresponding leaves (given by ω).

[0040] In order to learn the set of functions used in the model, we need to minimize the following regularization objective As shown in the following formula:

[0041]

[0042] Where: l(·) is a fine-tunable convex loss function that measures the predicted value and target value Y i the differences between; is a regularization term that penalizes model complexity, where γ and λ are hyperparameters. The model is trained in an additive manner, and GBDT minimizes the following objective function at the tth iteration

[0043]

[0044] Where: and are the first-order and second-order gradient statistics of the loss function, represents the first-order derivative of the loss function after the t-1th iteration; f t (x i ) represents the t-th tree model.

[0045] Step 5: Build a personalized model to improve the privacy-accuracy trade-off and balance the interests of different parties. Figure 3As shown, the personalized model takes the client's private data as input, then uses multi-granularity scanning to extract and process the input raw features to generate feature vectors, which are further sent to the cascade forest to complete the classification task. The original cascade forest in gcForest is defined by a multi-level integration of decision trees. In theory, it can be replaced by any other classifier that can output a class distribution vector. Therefore, the present invention uses GBDT instead of decision trees to build a personalized model. Detailed description: Each cascade level contains four GBDTs, and each level outputs a class distribution vector. The four vectors at the same level are then connected with the output of the multi-granularity scanner, and the result becomes the input vector of the next level. The final classification result is the maximum value of the average class vector output from the last level. After the federated training in step 4 is completed, each client weaves the resulting model GBDT into the local deep forest model, namely gcForest, in order to train the personalized model and further provide more accurate AD diagnosis. At this point, the present invention has successfully constructed a personalized federated learning model for auxiliary diagnosis of Alzheimer's disease.

Claims

1. A personalized federated learning method for auxiliary diagnosis of Alzheimer's disease, characterized by: The following steps are involved: Step 1: Obtain the subject's 1.5T T1-weighted structural magnetic resonance (sMRI) image data to form an image dataset; Step 2: Preprocess the image dataset; Step 3: Generate a dataset for auxiliary diagnosis based on the preprocessed image dataset, including the following steps: The image dataset processed by the preprocessing unit is fed into the 3DUnet-CBAM model for training. After the hippocampus is extracted by the 3DUnet-CBAM model, the voxels of the image are labeled in a neuroanatomical manner using IBASPM, and the volume is further automatically calculated to ultimately generate a dataset for auxiliary diagnosis. Step 4: Establish a federated GBDT framework for federated training: The federated GBDT framework enables distributed clients to collaboratively train a global model without sharing their local datasets generated in step 3. The federated GBDT framework has two participants, a central server and a client. Each client protects its security by preventing any other party from accessing its local private data. The client's local model selects a GBDT model for local training. In each training iteration, the central server broadcasts the GBDT model to each client, and then each client trains its own GBDT model using its local dataset and uploads the gradient to the central server. Finally, the server aggregates the gradient to train the global model. Step 5. After completing the federated training using the federated GBDT framework, each client weaves the resulting GBDT model into the local deep forest model and uses GBDT instead of the decision tree to build a personalized model. The personalized model takes the client's private data as input, and then uses multi-granularity scanning to extract and process the original features of the input to generate a feature vector, which is further sent to the cascade forest to complete the classification task.

2. A personalized federated learning method for auxiliary diagnosis of Alzheimer's disease according to claim 1, characterized in that: In step 1, all data obtained are from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The selected data are 1.5T T1-weighted structural magnetic resonance imaging (sMRI) image data acquired using fast gradient echo imaging technology, and the data are in NIFTI format.

3. The personalized federated learning method for auxiliary diagnosis of Alzheimer's disease according to claim 1, characterized in that: All image data in the image dataset are divided into three categories, namely, from subjects with normal cognition, patients with Alzheimer's disease, and patients with mild cognitive impairment.

4. The personalized federated learning method for auxiliary diagnosis of Alzheimer's disease according to claim 1, characterized in that: In step 2, the image dataset is preprocessed including skull stripping, bias field correction, standardization, and cropping to a uniform size.

5. The personalized federated learning method for auxiliary diagnosis of Alzheimer's disease according to claim 1, characterized in that: In the constructed personalized model, each cascade level contains four GBDTs, each level outputs a class distribution vector, and the four vectors at the same level are connected with the output of the multi-granularity scanner. The result becomes the input vector of the next level, and the final classification result is the maximum value of the average class vector output from the last level.

Images