CN115457339B

CN115457339B - AD prediction method, system and device based on deep ensemble learning

Info

Publication number: CN115457339B
Application number: CN202211416845.XA
Authority: CN
Inventors: 张弓; 苏进; 李学俊; 王华彬
Original assignee: China Canada Institute Of Health Engineering Hefei Co ltd
Current assignee: China Canada Institute Of Health Engineering Hefei Co ltd
Priority date: 2022-11-11
Filing date: 2022-11-11
Publication date: 2023-03-24
Anticipated expiration: 2042-11-11
Also published as: CN115457339A

Abstract

The invention discloses an AD prediction method, system and device based on deep ensemble learning, and belongs to the field of medical image processing. In order to solve the problems of low diagnosis efficiency, high misdiagnosis rate and the like in the prior art that the diagnosis of AD is carried out by visually observing the metabolic activity of the brain in a PET image of a patient through computer-aided marking software, the invention acquires the PET image from an ADNI database, slices the acquired PET image, carries out data amplification on the sliced PET image, inputs the amplified PET slice image into a ResNet-34 network and an EfficientNet-b1 network respectively for feature learning, classifies the learned features through MLP, fuses classification results and then carries out prediction output on the fused results, thereby realizing the enhancement of the characteristic learning capability of a prediction model on PET and improving the accuracy of early Alzheimer disease prediction.

Description

AD prediction method, system and device based on deep ensemble learning

Technical Field

The invention relates to the field of medical image processing, in particular to an AD prediction method, system and device based on deep ensemble learning.

Background

Alzheimer Disease (AD) is a progressive degenerative disease of the nervous system, and its clinical manifestations include memory impairment, aphasia, and agnosia. AD has become a global epidemic with over 5000 million people worldwide suffering from dementia, and the socio-economic burden placed on the world each year is conservatively estimated to exceed $ 1 trillion. The etiology of the AD is unknown so far, no specific medicine can cure or effectively reverse the disease process, only the development of the disease can be slowed down by using medicines and the like, and corresponding measures are taken as soon as possible to improve the nursing conditions in a targeted manner, so that the rehabilitation of the old patients or the delayed clinical attack time is facilitated.

Currently, a common method for diagnosing alzheimer's disease by clinicians is to use computer-aided labeling software to visually observe the metabolic activity of the brain in PET images of the brain of a patient. However, the method has the characteristics of low diagnosis efficiency, high misdiagnosis rate and complex diagnosis process, and is not beneficial to early discovery and later treatment of diseases.

With the rapid development of neuroimaging technology, neuroimaging diagnosis becomes an intuitive and reliable method for diagnosing alzheimer's disease. In neuroimaging diagnosis, positron Emission Tomography (PET) is an important neuroimaging technique for detecting alzheimer disease, and is an image reflecting pathological genes, molecules, metabolism and functional states of the brain. It can reflect the glucose consumption of the active area of the glutamatergic synapse and the astrocyte of the brain, and is a 'barometer' for the activity metabolism and the function of the neuron of the brain. Studies have shown that PET shows features of AD neuropathy earlier than MRI in individuals with mild cognitive impairment. Therefore, PET-based neuroimaging diagnosis is an important tool for early screening of AD.

With the development of computer-aided technology in recent years, more and more applications of deep learning and digital image processing technology in the medical field, especially in the aspect of Alzheimer's disease prediction, have achieved huge results. The brain PET neural image can effectively predict early Alzheimer's disease, but the early PET of a subject has the characteristic of unobvious lesion features, and the PET has the problems of low signal-to-noise ratio, small data volume and the like, so that the difficulty in distinguishing early lesion samples from normal samples is high. Therefore, in order to enhance the learning ability of the prediction model for characterization of PET and improve the accuracy of early stage alzheimer's disease prediction, it is necessary to develop a prediction method that can accurately diagnose AD before clinical symptoms of AD appear.

Disclosure of Invention

1. Technical problem to be solved

The invention provides an AD prediction method, system and device based on deep integrated learning, aiming at the problems that in the prior art, the metabolic activity of the brain in a PET image of the brain of a patient is observed by naked eyes through computer-aided marking software to diagnose AD, the diagnosis efficiency is low, the misdiagnosis rate is high, the diagnosis process is complex, and the early detection and later treatment of diseases are not facilitated.

2. Technical scheme

The purpose of the invention is realized by the following technical scheme.

A method for predicting AD based on deep ensemble learning comprises the steps of,

acquiring a PET image from an ADNI database, and slicing the acquired PET image;

carrying out data amplification on the sliced PET image;

respectively inputting the amplified PET slice images into a ResNet-34 network and an EfficientNet-b1 network for feature learning;

classifying the characteristics learned in the ResNet-34 network and the EfficientNet-b1 network through MLP;

and fusing the classification results, and predicting and outputting the fused results.

Further, the ResNet-34 Network is a Network with a residual structure ratio of (3,4,6,3), and a high way Network in the ResNet-34 Network structure is fused with low-order feature information in the amplified PET slice image.

Further, obtaining a scaling weight of the EfficientNet-b1 network, and performing transfer learning on the amplified PET slice image to obtain a new weight value of the EfficientNet-b1 network; the Efficientnet-b1 network automatically searches for the composite coefficient, and optimizes and predicts the network through three dimensions of depth, width and resolution of the network.

Further, the EfficientNet-b1 network structure comprises a backbone network, global average pooling, global maximum pooling, feature splicing and a classifier.

Further, after the amplified PET slice image is input into a backbone network, the EfficientNet-b1 network learns the feature information of the amplified PET slice image, generates a corresponding feature matrix, performs global average pooling and global maximum pooling to obtain two groups of one-dimensional feature information, performs information feature splicing, classifies the spliced feature information by a classifier, performs back propagation, and performs multiple iterations to form an optimal weight.

Further, model training is carried out on a plurality of processed PET slice images by adopting 10k-fold cross validation, one group of the processed PET slice images is used as a validation set during training, the rest images are used as training sets, the optimizer uses SGD, the learning rate attenuation strategy is StepLR, and the loss function adopts a cross entropy loss function.

Further, the cross entropy loss function L _CE The calculation formula of (2) is as follows:

wherein N is the number of samples, M is the number of categories, y _i Is the class of sample i, with the positive class being 1 and the negative class being 0,L _i For each training sample, the loss value, p, in the cross entropy operation with the corresponding class _i Predict probability of being a positive class, y, for sample i _ic As a function of the sign, p _ic The probability is predicted for sample i.

Further, a 10k-fold cross validation model is selected in the prediction training process; and four amplification modes of random scale transformation, random small-angle rotation, random horizontal mirror image and random vertical mirror image are used.

An AD prediction system based on deep ensemble learning is characterized in that a PET image is subjected to feature learning by adopting an AD prediction method based on deep ensemble learning; the AD prediction system includes:

the input module is used for inputting the PET image, carrying out slicing processing and then carrying out amplification and increasing samples of a training set;

the training module is used for respectively inputting the amplified PET slice images into a ResNet-34 network and an EfficientNet-b1 network for feature learning, and classifying the learned features through MLP;

and the output module is used for fusing the classification results and predicting and outputting the fused results.

An AD prediction device based on deep ensemble learning comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the processor realizes the steps of the AD prediction method based on deep ensemble learning when executing the program.

3. Advantageous effects

Compared with the prior art, the invention has the advantages that:

according to the AD prediction method, the system and the device based on the deep ensemble learning, the ResNet-34 Network can enable a deeper Network to train a good effect, the high-order characteristic information can be guaranteed not to be lost in the structure, the EfficientNet-b1 Network can expand the Network receptive field range, a model is strengthened to learn fine characteristics among PET image characteristics, the learned characteristic information is classified through MLP, classification results are fused, and the fused results are predicted and output. The characteristics of the PET image can be well learned by utilizing the common operation of the ResNet-34 network and the EfficientNet-b1 network, information complementation is formed, the AD can be accurately diagnosed before the clinical symptoms of the AD appear, the characteristic learning capability of a prediction model on the PET is enhanced, and the accuracy of predicting early Alzheimer's disease is improved.

Drawings

FIG. 1 is a network framework diagram of an AD prediction method based on deep ensemble learning according to the present invention;

FIG. 2 is a network training and testing flow chart of the deep ensemble learning-based AD prediction method of the present invention;

FIG. 3 is a schematic diagram of a ResNet-34 network structure in the deep ensemble learning-based AD prediction method of the present invention;

FIG. 4 is a schematic diagram of an EfficientNet-b1 network structure in the deep ensemble learning-based AD prediction method of the present invention.

The numbering in the figures illustrates: 1. random scale transformation; 2. rotating at a small angle randomly; 3. randomly horizontally mirroring; 4. random vertical mirroring.

Detailed Description

The invention is described in detail below with reference to the drawings and specific examples.

Examples

As shown in fig. 1, in the AD prediction method, system and apparatus based on deep ensemble learning provided in this embodiment, a PET image is obtained from an ADNI database, the obtained PET image is sliced, data amplification is performed on the sliced PET image, the amplified PET slice image is respectively input into a ResNet-34 network and an EfficientNet-b1 network for feature learning, features learned in the ResNet-34 network and the EfficientNet-b1 network are classified by MLP, classification results are fused, and then the fused result is predicted and output.

Specifically, in the present embodiment, as shown in fig. 1 and fig. 2, a network architecture diagram of an AD prediction method based on deep ensemble learning and a network training and testing flow diagram of the AD prediction method provided in the present embodiment are provided. The test flow is as follows:

(1) And (4) preprocessing data. Because of The specificity of AD Disease studies, this example study is based on The data set of Alzheimer's Disease Neuroactive Initiative (ADNI), which is publicly available on a website (www.loni.ucla.edu/ADNI). ADNI was initiated by non-profit organizations such as the national institute of aging problems (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB) and the like in 2003, and is dedicated to the common solution of AD diseases caused by the increasing aging population by medical experts all over the world. The neuroimaging brain PET images obtained from the ADNI database were taken from 5 size images of 128 × 128, 168 × 168, 256 × 256, 336 × 336 and a small number of 400 × 400, respectively, from devices manufactured by several medical companies, such as philips medical system and siemens. It can be seen that 256 × 256 and 336 × 336 are obtained by doubling 128 × 128 and 168 × 168, respectively, and 168 × 168 is obtained by 128 × 128 edge-extended black edges. Thus, in selecting the training data size, the compromise picks the 224 x 224 scale as the standard input image size. The advantage of selecting the size is that the PET image information is not excessively scaled to cause excessive morphological change of the image, and the original characteristic information of the PET slice image is furthest reserved.

(2) Data amplification and enhancement. And amplifying and enhancing the obtained sliced PET image through four modes of random scale transformation 1, random small-angle rotation 2, random horizontal mirror image 3 and random vertical mirror image 4. The amplification of the training data can increase samples of a training set, can effectively relieve the overfitting condition of the model, and can bring stronger generalization capability to the model. The random scale transformation 1 is obtained by calling a random scale transformation function mode, and is mainly used for deep learning of multi-scale PET images so as to strengthen network learning to obtain PET characteristic information of different scales; the random small-angle rotation 2 is obtained by calling a random small-angle rotation function mode and is mainly used for deep learning of multi-angle PET images so as to strengthen network learning to obtain PET characteristic information of different angles; the random horizontal mirror image 3 is obtained by calling a random horizontal and vertical translation function mode, is mainly used for deep learning of PET images in different horizontal and vertical directions, and is used for strengthening network learning to obtain PET characteristic information in different horizontal and vertical directions; the random vertical mirror image 4 is obtained by calling a random horizontal vertical mirror image function mode, and is mainly used for deep learning of batch copy of different horizontal and vertical PET images so as to strengthen network learning to obtain more PET characteristic information in different horizontal and vertical directions.

(3) And (5) network training. After memory, running time and robustness are considered, a residual network 34 type (ResNet-34 network) and a scaling network b1 type (EfficientNet-b 1 network) are selected as the prediction network structure based on deep ensemble learning in the embodiment. As shown in FIG. 3, the ResNet-34 network is a residual structure ratio (3,4,6,3) network. Specifically, there are 4 layer structures in the ResNet-34 network, the number of which is 3,4,6 and 3 in turn, and in this embodiment, the network only uses the ratio of the residual structure, the deeper the network is, the more information is acquired, and the richer the features are. However, as the network deepens, the optimization effect is worse, and the accuracy of the test data and the training data is reduced. The problem of gradient explosion and gradient disappearance can be caused by deepening the Network, while the ResNet-34 Network can enable a deeper Network to train a good effect, and a jump connection Network (high way Network) in the structure can ensure that high-order characteristic information is not lost, can be effectively fused with low-order characteristic information, and improves the characteristic learning capability of the Network. In this embodiment, the high-order feature information refers to a highly abstract feature obtained through multiple training in the ResNet-34 network training, and the low-order feature information refers to feature information obtained in the training start, such as characterization information of color change, shape size, and the like of an image.

The scaling network (EfficientNet network) is used for scaling according to a certain rule by combining the depth (depth), the width (width) and the resolution (resolution) of the network, so that the effect of the network is improved, the effect of the EfficientNet is obviously improved when the network is enlarged, and the upper limit of the precision is further improved. It is worth noting that in the prior art, the classification performance of the Efficientnet network is good, but the Efficientnet network depends on the pre-training weight, and the excellent scaling performance of the Efficientnet network cannot be shown on the premise that the pre-training weight is not available. Specifically, in this embodiment, a scaling weight of an open-type EfficientNet network is obtained, and a new weight value is obtained by performing transfer learning on an existing brain PET slice image, so that excellent classification performance of the EfficientNet network is finally shown. In this embodiment, the scaling of the Efficientnet-b1 network is width _ coeffient =1.0, depth _coeffient =1.1, resolution =240, and drop _rate =0.2, and the prediction network is optimized from three dimensions of depth, width, and resolution of the network by automatically searching for a composite coefficient. An optimal set of parameters or composite coefficients can be obtained based on a neural structure search technique to achieve a tradeoff between accuracy and computational complexity. The EfficientNet-b1 network structure comprises a backbone network, global average pooling, global maximum pooling, feature splicing and a classifier. As shown in fig. 4, in the network training process, after the sliced PET image is input to the backbone network, the network learns the feature information of the sliced PET image, generates a corresponding feature matrix, performs global average pooling and global maximum pooling, obtains two sets of one-dimensional feature information, performs feature concatenation, classifies the feature information after feature concatenation by a multilayer perceptron (MLP), performs back propagation, performs multiple iterations to form an optimal weight, and predicts that the network completes training.

Therefore, the ResNet-34 Network can enable a deeper Network to train a good effect, the high-order characteristic information can be guaranteed not to be lost in the Network, the EfficientNet-b1 Network can expand the Network sensing visual field range, the model can be strengthened to learn the fine characteristics among the PET image characteristics, the two Network structures work in a combined mode, the characteristics of the PET image can be well learned, information complementation is formed, and the prediction accuracy is improved.

(4) MLP perceptual classification. The MLP comprises three neural networks, namely an input layer, a hidden layer and an output layer, so that the problem of nonlinearity which cannot be solved by a single-layer perceptron can be effectively solved, and the classification performance and the prediction precision of the prediction network are improved. In the embodiment, the sliced PET image is input into a ResNet-34 network, and the network obtains a characteristic matrix y after convolution, pooling and residual error; inputting the sliced PET image into an EfficientNet-b1 network, obtaining a feature matrix y after the network is subjected to convolution and pooling, finally inputting the feature matrix y into an MLP to obtain a low-dimensional characterization vector, inputting the feature matrix y into the MLP to obtain a low-dimensional characterization vector, adding the two low-dimensional characterization vectors element by element to obtain a new low-dimensional characterization vector, namely a new feature matrix, then performing MLP training on the new feature matrix, obtaining more appropriate parameters and threshold values through continuous training iteration, and finally outputting the corresponding low-dimensional characterization vector by an output layer, namely representing the characterization information of a training set which is learned by the network.

(5) And (5) training a strategy. Model training is performed by using 10k-fold cross validation, in this embodiment, 8000 acquired PET images are divided into 10 groups, each group includes 800 images, one group is used as a validation set during training, and the remaining images are used as a training set. The optimizer uses SGD, the learning rate decay strategy is StepLR, and the loss function uses cross entropy loss. The purpose of using 10k-fold cross validation is to make the predictive network fully utilize the data resources to train to optimal weights within a limited data set.The ten folds divide the data set evenly so that the amount of data per fold is the same. In the network training stage, cross entropy loss continuously iterates and optimizes the prediction weight, and the method is an important link for predicting the information of the focus from network learning. Cross entropy loss function L _CE The calculation formula of (a) is as follows:

wherein N is the number of samples, M is the number of categories, y _i Is the class of sample i, with the positive class being 1 and the negative class being 0,L _i For each training sample, the loss value, p, in the cross entropy operation with the corresponding class _i Predict probability of being positive class, y, for sample i _ic As a function of the sign, p _ic The probability is predicted for sample i.

(6) And (5) enhancing the test. In the testing process, a large number of effective data samples are obtained by using four enhancing modes of random scale transformation 1, random small-angle rotation 2, random horizontal mirror image 3 and random vertical mirror image 4, and various data information is provided for predicting network accuracy so as to improve the robustness and prediction accuracy of a prediction model. And simultaneously, selecting a model with the best effect of 10k-fold cross validation training for testing in the test, and selecting the prediction type with the highest confidence coefficient as the final prediction result of the image.

(7) And (6) testing results. In the training process, the model training and testing operation environment is as follows: windows system, 6GB video memory, python3.8, using Pythrch deep learning frame, optimizer Adam, training learning rate 0.0005, maximum round epoch 100, batch size 64, adopting ten-fold cross training, and using mixed precision training accelerated library training, accelerating library parameter setting accumulation _ steps =4, opt_level = "O1". During the training process, when the loss value is smaller than the optimal value, the model is saved until the round is finished.

As shown in table 1, a comparison table between the AD prediction method based on deep ensemble learning and the AD prediction method advanced in recent years is provided for this embodiment. Where NC indicates normal cognition, SEN (Sensitivity) indicates Sensitivity, SPE (Specificity) indicates Specificity, and ACC (Accuracy) indicates Accuracy. For sensitivity SEN, i.e., true positive rate, it is the proportion of truly diseased samples in all samples predicted to be diseased. For SPE, the rate is true negative, indicating the proportion of truly disease-free samples in all samples predicted to be disease-free. Since in sensitivity SEN, if a disease is predicted but not, the rate of samples predicted to be disease-free but actually diseased is not as small as possible. This illustrates that in a practical diagnostic expert system, the more important of the two indicators is the SPE specificity.

The ROC Curve is also called a receiver operating characteristic Curve (receiver operating characteristic Curve), and is a comprehensive index reflecting continuous variables of the sensitivity SEN and the specificity SPE, the ROC Curve can reflect the classification effect of the classifier to a certain extent, in this embodiment, an AUC (Area Under Curve) is defined, and the classification capability expressed by the ROC Curve can be reflected by an AUC value. The training process verifies that the set Loss =0.043, auc =98.125%, and the test set F1=92.7%. Wherein Loss =0.043 is that the network tends to be stable around 0.043 after 200 iterations in the training set, indicating that the network has fitted the training set data. The test set FI represents the value of the test index F1 in the network test set, is a weighted average of the model accuracy rate and the recall rate, is the comprehensive evaluation of the classification recall rate and accuracy rate, and is an important index for evaluating the classification performance of the model.

Table 1 comparison table of AD prediction method provided in this embodiment and AD prediction method advanced in recent years

As can be seen from table 1, compared with the AD prediction method advanced in recent years, the accuracy ACC of the AD prediction method provided in this embodiment is 9.24% higher than that of the latest literature, 1.57% higher than that of the literature with the best experimental effect, and all other parameters are higher than that of the current literature.

In addition, according to the AD prediction system based on the deep ensemble learning provided by the embodiment, the AD prediction method based on the deep ensemble learning is adopted to perform the feature learning of the PET image; the AD prediction system includes:

the input module is used for inputting the PET image to carry out slicing processing, and then amplifying the PET image through four modes of random scale transformation 1, random small-angle rotation 2, random horizontal mirror image 3 and random vertical mirror image 4, and is used for increasing samples of a training set;

The embodiment also provides an AD prediction apparatus based on deep ensemble learning, and the AD prediction apparatus is a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor implements the steps of the AD prediction method based on deep ensemble learning as in the present embodiment when executing the program. The computer device may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a rack server, a blade server, a tower server or a cabinet server (including an independent server or a server cluster composed of a plurality of servers) capable of executing programs, and the like. The computer device of the embodiment at least includes but is not limited to: a memory, a processor communicatively coupled to each other via a system bus. The memory (i.e., readable storage medium) includes flash memory, hard disks, multimedia cards, card-type memory (e.g., SD or DX memory, etc.), random Access Memory (RAM), static Random Access Memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disks, optical disks, and the like. The memory may be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device, or an external storage unit of the computer device, such as a plug-in hard disk provided on the computer device, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like. Of course, the memory may also include both internal and external storage devices for the computer device. In this embodiment, the memory is generally used for storing an operating system, various types of application software, and the like installed in the computer device. In addition, the memory may also be used to temporarily store various types of data that have been output or are to be output. The processor may be a Central Processing Unit (CPU), controller, microcontroller, microprocessor, or other data Processing chip in some embodiments. The processor is typically used to control the overall operation of the computer device and, in this embodiment, is used to execute program code stored in memory or to process data.

The invention and its embodiments have been described above schematically, without limitation, and the invention can be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The representation in the drawings is only one of the embodiments of the invention, the actual construction is not limited thereto, and any reference signs in the claims shall not limit the claims concerned. Therefore, without departing from the spirit of the present invention, a person of ordinary skill in the art should also understand that the present invention shall not be limited to the embodiments and the similar structural modes of the present invention. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Several of the elements recited in the product claims may also be implemented by one element in software or hardware. The terms first, second, etc. are used to denote names, but not to denote any particular order.

Claims

1. A method for predicting AD based on deep ensemble learning comprises the steps of,

carrying out data amplification on the sliced PET image;

and fusing the classification results, and predicting and outputting the fused result, wherein the fusion means that the features in the ResNet-34 network and the EfficientNet-b1 network classified by the MLP are subjected to element-by-element addition fusion to obtain a new feature.

2. The method of claim 1, wherein the ResNet-34 network is a residual structure ratio (3,4,6,3) network.

3. The AD prediction method based on the deep ensemble learning of claim 2, wherein the scaling weight of the EfficientNet-b1 network is obtained, and the new weight value of the EfficientNet-b1 network is obtained by performing transfer learning on the amplified PET slice image; the EfficientNet-b1 network automatically searches for a composite coefficient, and optimizes a prediction network through three dimensions of depth, width and resolution of the network, wherein the prediction network is a ResNet-34 network and an EfficientNet-b1 network.

4. The deep ensemble learning-based AD prediction method according to claim 3, wherein the EfficientNet-b1 network structure comprises a backbone network, global average pooling, global maximum pooling, feature stitching and a classifier.

5. The AD prediction method based on the deep ensemble learning of claim 4, wherein after the amplified PET slice image is input to a backbone network, the EfficientNet-b1 network learns the feature information of the amplified PET slice image, generates a corresponding feature matrix, performs global average pooling and global maximum pooling, obtains two groups of one-dimensional feature information for information feature splicing, classifies the spliced feature information by a classifier, performs back propagation, and performs multiple iterations to form the optimal weight.

6. The method of claim 5, wherein 10k-fold cross validation is used to perform model training on a plurality of processed PET slice images, one of the sets is used as a validation set during training, the remaining images are used as a training set, the optimizer uses SGD, the learning rate decay strategy is StepLR, and the loss function is a cross entropy loss function.

7. The method according to claim 6, wherein the cross entropy loss function L is a function of the depth ensemble learning (AD) _CE The calculation formula of (2) is as follows:

8. The AD prediction method based on the deep ensemble learning of claim 7, wherein a 10k-fold cross validation model is selected in the prediction training process; four amplification modes of random scale transformation (1), random small-angle rotation (2), random horizontal mirror image (3) and random vertical mirror image (4) are used.

9. An AD prediction system based on deep ensemble learning, which is characterized in that the AD prediction method based on deep ensemble learning of any one of claims 1-8 is adopted to carry out feature learning on a PET image; the AD prediction system includes:

the input module is used for inputting the PET image, performing slicing processing on the PET image, and performing amplification on the PET image to increase samples of a training set;

10. An AD prediction apparatus based on deep ensemble learning, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the AD prediction method based on deep ensemble learning according to any one of claims 1 to 8 when executing the program.