Parkinson prediction method based on medical images
By improving the ResNet18 network model, combining 3×3×3 ordinary convolutions and 3D depthwise separable convolutions, and using ECA attention mechanism and dual-channel pooling layers, the accuracy of Parkinson's disease diagnosis based on medical images has been improved, achieving early detection and accurate classification, and solving the problem of low accuracy in existing technologies.
Patent Information
- Application Number
- CN202310751471.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-25
- Publication Date
- 2026-01-16
- Estimated Expiration
- 2043-06-25
AI Technical Summary
Existing AI-based classification and prediction models for Parkinson's disease diagnosis based on medical images are not very accurate, making it difficult to detect Parkinson's disease in its early stages and causing patients to miss the golden period for treatment.
An improved ResNet18 network model is used, which combines 3×3×3 ordinary convolution and 3D depthwise separable convolution for feature extraction. The ECA attention mechanism is used for feature fusion, and a dual-channel pooling layer and two fully connected layers are added to the output module to automatically classify Parkinson's disease.
It has improved the accuracy of Parkinson's disease diagnosis, enabled early detection, reduced the economic and medical burden on patients and their families, and promoted the common progress of medicine and society.
Smart Images
Figure CN116778245B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical image processing, in particular to a Parkinson prediction method based on medical images. BACKGROUND
[0002] With the arrival of social aging, the number of patients and the prevalence rate of Parkinson and other common neurological diseases in the middle-aged and elderly are also increasing year by year. However, the early symptoms of Parkinson's disease are not obvious, and it is easy to be ignored when seeking medical treatment. When diagnosing, doctors need to rely on clinical experience and complex procedures for final assessment. The evaluation process is long and there is a certain personal subjective factor. Therefore, many Parkinson patients have reached the middle and late stages when they are diagnosed, missing the golden period of early treatment, which brings great economic burden and medical pressure to patients, families and society.
[0003] In recent years, with the gradual maturity of artificial intelligence and the rise of deep learning, more and more people have begun to study its application in the medical field, such as pathological detection, patient service, auxiliary diagnosis, etc. Not only does it save patients' time and improve the quality of diagnosis, but also reduces medical errors. Similarly, it has also made certain progress in the diagnosis of Parkinson's disease and other neurological diseases. Various models of deep learning can not only assist doctors in intelligent diagnosis and health assessment of diseases, but also help detect early Parkinson's disease patients, facilitate early detection, early intervention and early treatment of patients, and grasp the golden period of Parkinson's disease treatment. Not only does it reduce the economic and medical burden of patients and families, but also promotes the joint progress of medicine and society.
[0004] At present, the artificial intelligence classification prediction model for auxiliary diagnosis of Parkinson mainly based on motor symptoms (speech disorders, frozen gait, tremors, facial expressions, etc.) and non-motor symptoms (cognition, imaging, scales, clinical indicators, etc.). Common classification data types include scales, speech, gait, electroencephalogram (EEG) and medical images. Compared with scales, speech, gait, EEG (Electroencephalogram, brain waves) and other data types, image data has fewer restrictions when collecting, is more convenient to collect, and can directly observe the changes in brain internal structure and function. It is also non-invasive, has high spatial resolution, good tissue contrast, etc. Although there are artificial intelligence classification prediction models for auxiliary diagnosis of Parkinson based on image data in the prior art, the accuracy of such artificial intelligence classification prediction models still needs to be further improved. SUMMARY
[0005] The technical problem to be solved by the present application is to provide a Parkinson prediction method based on medical images with high accuracy.
[0006] To solve the above technical problems, the technical solutions of the present application are as follows:
[0007] A Parkinson prediction method based on medical images, comprising:
[0008] obtaining an MRI image to be evaluated;
[0009] preprocessing the MRI image;
[0010] inputting the preprocessed MRI image into a pre-constructed artificial intelligence classification prediction model to obtain a prediction result, wherein the artificial intelligence classification prediction model is an improved ResNet18 network model.
[0011] Further, the improved ResNet18 network model comprises an input module, a feature extraction module and an output module connected in sequence, wherein the input module comprises a 3*3*3 ordinary convolution and a three-dimensional depth separable convolution.
[0012] Further, the input module comprises a 3*3*3 ordinary convolution, a batch normalization layer, a three-dimensional depth separable convolution and a maximum pooling layer connected in sequence.
[0013] Further, in the feature extraction module, N feature maps obtained by N times of feature extraction from shallow to deep are retained, and N-1 feature maps of a hierarchical depth are up-sampled to keep the shape consistent with the original feature map of the previous layer, then the original feature map of the previous layer is adjusted by an ECA attention mechanism and fused with the up-sampled feature map to obtain N-1 new feature maps, and the deepest feature map is adjusted by an ECA attention mechanism to obtain the Nth new feature map, thereby obtaining N new feature maps; thereafter, the N new feature maps are adjusted to the same size and concatenated in the channel dimension.
[0014] Further, the up-sampling comprises a 1*1*1 convolution and a 3D deconvolution.
[0015] And / or, the ECA attention mechanism is a 3D ECA attention mechanism, which first performs spatial feature compression on the C*H*W*D feature map through three-dimensional average pooling to obtain a new C*1*1*1 feature map, then learns channel features using a weight-shared three-dimensional convolution, and finally fuses the channel feature map and the original feature map through a channel-by-channel multiplication method to output a feature map with channel attention;
[0016] And / or, the adjusting the N new feature maps to the same size comprises: performing a 3*3*3 convolution operation on the N-1 new feature maps respectively to eliminate the aliasing effect generated in the operation, and then adjusting the N-1 new feature maps to the same size as the Nth new feature map through a pooling layer. Figure 1Sample size feature map.
[0017] Further, the output module comprises a double-channel pooling layer, a first full connection layer, a hidden layer and a second full connection layer connected in sequence.
[0018] Further, the double-channel pooling layer comprises a maximum pooling layer and an average pooling layer connected in parallel.
[0019] Further, the preprocessing of the MRI image comprises:
[0020] Converting the format of the MRI image into NIfTI format.
[0021] Further, the preprocessing of the MRI image comprises:
[0022] After adjusting the appropriate fractional intensity threshold, threshold gradient and initial brain center coordinates, the MRI image in NIfTI format is automatically separated to remove irrelevant tissues including skull, fat and cerebellum in the brain.
[0023] Further, the preprocessing of the MRI image comprises:
[0024] The MRI image is automatically registered into the MNI152 standard template to align the brain image in spatial position.
[0025] The present application has the following beneficial effects:
[0026] The present application combines the gradually mature medical image and deep learning, plays the advantages of artificial intelligence and applies it to actual life. Based on the residual network as the basic framework, the medical image characteristics of Parkinson's disease are improved to realize the automatic classification of Parkinson's disease, and the accuracy is high. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 It is a flowchart of the Parkinson prediction method based on medical image of the present application;
[0028] Figure 2 It is a structural schematic diagram of the original ResNet18 network model of the prior art;
[0029] Figure 3 It is a structural schematic diagram of the improved ResNet18 network model in the present application;
[0030] Figure 4 It is a structural schematic diagram of the three-dimensional depth separable convolution in the present application;
[0031] Figure 5 It is a structural schematic diagram of the input module before modification of the prior art;
[0032] Figure 6 Figure 1 is a structural schematic diagram of the input module in the present application after modification;
[0033] Figure 7 Figure 2 is a structural schematic diagram of the feature extraction module in the present application;
[0034] Figure 8 Figure 3 is a structural schematic diagram of the ECA attention mechanism in the prior art before modification;
[0035] Figure 9 Figure 4 is a structural schematic diagram of the ECA attention mechanism in the present application after modification;
[0036] Figure 10 Figure 5 is a structural schematic diagram of the output module in the prior art before modification;
[0037] Figure 11 Figure 6 is a structural schematic diagram of the output module in the present application after modification;
[0038] Figure 12 Figure 7 is a comparison schematic diagram of the MRI image before and after skull separation in the present application;
[0039] Figure 13 Figure 8 is a comparison schematic diagram of the MRI image before and after standard position registration in the present application. DETAILED DESCRIPTION
[0040] To make the technical problems, technical solutions and advantages of the present application clearer, specific embodiments will be described in detail below with reference to the accompanying drawings.
[0041] The present application provides a Parkinson prediction method based on medical images, as shown in Figure 1 , which comprises:
[0042] Step 1: obtaining an MRI (Magnetic Resonance Imaging) image to be evaluated;
[0043] In this step, the MRI image of the patient to be evaluated is obtained.
[0044] Step 2: pre-processing the MRI image;
[0045] Step 3: inputting the pre-processed MRI image into a pre-constructed artificial intelligence classification prediction model to obtain a prediction result, wherein the artificial intelligence classification prediction model is an improved ResNet18 network model.
[0046] In this step, the artificial intelligence classification prediction model is based on an improved ResNet18 network model, and the original ResNet18 network model is as shown in Figure 2As shown, the improved ResNet18 network model is based on a 3D_Resnet18 architecture, which is adjusted according to the characteristics of Parkinson's images, and the improved ResNet18 network model is as shown in Figure 3 As shown, the model as a whole includes an input module, a feature extraction module and an output module connected in sequence, and each module of the model is adjusted / modified as follows:
[0047] A. Improvement of the input module
[0048] In the input module, the original model performs initial feature extraction through a 7x7x7 convolution (see Figure 2 ). In the new model, it is modified to a 3x3x3 ordinary convolution and a three-dimensional depth separable convolution (see Figure 3 ). That is, the input module includes a 3x3x3 ordinary convolution and a three-dimensional depth separable convolution.
[0049] In the input module, as shown in Figure 5 , the original network is composed of a 7x7x7 convolution layer, a batch normalization processing layer and a maximum pooling layer, and the calculation amount is proportional to the lengthxwidthxheight of the convolution kernel. Considering the characteristics of low resolution of MRI and other medical images, more attention to details, large amount of three-dimensional image operation, etc., the 7x7x7 convolution kernel is replaced by a 3x3x3 convolution kernel for shallow feature information extraction. A group of three-dimensional depth separable convolution is added behind it, which reduces the parameter amount and calculation amount while ensuring the feature extraction effect of shallow effective information. The three-dimensional depth separable convolution is as shown in Figure 4 , and the structure of the input module before and after modification is as shown in Figures 5-6 , and the parameter amount formula is as shown below:
[0050] Parameter amount before modification:
[0051] Parameter amount after modification:
[0052] Where K7 represents a convolution kernel with a length, width and height of 7, K3 represents a convolution kernel with a length, width and height of 3, M represents the number of input channels, N represents the number of output channels, and M ′ represents the input channel number of point convolution. Since the medical image has only one channel, the original input channel number is 1, i.e. M=1. After the first layer of convolution, the output channel number is adjusted to half of the original output channel number, and the input channel number of the grouped convolution is
[0053] In specific implementation, as shown in Figure 6As shown, the input module can include a 3x3x3 normal convolution, a batch normalization layer, a three-dimensional depth separable convolution, and a max pooling layer connected in sequence, wherein the three-dimensional depth separable convolution can include a channel-wise convolution and a point-wise convolution connected in sequence.
[0054] B. Improvement of the feature extraction module
[0055] In the feature extraction module, the original model inputs the feature map into the output module after feature extraction. The new model retains the N feature maps obtained by N times of feature extraction from shallow to deep (N is an integer) in this part, and up-samples the N-1 feature maps with a higher level to keep the shape consistent with the original feature map of the previous layer. Then, the original feature map of the previous layer is adjusted by the ECA attention mechanism and fused with the up-sampled feature map to obtain N-1 new feature maps. Meanwhile, the deepest feature map is adjusted by the ECA attention mechanism to obtain the Nth new feature map, thereby obtaining a total of N new feature maps. Then, the N new feature maps are adjusted to the same size and concatenated in the channel dimension. Finally, the fused feature map is input into the output module.
[0056] In Figure 3 In the embodiment shown, taking N=4 times of feature extraction as an example, the new model retains the 4 feature maps obtained by 4 times of feature extraction from shallow to deep in this part, and up-samples the 3 feature maps with a higher level to keep the shape consistent with the original feature map of the previous layer. Then, the original feature map of the previous layer is adjusted by the ECA attention mechanism and fused with the up-sampled feature map to obtain 3 new feature maps. Meanwhile, the deepest feature map is adjusted by the ECA attention mechanism to obtain the 4th new feature map, thereby obtaining a total of 4 new feature maps. Then, the 4 new feature maps are adjusted to the same size and concatenated in the channel dimension. Finally, the fused feature map is input into the output module.
[0057] In specific implementation, the feature extraction module can be divided into four stages, namely stage1-stage4, and each stage uses a residual structure to integrate high-dimensional and low-dimensional information. Since medical images themselves have the characteristics of small image size and low resolution, and three-dimensional images also need to pay more attention to spatial information. However, in the original down-sampling process, although the semantic information is gradually enhanced, the spatial information and the ability to perceive details are reduced. Therefore, the overall architecture of the feature extraction module is based on a residual network, and is adjusted according to the characteristics of Parkinson's images as follows:
[0058] (1) Retain the 4 feature maps obtained from shallow to deep in the stage1-stage4 process, namely feature1-feature4.
[0059] (2) featureN and feature(N-1) are made consistent in the H (Height), W (Width), and D (Depth) dimensions through 1x1x1 convolution and 3D deconvolution. That is, the upsampling described above includes 1x1x1 convolution and 3D deconvolution. In a specific implementation, the upsampling can include, in sequence, 1x1x1 convolution, a first batch normalization layer, 3D deconvolution, and a second batch normalization layer (see Figure 7 ).
[0060] (3) Channel weight assignment is performed on feature1-feature4 using the modified 3D ECA attention mechanism to adjust the weight dependency relationship between channels.
[0061] (4) featureN in step (2) and feature(N-1) in step (3) are added to obtain feature3', feature2', feature1', and feature4 with adjusted channel weights, respectively.
[0062] (5) 3x3x3 convolution operations are performed on feature1', feature2', and feature3' to eliminate the aliasing effects generated in the above operations. Then, feature3', feature2', and feature1' are adjusted to feature maps of the same size as feature4 through a pooling layer. That is, the aforementioned adjustment of the N new feature maps to the same size can include: performing 3x3x3 convolution operations on the N-1 new feature maps (i.e., the N-1 feature maps at a shallow level) to eliminate the aliasing effects generated in the operations, and then adjusting the N-1 new feature maps to feature maps of the same size as the Nth new feature map through a pooling layer. Figure 1
[0063] (6) Finally, the adjusted feature1', feature2', feature3', and feature4 are subjected to a concat operation in the channel dimension to obtain a three-dimensional feature map that contains both spatial information and semantic information. The overall process is shown in Figure 7 .
[0064] In the above step (3), in order to enable the ECA attention mechanism to better use the 3D image, one-dimensional convolution is changed to three-dimensional convolution, so that it can learn three-dimensional channel attention information.
[0065] As shown in Figure 9 As shown, the ECA attention mechanism is preferably a 3D ECA attention mechanism, which first performs spatial feature compression on the feature map of C*H*W*D through three-dimensional average pooling to obtain a new feature map of C*1*1*1, then performs channel feature learning using a weight-shared three-dimensional convolution, and finally performs channel fusion on the channel feature map and the original feature map through a channel-by-channel multiplication method to output a feature map with channel attention. Wherein, C is the number of channels, H is the height, W is the width, and D is the depth. The method of adaptively selecting the three-dimensional convolution kernel size can effectively determine the coverage range of local cross-channel interaction, and the formula is as follows, wherein γ = 2, b = 1, and C represents the number of channels.
[0066]
[0067] C. Improvement of the output module
[0068] In the output module, the original model is classified through an average pooling layer and a fully connected layer (see Figure 10 ). In the new model, the feature map is outputted through a double-channel pooling layer, a hidden layer and two fully connected layers in the output module (see Figure 11 ) to output the final classification result.
[0069] In the original output module, the feature map after feature extraction is adjusted in shape through an average pooling layer and outputted through a fully connected layer to output the classification result. However, in the new model, the final output feature map is larger due to the modification of the feature extraction module, so a three-dimensional double-channel pooling layer is used here to adjust the size of the feature map, and then a hidden layer and two fully connected layers are used to output the final classification result. That is, as shown in Figure 11 , the output module can include a double-channel pooling layer, a first fully connected layer, a hidden layer and a second fully connected layer connected in sequence, wherein the double-channel pooling layer can include a maximum pooling layer and an average pooling layer connected in parallel.
[0070] As an optional embodiment, the pre-processing of the MRI image (step 2) can include:
[0071] Step 21: converting the format of the MRI image into NIfTI format;
[0072] This step is format conversion. In specific implementation, all MRI images can be converted into NIfTI format using dcm2niix in MRIcron software to convert the DICOM format image into NIfTI format, and the converted NIfTI image can be viewed using visualization browsing tools such as MRIcron.
[0073] Step 22: The NIfTI format MRI image is automatically separated after adjusting the appropriate fractional intensity threshold, threshold gradient and initial brain center coordinates to remove irrelevant tissues including skull, fat and cerebellum in the brain;
[0074] This step is skull separation. In implementation, skull separation can use the BET component in FSL to automatically separate each NIfTI image after adjusting the appropriate fractional intensity threshold, threshold gradient and initial brain center coordinates, etc. The purpose of this step is to remove irrelevant tissues such as skull, fat and cerebellum in the brain to avoid their possible impact.
[0075] Step 23: The MRI image is automatically registered into the MNI152 standard template to align the brain image in spatial position.
[0076] This step is standard position registration. In implementation, standard position registration can use the FLIRT tool in FSL to automatically register the MRI image into the MNI152 standard template. The purpose of this step is to align different brain images in spatial position to avoid the impact of different brain shapes and positions.
[0077] The accuracy of the Parkinson prediction method based on medical images of the present application is verified by experiment as follows:
[0078] Data set: The experimental data is provided by the Second Affiliated Hospital of Shandong University, which includes 122 MRI images of Parkinson's disease patients and 136 MRI images of normal people. The experimental group, i.e. the PD (Parkinson's disease) group, has 75 males and 47 females, with an age of 50-75 years old and an average age of 62.73±9.53 years old. All patients are right-handed. The control group has 76 males and 60 females, with an age of 48-73 years old and an average age of 60.74±11.33 years old, all of whom are right-handed. The population in the PD group and the control group has no history of alcohol abuse, no history of drug abuse, no history of seizures, no brain parenchymal lesions, no history of head trauma, and no contraindications for MRI examination.
[0079] Data preprocessing:
[0080] A. Format conversion
[0081] Format conversion is performed on all MRI images. The dcm2niix in MRIcron software is used to convert the DICOM format image to NIfTI format. The converted NIfTI image can be viewed using visualization tools such as MRIcron.
[0082] B. Skull separation
[0083] The skull separation uses the BET component in FSL to automatically separate each NIfTI image after adjusting the appropriate fractional intensity threshold, threshold gradient and initial brain center coordinates and the like. The purpose of this step is to remove the experimental irrelevant tissues such as brain skull, fat and cerebellum, and to avoid the influence on the experiment. The comparison before and after the skull separation of the MRI image is shown in FIG. 1. Figure 12
[0084] C. Standard position registration
[0085] The standard position registration uses the FLIRT tool in FSL to automatically register the MRI image into the MNI152 standard template. The purpose of this step is to align the brain images taken at different times, by different individuals and by different machines in the spatial position, so as to avoid the influence of the different brain shapes and positions on the experiment. The comparison before and after the standard position registration of the MRI image is shown in FIG. 2. Figure 13
[0086] Experimental parameters: The experiment is trained on a cloud server, the number of GPUs is 2, the type of GPU is gpu_class_6, and the memory size is 120G. The experiment is trained for 120 rounds, and the cross-entropy loss function and the SGD (gradient descent method) optimizer are used to optimize various parameters. Among them, momentum = 0.9, weight_decay = 0.001, nesterov = True. The initial value of the learning rate is 0.006, and every 30 rounds, the learning rate is adjusted to half of the original value.
[0087] Evaluation index: The classification performance evaluation index of the experiment is Accuracy, Precision, Recall and F1, and the formulas are as follows:
[0088]
[0089]
[0090]
[0091]
[0092] Among them, TP, TN, FP and FN respectively represent the correct positive sample, the correct negative sample, the incorrect positive sample and the incorrect negative sample. TP is the true positive, that is, the PD patient is correctly divided into the PD patient. TN is the true negative, that is, the healthy person is correctly divided into the healthy person. FP is the false negative, that is, the healthy person is incorrectly divided into the PD patient. FN is the false negative, that is, the PD patient is incorrectly divided into the normal person.
[0093] The accuracy and recall rate of the present application and the original model before improvement are compared, and the table is shown as follows:
[0094] Table 1
[0095] Model Accuracy / % Precision / % Recall / % F1 / ％ ResNet18 model 70.35 66.60 59.64 60.10 Improved model of the present invention 85.68 81.63 68.80 69.60
[0096] In summary, the Parkinson prediction method based on medical images of the application first acquires the MRI image to be evaluated, then pre-processes the MRI image, and finally inputs the pre-processed MRI image into a pre-constructed artificial intelligence classification prediction model to obtain a prediction result, wherein the artificial intelligence classification prediction model is an improved ResNet18 network model. The application combines the mature medical images and deep learning, takes advantage of artificial intelligence and applies it to real life. Based on the residual network architecture, the Parkinson disease is improved according to the medical image characteristics, and the automatic classification of Parkinson disease is realized, with high accuracy.
[0097] The characteristics of Parkinson's disease, such as difficulty in discovery, diagnosis, long course, high disability rate, and high treatment cost, not only bring spiritual and economic pressure to patients and their families, but also bring heavy medical burden to society. The maturity of artificial intelligence and the rise of deep learning make people begin to try its application in the medical field. Through the deep learning model, intelligent diagnosis of Parkinson's disease is realized, which not only provides certain objective standard for the diagnosis result of doctors, but also greatly reduces the economic burden of patients and families and the diagnosis and treatment pressure of society due to its low cost and machine fatigue resistance. In the future, with the continuous deepening of research, there will be better application prospects.
[0098] The existing Parkinson deep learning auxiliary diagnosis and treatment method is mostly based on scales, voices, gaits, etc. The data set in the collection process often has various problems, such as strong personal subjectivity, high collection environment requirement, weak applicability, etc. The application is based on medical image data, which has fewer restrictions during collection and is more convenient to collect. Moreover, it can directly observe the changes of internal structure and function of the brain, and has the advantages of non-invasiveness, high spatial resolution, good tissue contrast, etc. Classification of Parkinson's disease through medical images is more conducive to the accuracy and objectivity of classification.
[0099] At present, the auxiliary diagnosis method based on medical images is mostly based on traditional machine learning methods, without fully considering the combination of Parkinson three-dimensional image spatial features and semantic information, and lacking targeted adjustment for Parkinson three-dimensional image characteristics. The application fully considers the changes of spatial information and semantic information of three-dimensional images, and performs deep and shallow feature fusion in the feature extraction module. At the same time, in the initial feature extraction, considering the characteristics of low image resolution and large three-dimensional image calculation, the initial convolution layer is adjusted. In the output module, double-channel pooling is added for shape adjustment, the hidden layer is prevented from overfitting, and two fully connected layers are added to improve the final classification effect.
[0100] In addition, the present application also has the following beneficial effects:
[0101] (1) In the initial feature extraction process, the problem of low resolution of medical images is considered, the 7*7*7 convolution is modified to 3*3*3 convolution, and a set of 7*7*7 deep separable convolution is added behind it, which reduces the parameter amount to the maximum while ensuring that the feature extraction effect does not change.
[0102] (2) In the feature extraction process, the problem of spatial information loss is considered, and deep and shallow feature fusion is performed, fully utilizing the spatial information and semantic information in three-dimensional images.
[0103] (3) In the output module, three-dimensional double-channel pooling layer is used for adjusting the size of feature map, hidden layer is used to prevent overfitting, and two fully connected layers are used for classification, so as to improve the classification effect as much as possible.
[0104] (4) Artificial intelligence and deep learning are applied to practical life, a deep learning model is built according to Parkinson's image, and classification prediction of Parkinson's disease is realized. It can give doctors some objective reference, and assist doctors to realize the discovery and diagnosis of Parkinson's disease.
[0105] The above is the preferred embodiment of the present application, it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.
Claims
1. A method for predicting Parkinson's disease based on medical images, characterized by, The method comprises: acquiring an MRI image to be evaluated; preprocessing the MRI image; inputting the preprocessed MRI image into a pre-constructed artificial intelligence classification prediction model to obtain a prediction result, wherein the artificial intelligence classification prediction model is an improved ResNet18 network model; the improved ResNet18 network model comprises an input module, a feature extraction module and an output module connected in sequence, wherein the input module comprises a 3*3*3 ordinary convolution and a three-dimensional depth separable convolution; the input module comprises a 3*3*3 ordinary convolution, a batch normalization layer, a three-dimensional depth separable convolution and a maximum pooling layer connected in sequence; in the feature extraction module, N feature maps obtained by N times of feature extraction from shallow to deep are retained, and N-1 feature maps with a deeper level are up-sampled to keep the shape consistent with the original feature map of the previous layer, then the original feature map of the previous layer is adjusted by an ECA attention mechanism and fused with the up-sampled feature map to obtain N-1 new feature maps, and the feature map with the deepest level is adjusted by the ECA attention mechanism to obtain the Nth new feature map, thereby obtaining N new feature maps; then, the N new feature maps are adjusted to the same size and concatenated in the channel dimension; the up-sampling comprises a 1*1*1 convolution and a 3D deconvolution; and / or, the ECA attention mechanism is a 3D ECA attention mechanism, which first performs spatial feature compression on the C*H*W*D feature map through three-dimensional average pooling to obtain a new C*1*1*1 feature map, then learns channel features using a weight-shared three-dimensional convolution, and finally fuses the channel feature map and the original feature map through a channel-by-channel multiplication method to output a feature map with channel attention; and / or, the step of adjusting the N new feature maps to the same size comprises: performing a 3*3*3 convolution operation on the N-1 new feature maps respectively to eliminate aliasing effects generated in the operation, and then adjusting the N-1 new feature maps to feature maps with the same size as the Nth new feature map through a pooling layer; the output module comprises a double-channel pooling layer, a first full connection layer, a hidden layer and a second full connection layer connected in sequence; the double-channel pooling layer comprises a maximum pooling layer and an average pooling layer connected in parallel.
2. The method of any one of claim 1, wherein, The preprocessing of the MRI image comprises: converting the format of the MRI image to NIfTI format.
3. The method of claim 2, wherein, The preprocessing of the MRI image comprises: automatically separating the MRI image in NIfTI format after adjusting appropriate fractional intensity threshold, threshold gradient and initial brain center coordinates to remove irrelevant tissues including skull, fat and cerebellum in the brain.
4. The method of claim 2, wherein, The preprocessing of the MRI image comprises: automatically registering the MRI image to the MNI152 standard template to align the brain image in spatial position.