A Parkinson's disease progression prediction method based on the BCL model

Abstract

The present invention discloses a Parkinson's disease progression prediction method based on the BCL model, which relates to the technical field of medical imaging data, and includes: S1, obtaining clinical data including the H&Y score and basic patient information, as well as image data including MRI images; S2, performing format conversion on the MRI images and performing data annotation in combination with the H&Y score; S3, preprocessing the images after data annotation; S4, using principal component analysis to reduce the dimension of the preprocessed images and extract the low-dimensional features of the images; S5, constructing a curriculum learning framework and a classification model based on a balanced weight scheduler to train the features after dimension reduction; S6, predicting the progression of Parkinson's disease in patients and evaluating the prediction performance of each classification model. The present invention can provide a more accurate assessment of the patient's disease stage, reduce the data dimension while improving the computational efficiency and generalization ability of the model, and enhance the model's learning ability for difficult samples.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging data, and in particular to a method for predicting Parkinson's disease progression based on a BCL model. Background Art

[0002] Parkinson's disease is a neurodegenerative disease that affects millions of people worldwide. Early diagnosis and monitoring of disease progression are crucial for effective treatment. Traditionally, the diagnosis of Parkinson's disease mainly relies on clinical evaluation and observation of patient symptoms. However, this method may lack accuracy in the early stages of the disease and is difficult to predict the progression speed and severity of the disease. Although some studies have attempted to use machine learning techniques to assist in the diagnosis of Parkinson's disease, most studies have focused on initial diagnosis, and relatively few studies have been conducted on the prediction of disease progression.

[0003] With the rapid development of medical imaging technology, neuroimaging methods such as MRI have been widely used in the study of neurodegenerative diseases. MRI images can provide detailed brain structure information and provide an important tool for the study of Parkinson's disease. However, due to the high data dimension and complex processing of MRI images, how to effectively extract useful features from them and apply them to machine learning models remains a challenge. Therefore, the prediction of Parkinson's disease progression based on MRI images is relatively rare in the research.

[0004] In the prior art, researchers have begun to explore how to combine MRI images with machine learning techniques to better perform Parkinson's disease classification tasks. However, most of the existing studies have focused on the binary classification task of Parkinson's disease, lacking in-depth analysis of different disease progression stages, and in the multi-classification task, there are still problems of data imbalance. In addition, the existing methods have not fully utilized the feature differences of MRI images at each stage, resulting in the inability to accurately reflect the severity and progression of the patient's condition. Summary of the Invention

[0005] In order to solve the problem that the prior art cannot accurately predict the progression of Parkinson's disease, the present invention proposes a method for predicting Parkinson's disease progression based on a BCL model, effectively combining MRI images with the H&Y score, and based on unbalanced data, making full use of the multi-stage information of MRI images to more comprehensively evaluate the development of the patient's condition to solve the above problems.

[0006] The present application discloses a method for predicting Parkinson's disease progression based on a BCL model, including the following steps:

[0007] S1, obtaining clinical data including the H&Y score and patient basic information, and image data including MRI images;

[0008] S2. Convert the format of the MRI images and perform data annotation in combination with the H&Y score;

[0009] S3. Preprocess the images after data annotation;

[0010] S4. Use principal component analysis to reduce the dimension of the preprocessed images and extract the low-dimensional features of the images;

[0011] S5. Construct a curriculum learning framework based on a balanced weight scheduler, gradually introduce samples of different difficulties, construct different classification models and train the reduced-dimensional features to obtain a curriculum learning model based on a balanced weight scheduler;

[0012] S6. Use the curriculum learning model based on a balanced weight scheduler to predict the progression of Parkinson's disease in patients and evaluate the prediction performance of each classification model.

[0013] Preferably, the S2 includes the following steps:

[0014] S21. Read the clinical data file containing the H&Y score and the image data file containing the basic information of the MRI images, integrate the two data files of the H&Y score and the basic information of the MRI images into a new field, and store them in a formatted manner to achieve the mapping between the image ID and the H&Y score;

[0015] S22. Use MRIcroGL to convert the DICOM format MRI images into NIFTI format and save the basic metadata related to the images, including image size, pixel size, image orientation, and time point information;

[0016] S23. For each NIfTI image, extract the corresponding image ID, obtain the H&Y score corresponding to this image according to the mapping formed in S21, save the image path of this MRI and the corresponding H&Y score, and achieve the data annotation of the MRI images.

[0017] Preferably, the S3 includes the following steps:

[0018] Read the image path and convert the format to float32. For the image data that meets the dimension conditions, use OpenCV for resampling and normalization processing to standardize the images to the same resolution and size. The dimension conditions are: the image data must be three-dimensional data, and the dimension of each slice is the image data of a valid two-dimensional matrix.

[0019] Preferably, the resampling includes the following steps:

[0020] Set the target affine matrix size to , resample each slice of each MRI image using the resize function of cv2, and return the resampled image object;

[0021] The normalization process includes the following steps:

[0022] Calculate the mean and standard deviation of the pixel values of the input image. Subtract the image mean from each pixel value to eliminate the influence of overall brightness, and divide by the sum of the standard deviation and a smoothing term to prevent division by zero when the standard deviation is zero. Each normalized pixel value represents the number of standard deviations relative to the original data distribution. The calculation formula is as follows:

[0023]

[0024] where,

[0025]

[0026]

[0027] is the image pixel value, is the pixel value of the input image, is the image mean, is the image standard deviation, is the smoothing term, is the total number of samples.

[0028] Preferably, S4 includes the following steps:

[0029] For all images of size flatten the image data into a one-dimensional vector;

[0030] Use principal component analysis for feature extraction to achieve data dimensionality reduction. Re-integrate the extracted principal components into the original data frame to obtain a data set, and divide the data set into a training set and a test set. Each principal component serves as a new feature column.

[0031] Preferably, the flattening of the image data into a one-dimensional vector includes the following steps:

[0032] For each image, straighten its pixel values to form a one-dimensional array, and combine all the one-dimensional arrays into a two-dimensional array with the shape: (number of samples, number of pixels), achieving the conversion of a high-dimensional image to a two-dimensional data matrix;

[0033] The use of principal component analysis for feature extraction includes the following steps:

[0034] Use principal component analysis to reduce the flattened high-dimensional image data to 50 dimensions, reducing the data dimension while retaining most of the information.

[0035] Preferably, S5 includes the following steps:

[0036] S51. Divide the training set samples into 5 levels according to the H&Y score, and divide the training set after dividing the sample levels into 4 stages;

[0037] S52. In the framework of curriculum learning, introduce a balanced weight scheduler to adjust the weights of samples in different stages to obtain a curriculum learning framework based on the balanced weight scheduler;

[0038] S53. Construct a classification model including a random forest model, a support vector machine model, and a multi-layer perceptron model;

[0039] S54. Train the model stage by stage. Before the start of each stage, first obtain the number of classes and the number of samples of each class in the training set of the current stage. Secondly, call the balanced weight scheduler to calculate the class weights and generate a weight dictionary. Finally, call the classification model for training respectively, and use the weight dictionary as the loss function to be passed into the classification model. Repeat the above training process until all four stages are trained to obtain a curriculum learning model based on the balanced weight scheduler.

[0040] Preferably, the 4 stages of the training set division include:

[0041] The first stage: only includes the control group samples with a score of 0 and the easiest samples with a score of 4;

[0042] The second stage: on the basis of the first stage, add the relatively easy samples with a score of 3;

[0043] The third stage: on the basis of the second stage, add the relatively difficult samples with a score of 2;

[0044] The fourth stage: on the basis of the third stage, add the most difficult samples with a score of 1.

[0045] Preferably, S6 includes the following steps:

[0046] Apply the trained curriculum learning model based on the balanced weight scheduler to the test set, perform classification prediction on the test samples, calculate each sample in the test set according to the input features, and the output result is the predicted class label of each test sample. The classes include five levels of 0, 1, 2, 3, and 4. The predicted class label corresponds to the H&Y score of the patient's MRI image, which is used to reflect the current stage of Parkinson's disease of the patient.

[0047] Preferably, the evaluation indexes for evaluating the prediction performance of each classification model include accuracy, precision, recall rate, F1 score, mean square error, and specificity.

[0048] Advantages of the present invention:

[0049] (1) The present invention uses the H&Y score combined with MRI images to predict the progression of Parkinson's disease. Different from most of the existing technologies that only focus on the diagnosis of Parkinson's disease, this method can provide a more accurate assessment of the patient's disease stage, which helps to formulate personalized treatment plans.

[0050] (2) The present invention introduces the PCA dimensionality reduction technology in the traditional machine learning model to extract features, which can effectively reduce the redundant information in the MRI images, improve the computational efficiency and generalization ability of the model while reducing the data dimension, and at the same time help reduce the risk of model overfitting.

[0051] (3) The present invention introduces a dynamic class weight adjustment mechanism in the framework of curriculum learning, dynamically calculates the class weights according to the data distribution of each stage, effectively solves the problem of sample class imbalance, enhances the model's learning ability for difficult samples, and at the same time avoids the problem of over-concentration of training on simple classification samples. Description of the Drawings

[0052] Figure 1 It is a schematic flowchart of the Parkinson's progression prediction method based on the BCL model according to the embodiment of the present invention;

[0053] Figure 2 It is a flowchart of the MRI image preprocessing and feature extraction according to the embodiment of the present invention;

[0054] Figure 3 It is a flowchart of constructing a classification model to predict the progression of Parkinson's disease according to the embodiment of the present invention;

[0055] Figure 4 It is a structural diagram of the curriculum learning framework based on the balanced weight scheduler according to the embodiment of the present invention. Detailed Embodiment

[0056] To make the purpose, technical solutions and advantages of the present application more clear, the following examples are given with reference to the accompanying drawings to further elaborate on the present application in detail.

[0057] The embodiment of the present application discloses a Parkinson's progression prediction method based on the BCL model, and the process is as Figure 1 shown, including the following steps:

[0058] S1. Obtain the MRI image data of the patient and the clinical data of the H&Y score.

[0059] Obtain the clinical data including the H&Y score and the basic information of the patient from the PPMI platform, as well as the image data including the MRI images. In this embodiment, the downloaded image data includes 2,781 MRI images of 242 patients, and the downloaded data is stored separately in the clinical data file and the image data file.

[0060] S2. Convert the format of the MRI images and perform data annotation in combination with the H&Y score.

[0061] S21. Read the clinical data file containing the H&Y score, extract basic information such as patient number, visit time, etc. and the H&Y score, and store them in a formatted manner. Read the image data file containing the basic information of the MRI images, extract basic information such as image ID, patient number, gender, age, and visit time, and store them in a formatted manner. Integrate the two data files of the H&Y score and the basic information of the MRI images into a new field through the patient number and visit time, and store them in a formatted manner to achieve the mapping between the image ID and the H&Y score.

[0062] S22. Use MRIcroGL to convert the DICOM format MRI images into NIFTI format, and save the basic metadata related to the images, including image size, pixel size, image orientation, and time point information. Name each MRI image in the way of "image number - MRI imaging information - image acquisition time - image scan serial number.nii.gz".

[0063] S23. For each NIfTI image, extract the corresponding image ID, and obtain the H&Y score corresponding to this image according to the mapping formed in S21. Save the image path of the MRI and the corresponding H&Y score to achieve the data annotation of the MRI images. After the data annotation, there are 2,759 MRI images in total, including 1,892 images with an H&Y score of 0, 195 images with an H&Y score of 1, 603 images with an H&Y score of 2, 52 images with an H&Y score of 3, and 17 images with an H&Y score of 4.

[0064] S3. Preprocess the images after data annotation, and the process is as Figure 2 shown.

[0065] Read the image path and convert the format to float32. For the image data that meets the dimension conditions, use OpenCV for resampling and normalization processing to standardize the images to the same resolution and size. The dimension conditions are: the image data must be three-dimensional data, and the dimension of each slice is the image data of a valid two-dimensional matrix.

[0066] The resampling includes the following steps: Set the target affine matrix size to , use the resize function of cv2 to resample each slice of each MRI image, adjust all the MRI images to a unified size, and return the resampled image object.

[0067] The normalization process includes the following steps: Calculate the mean and standard deviation of the pixel values of the input image, subtract the image mean from each pixel value to eliminate the overall brightness effect, and divide by the sum of the standard deviation and a smoothing term to prevent division by zero errors when the standard deviation is zero. Each normalized pixel value is expressed as the number of standard deviations relative to the original data distribution, and the calculation formula is as follows:

[0068]

[0069] Among them,

[0070]

[0071]

[0072] is the image pixel value, is the pixel value of the input image, is the image mean, is the image standard deviation, is the smoothing term, is the total number of samples.

[0073] S4. Use principal component analysis to reduce the dimension of the preprocessed image and extract the low-dimensional features of the image.

[0074] As Figure 2 shown, for all images of size , flatten the image data into a one-dimensional vector. Specifically, for each image, straighten its pixel values to form a one-dimensional array, and combine all the one-dimensional arrays into a two-dimensional array with the shape: (number of samples, number of pixels), realizing the conversion of the high-dimensional image into a two-dimensional data matrix.

[0075] Use principal component analysis (PCA) for feature extraction to achieve data dimension reduction. Specifically, use principal component analysis to reduce the flattened high-dimensional image data to 50 dimensions, significantly reducing the data dimension while retaining most of the information. Re-integrate the extracted principal components into the original data frame to obtain a data set, and divide the data set into a training set and a test set. Each principal component serves as a new feature column and is used as the model input.

[0076] The data set after data preprocessing and feature extraction includes 2721 images, among which there are 1886 images with an H&Y score of 0, 195 images with an H&Y score of 1, 595 images with an H&Y score of 2, 30 images with an H&Y score of 3, and 15 images with an H&Y score of 4. Among them, the training set contains 2174 MRIs, and the test set contains 547 MRIs, which are saved in the form of pickle files respectively.

[0077] After saving the dataset, load the training set and test set from the pickle file and remove the NaN values. Then, extract 50 features from each MRI image as input features, and extract the H&Y score as the output class label.

[0078] S5. Construct a curriculum learning framework based on a balanced weight scheduler, gradually introduce samples of different difficulties, construct different classification models, and train them on the dimensionality-reduced features to obtain a curriculum learning model based on the balanced weight scheduler. The flowchart for constructing a classification model to predict the progression of Parkinson's disease is as Figure 3 shown.

[0079] S51. Divide the training set samples into 5 difficulties according to the H&Y score, and divide the training set after dividing the sample difficulties into 4 stages. Since the MRI images of patients with more severe diseases will contain more obvious Parkinson's features, the MRI features with a score of 4 (H&Y4) are defined as the easiest samples, the MRI features with a score of 3 (H&Y3) are defined as the relatively easy samples, the MRI features with a score of 2 (H&Y2) are defined as the relatively difficult samples, the MRI features with a score of 1 (H&Y1) are the most difficult samples, and the MRI features with a score of 0 are the control group samples. The 4 stages of dividing the training set samples include:

[0080] The first stage: only contains the control group samples with a score of 0 and the easiest samples with a score of 4;

[0081] The second stage: on the basis of the first stage, add the relatively easy samples with a score of 3;

[0082] The third stage: on the basis of the second stage, add the relatively difficult samples with a score of 2;

[0083] The fourth stage: on the basis of the third stage, add the most difficult samples with a score of 1.

[0084] S52. In the curriculum learning framework, introduce a balanced weight scheduler to adjust the weights of samples in different stages, and obtain a curriculum learning framework based on the balanced weight scheduler (BCL). Give greater weights to the high-difficulty samples with a smaller number, and give smaller weights to the low-difficulty samples with a larger number to alleviate the impact of the imbalance in the number of samples in each stage on the model. The expression is as follows:

[0085]

[0086] Among them, represents the weight of the current class, is the total number of samples in the training set, is the number of samples in the current class, is the number of samples in each class, represents the total number of samples in all classes.

[0087] Specifically, before the start of each stage, first count the number of types of samples included in the current stage, that is, the number of categories of samples in the current stage. Secondly, perform weight conversion. According to the number of samples in each category and the total number of samples, calculate the balanced weights, and pair each category with its corresponding weight. Finally, convert the pairing result into a weight dictionary and pass it as the weight parameter of the loss function into the subsequent classification model.

[0088] S53. Construct a classification model including a Random Forest (RF) model, a Support Vector Machine (SVM) model, and a Multi-Layer Perceptron (MLP) model.

[0089] For the Random Forest model, construct multiple decision trees and determine the final classification result through a voting mechanism. Set the number of decision trees n_estimators = 50, the maximum depth of each decision tree max_depth = 10 to prevent overfitting caused by a single tree being too complex, the minimum number of samples required for each node to split min_samples_split = 10, the minimum number of samples that must be included in a leaf node min_samples_leaf = 5, the total number of features considered by the model each time a decision tree splits max_fearures ='sqrt', that is, the square root of the number of features, to introduce additional randomness to prevent all trees from tending to the same structure, and the random seed random_state = 35.

[0090] For the Support Vector Machine model, achieve the classification prediction of different progression stages of Parkinson's disease by maximizing the classification boundary. Select the Gaussian kernel function as the kernel function, which is used to map non-linear data to a high-dimensional space to find the optimal classification hyperplane. The expression of the Gaussian kernel function is:

[0091]

[0092] Among them, and are the feature vectors of two input samples, is the Euclidean distance between the two, is the hyperparameter of the kernel function, which controls the attenuation rate between samples.

[0093] For the Multi-Layer Perceptron model, perform modeling and achieve multi-classification through a feedforward neural network, specifically including two hidden layers. The first layer contains 100 neurons, the second layer contains 50 neurons, and set the maximum number of training iterations max_iter = 20.

[0094] S54. Train the model stage by stage. The curriculum learning framework structure of the balanced weight scheduler in the embodiment of the present application is as Figure 4As shown, the model is trained stage by stage from the first stage to the fourth stage. Before the start of each stage, the number of categories and the number of samples in each category in the training set of the current stage are obtained. Secondly, the balanced weight scheduler is called to calculate the category weights and generate a weight dictionary. Finally, the random forest model, the support vector machine model, and the multi-layer perceptron model are called for training respectively, and the weight dictionary is passed into the classification model as the loss function. The above training process is repeated until all four stages are trained to obtain a curriculum learning model based on the balanced weight scheduler.

[0095] S6. Use the curriculum learning model based on the balanced weight scheduler to predict the Parkinson's disease stage of the patient.

[0096] Apply the trained curriculum learning model based on the balanced weight scheduler to the test set, perform classification prediction on the test samples, calculate each sample in the test set according to the input features, and the output result is the predicted category label of each test sample. The categories include five levels: 0, 1, 2, 3, and 4. The predicted category label corresponds to the H&Y score of the patient's MRI image and is used to reflect the current Parkinson's disease stage of the patient.

[0097] After the prediction is completed, the performance of each model is evaluated by comparing the prediction results with the true labels. The evaluation metrics used include accuracy, precision, recall, F1-score, mean squared error, and specificity. The calculation formulas are as follows:

[0098] Accuracy:

[0099]

[0100] Precision:

[0101]

[0102] Recall:

[0103]

[0104] F1-score:

[0105]

[0106] Mean squared error:

[0107]

[0108] Specificity:

[0109]

[0110] Among them, is the number of samples correctly predicted as the positive class, is the number of samples correctly predicted as the negative class, is the number of negative samples mispredicted as positive classes, is the number of positive samples mispredicted as negative classes; is the true value of the th sample, is the predicted value of the th sample.

[0111] In a specific embodiment, the performance of different classification models in predicting the progression of Parkinson's disease is compared and analyzed to determine the optimal model in the BCL framework. In this embodiment, the accuracy, precision, recall, F1 score, and mean squared error of three classification models are compared, and the comparison results are shown in Table 1:

[0112] Table 1 Performance comparison table of each classification model in predicting the progression of Parkinson's disease

[0113] Model Accuracy / % Precision / % Recall / % F1 Score MSE RF 92.50 92.85 92.50 0.92 0.37 SVM 97.99 98.08 97.99 0.98 0.08 MLP 98.72 98.73 98.72 0.99 0.04

[0114] As can be seen from Table 1, the MLP model performs optimally in terms of accuracy, mean squared error, precision, recall, F1 score, and mean squared error, proving that in the BCL framework, the prediction of MLP is more accurate, the model is more precise, and the accuracy can reach 98.72%. The performance of SVM is second only to MLP, and the overall performance of RF is relatively poor. Generally speaking, the accuracy, precision, and recall of the three models in the BCL framework can all reach at least 92%, proving that the BCL framework can have good performance in multi-classification tasks based on imbalanced samples.

[0115] The specificity comparison of the three classification models for each H&Y score classification is shown in Table 2:

[0116] Model Class 0 Class 1 Class 2 Class 3 Class 4 Average RF 0.74 1.00 1.00 1.00 1.00 0.95 SVM 0.99 1.00 0.98 1.00 1.00 0.99 MLP 0.98 1.00 0.99 1.00 1.00 0.99

[0117] As can be seen from Table 2, MLP and SVM have similar precision and the lowest misclassification rate, especially showing good performance in the recognition of class 0. Although it is slightly lower than RF in the recognition of class 2, the overall performance is significantly better; although RF performs well in other classifications, it has more misclassifications in class 0, and the overall performance is also slightly worse than the other two models.

[0118] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A Parkinson's progression prediction method based on the BCL model, characterized in that It includes the following steps: S1. Obtain clinical data including the H&Y score and basic patient information, as well as image data including MRI images; S2. Perform format conversion on the MRI images and perform data annotation in combination with the H&Y score; S3. Preprocess the images after data annotation; S4. Use principal component analysis to reduce the dimension of the preprocessed images and extract the low-dimensional features of the images; S5. Construct a curriculum learning framework based on a balanced weight scheduler, gradually introduce samples of different difficulties, construct different classification models and train the reduced-dimensional features, and obtain a curriculum learning model based on the balanced weight scheduler; S51. Divide the training set samples into 5 difficulties according to the H&Y score, and divide the training set after dividing the sample difficulties into 4 stages; S52. In the curriculum learning framework, introduce a balanced weight scheduler to adjust the weights of samples in different stages to obtain a curriculum learning framework based on the balanced weight scheduler; S53. Construct classification models including a random forest model, a support vector machine model, and a multi-layer perceptron model; S54. Train the model stage by stage. Before starting each stage, first obtain the number of categories and the number of samples in each category in the current stage training set. Secondly, call the balanced weight scheduler to calculate the category weights, generate a weight dictionary. Finally, call the classification models respectively for training, and pass the weight dictionary as the loss function into the classification models. Repeat the above training process until all four stages are trained to obtain a curriculum learning model based on the balanced weight scheduler; S6. Use the curriculum learning model based on the balanced weight scheduler to predict the progression of Parkinson's disease in patients and evaluate the prediction performance of each classification model.

2. The Parkinson's progression prediction method based on the BCL model according to claim 1, wherein The S2 includes the following steps: S21. Read the clinical data file containing the H&Y score and the image data file containing the basic information of the MRI images, integrate these two parts of data files into a new field, and store them in a formatted manner to achieve the mapping between the image ID and the H&Y score; S22. Use MRIcroGL to convert the DICOM format MRI images into NIFTI format, and save the basic metadata related to the images, including image size, pixel size, image orientation, and time point information; S23. For each NIfTI image, extract the corresponding image ID, and obtain the H&Y score corresponding to this image according to the mapping formed in S21, save the image path of this MRI and the corresponding H&Y score, and achieve data annotation for the MRI images.

3. The Parkinson's progression prediction method based on the BCL model according to claim 2, wherein, The S3 includes the following steps: Read the image path and convert the format to float32. For the image data that meets the dimension conditions, use OpenCV for resampling and normalization processing to standardize the images to the same resolution and size.

4. The Parkinson's progression prediction method based on the BCL model according to claim 3, wherein The resampling includes the following steps: Set the target affine matrix size to , resample each slice of each MRI image using the resize function of cv2, and return the resampled image object; The normalization processing includes the following steps: Calculate the mean and standard deviation of the pixel values of the input image. Subtract the image mean from each pixel value to eliminate the influence of overall brightness, and divide by the sum of the standard deviation and a smoothing term to prevent division-by-zero errors when the standard deviation is zero. Each pixel value after standardization represents the number of standard deviations relative to the original data distribution. The calculation formula is as follows: Wherein, is the image pixel value, is the pixel value of the input image, is the image mean, is the image standard deviation, is the smoothing term, is the total number of samples.

5. The Parkinson's progression prediction method based on the BCL model according to claim 4, wherein S4 includes the following steps: For all images of size flatten the image data into a one-dimensional vector; Use principal component analysis for feature extraction to achieve data dimensionality reduction. Re-integrate the extracted principal components into the original data frame to obtain a data set, and divide the data set into a training set and a test set. Each principal component serves as a new feature column.

6. The Parkinson's progression prediction method based on the BCL model according to claim 5, wherein The step of flattening the image data into a one-dimensional vector includes the following steps: For each image, straighten its pixel values to form a one-dimensional array, and combine all the one-dimensional arrays into a two-dimensional array with the shape: (number of samples, number of pixels), realizing the conversion of the high-dimensional image into a two-dimensional data matrix; The step of using principal component analysis for feature extraction includes the following steps: Use principal component analysis to reduce the flattened high-dimensional image data to 50 dimensions, reducing the data dimension while retaining most of the information.

7. The Parkinson's progression prediction method based on the BCL model according to claim 6, characterized in that, The 4 stages of the division of the training set include: The first stage: only includes the control group samples with a score of 0 and the easiest samples with a score of 4; The second stage: on the basis of the first stage, add the relatively easy samples with a score of 3; The third stage: on the basis of the second stage, add the relatively difficult samples with a score of 2; The fourth stage: on the basis of the third stage, add the most difficult samples with a score of 1.

8. The Parkinson's progression prediction method based on the BCL model according to claim 7, wherein S6 includes the following steps: Apply the trained curriculum learning model based on the balanced weight scheduler to the test set to perform classification prediction on the test samples. Calculate for each sample in the test set according to the input features, and the output result is the predicted class label of each test sample. The classes include five levels: 0, 1, 2, 3, and 4. The predicted class label corresponds to the H&Y score of the patient's MRI image, which is used to reflect the current stage of Parkinson's disease of the patient.

9. The Parkinson's disease progression prediction method based on the BCL model according to claim 8, wherein, The evaluation metrics for evaluating the prediction performance of each classification model include accuracy, precision, recall, F1 score, mean squared error, and specificity.

Images