A method and system for constructing an accurate identification model for early Parkinson's disease

Abstract

The present invention relates to the field of Parkinson's disease screening, and specifically to a method and system for constructing an early Parkinson's disease accurate identification model, comprising the following steps: S1 collecting motion data of humans or animals to draw motion trajectory diagrams and heat maps; S2 preprocessing the motion trajectory diagrams and heat maps to obtain standardized data sets; S3 constructing a feature extractor based on a twin DenseNet neural network, and constructing a classifier based on a multi-layer perceptron and Softmax, using the feature extractor to extract key features from the motion trajectory diagram and the heat map respectively, and then performing multi-source feature fusion on the extracted features, and using the fused features to classify using a predetermined classifier, thereby constructing an early Parkinson's disease accurate identification preliminary model; S4 training and tuning the preliminary model to obtain an early Parkinson's disease accurate identification model; S5 using test set data to test the model and complete early Parkinson's disease identification. This is to achieve large-scale promotion of Parkinson's disease screening and improve the accuracy and efficiency of screening.

Description

[Technical field]

[0001] The present invention relates to the field of Parkinson's disease screening, and in particular to a method and system for constructing an accurate identification model for early Parkinson's disease. [Background technology]

[0002] Parkinson's disease is a neurodegenerative disease. Although early screening can help protect nerves and develop personalized treatments, its hidden initial symptoms and complex pathophysiological mechanisms make traditional screening methods challenging. However, subtle changes in gait and motor behavior may reveal its early signs.

[0003] Traditional Parkinson's disease monitoring and screening mainly rely on hardware devices such as accelerometers, gyroscopes, magnetometers, and physiological and sound sensors. However, these devices have their limitations, such as incomplete data, susceptibility to external environmental interference, data offset in long-term monitoring, and data quality degradation. These limitations affect the effectiveness of sensors in Parkinson's disease screening and pose an obstacle to improving the accuracy and efficiency of screening. In addition, these data collection devices are not easy to popularize, making large-scale screening difficult to achieve.

[0004] In the field of Parkinson's disease research, motor behavior has important diagnostic value as a tool for evaluating movement characteristics and behavior patterns. This invention aims to explore the behavior and movement characteristics of patients through comprehensive analysis of the patient's movement trajectory and retention time heat map, and provide a reference for early screening. [Summary of the invention]

[0005] The present invention provides a method for constructing an accurate identification model for early Parkinson's disease, so as to solve the problems in the prior art that Parkinson's disease screening is difficult to promote on a large scale, and has poor screening accuracy and low efficiency.

[0006] The present invention is realized by the following technical scheme, and provides a method for constructing an early Parkinson's disease accurate identification model, comprising the following steps: S1, collecting motion data of humans or animals to draw motion trajectory diagrams and heat maps; S2, preprocessing the motion trajectory diagram and heat map in S1 to obtain a standardized data set, and then dividing the first collected data into a training set and a validation set according to a ratio of 6:1:3 for the standardized data set, and dividing the later collected data into a test set; S3, constructing a feature extractor based on a twin DenseNet neural network, and constructing a classifier based on a multi-layer perceptron and Softmax, using the feature extractor to extract key features of the motion trajectory diagram and the heat map respectively, and then performing multi-source feature fusion on the extracted key features through a fusion module, and then using the fused features as a basis, using the classifier to perform accurate classification, so as to construct a preliminary model for accurate identification of early Parkinson's disease; S4, using the training set and validation set in S2 to train and tune the preliminary model, and obtain an early Parkinson's disease accurate identification model; S5, using the test set data to test the model, completing the early Parkinson's disease identification model test, and calculating the model performance.

[0007] Furthermore, the human motion data collected in S1 is the data of healthy people and Parkinson's patients, which is specifically implemented in the following steps: S11, demarcate an experimental area of ​​5m×5m; S12, use a calibrated camera to record the daily activities of healthy people and Parkinson's patients in the experimental area for 5 minutes, and use a single target tracking algorithm to track the daily activities of healthy people and Parkinson's patients in real time; S13, based on the conversion matrix obtained after camera calibration, use a Python program to convert the pixel coordinates of the video taken by the camera into actual physical coordinates; S14, sort the physical coordinates according to the time sequence of the video to form a motion trajectory diagram, and count the residence time at different locations to generate a heat map.

[0008] The data of healthy people and Parkinson's patients can be collected by collecting location information through GPS or Beidou positioning, and then extracting the location data using Python. It can also be located through WiFi or Bluetooth and the location information can be obtained using the three-sided positioning algorithm.

[0009] Furthermore, the animal data collected in S1 are the data of healthy mice and early Parkinson's mice. The specific process is: placing the early Parkinson's disease mice induced by injection of drugs and healthy mice in an experimental box for 5 minutes, using animal behavior analysis software to automatically record the activities of the mice in the experimental box, and obtain the movement trajectory diagrams of the early Parkinson's disease mice and healthy mice and the corresponding heat maps.

[0010] Furthermore, the preprocessing of the motion trajectory map in S2 includes: cropping, scaling and denoising. The specific process is as follows: first, a target pixel set containing specific RGB values ​​is constructed, and the specific RGB values ​​represent the color features of interest. Then, each pixel in the motion trajectory map is screened to check whether its color matches any one of the target pixel set. According to the screening result, the pixels that do not belong to the color features of interest are replaced with interpolated pixels through bilinear interpolation to achieve denoising. Finally, the bounding box of the retained area is calculated, and these areas are cropped out from the original motion trajectory map.

[0011] The preprocessing of the heat map in S2 includes denoising, cropping, scaling and time chromatogram mapping. The cropping of the heat map needs to be set before scaling and time chromatogram mapping.

[0012] Gaussian smoothing is used to denoise the heat map. The specific process is as follows: for each pixel value I(x, y) on the original image, after smoothing, the new pixel value I'(x, y) is calculated by the following formula:

[0013]

[0014] Among them, G(i,j) is the value of the Gaussian filter at position (i,j), that is, k is half the size of the filter, σ is the standard deviation of the Gaussian function, which determines the degree of filtering. I(xi,yj)·G(i,j) calculates the product of the pixel value at a certain position in the original image and the corresponding Gaussian kernel. I'(x,y) is the sum of all these products in the sliding window, which achieves the purpose of fusing the neighborhood pixel values ​​according to the weight of the Gaussian function.

[0015] The specific process of cropping and scaling the heat map in S2 is to crop the heat map to remove redundant areas and scale it to a size of 224×224 pixels.

[0016] Furthermore, S2 uses a heatmap normalization method based on Manhattan distance for the temporal chromatogram mapping of the heatmap, and the steps are as follows:

[0017] Step 1, calculate the t value and s value of each pixel in the thermal map, t is the local residence time represented by each pixel, and s is the sequence number of the standardized pixel in the time mapping chromatogram;

[0018] Step 2, according to the s value, obtain the RGB value closest to the calculated RGB value from the time mapping color spectrum;

[0019] Step 3, overwrite the pixel value in the original heat map with the minimum RGB value of the Manhattan distance;

[0020] distance=|R1-R2|+|G1-G2|+|B1-B2| Formula 1-1

[0021] The Manhattan distance between two pixels in the three-dimensional RGB color space is calculated by formula 1-1. The smaller the color distance, the more similar the colors of the two pixels are. R1 and R2 represent the red channel values ​​of the first pixel and the second pixel respectively, G1 and G2 represent the green channel values ​​of the first pixel and the second pixel respectively, B1 and B2 represent the blue channel values ​​of the first pixel and the second pixel respectively, and the value ranges of R1, R2, G1, G2, B1 and B2 are all 0 to 255;

[0022]

[0023] Formula 1-2 is used to calculate the local residence time of each pixel, where T represents the maximum local residence time, i represents the serial number of each pixel in the time mapping color spectrum, and N represents the total number of RGB values ​​in the time mapping color spectrum;

[0024]

[0025] Formula 1-3 is used to map the local residence time to the normalized value of the heat map, T total It is the longest residence time of all the measured parts.

[0026] Furthermore, the feature extractor in S3 includes two parallel connected DenseNet networks with the same structure. Each layer of the DenseNet network is directly connected to all subsequent layers to ensure maximum information flow. The DenseNet network starts with 1 initial convolution layer, followed by 4 dense blocks and 3 transition layers. The initial convolution layer has 64 7×7 filters with a step size of 2. The dense block consists of multiple dense layers, each dense layer includes a 1×1 convolution and a 3×3 convolution. The dense block configuration is (6, 12, 24, 16). The transition layer is used to reduce the depth of the feature map, which includes 1 1×1 convolution layer and 1 2×2 average pooling layer. After the last dense block, there is a batch normalization layer for batch normalization. There is a batch normalization operation after each convolution layer. The global average pooling layer is after the last dense block, followed by the output layer for classification.

[0027] Furthermore, the classifier in S3 includes: a linear layer and a Softmax activation function, the linear layer is used to map the output of the feedforward neural network to the category space, and the Softmax activation function is applied to obtain the probability distribution of each category.

[0028] Furthermore, the fusion module in S3 uses the feature splicing method to achieve multi-source feature fusion. The process is that the two feature vectors output by the twin DenseNet network and Where d1 and d2 represent the dimensions of two feature vectors respectively, and a new feature vector is obtained by feature concatenation Its definition is as follows: concat =[F1, F2], where [F1, F2] means connecting the feature vectors F1 and F2 end to end.

[0029] Furthermore, the specific steps of S4 are: passing the predicted output generated by the preliminary model for accurate identification of early Parkinson's disease and the actual classification label to the loss function, and then adjusting the parameters of the preliminary model for accurate identification of early Parkinson's disease through back propagation according to the loss function value to obtain an accurate identification model for early Parkinson's disease. The Adam optimizer is used in the training process, and the learning rate is dynamically adjusted using exponential decay.

[0030] The present invention provides an early Parkinson's disease intelligent screening system based on behavioral analysis, comprising the following modules: a data acquisition module for acquiring human or animal data; a data conversion module for converting the acquired human or animal data into motion trajectory diagrams and thermal maps; a data processing module for preprocessing the motion trajectory diagrams and thermal maps to obtain standardized data sets; an extractor module for extracting features from the motion trajectory diagrams and thermal maps; a fusion module for fusing multi-source features of the extracted features; and a classifier module for classifying the fused features.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) By comprehensively analyzing the patient's movement trajectory and residence time heat map, we can deeply explore the patient's behavior and movement characteristics and provide a reference for early screening;

[0033] (2) By using the heat map normalization method based on Manhattan distance through time chromatogram mapping, the time chromatogram mapping and heat map normalization can be achieved to reduce the impact of deviations and outliers in the original data of the heat map on the overall;

[0034] (3) Building a feature extractor based on the twin DenseNet neural network. Through two parallel DenseNet networks, the dense connection characteristics of DenseNet are fully utilized to capture richer Parkinson's related features and patterns;

[0035] (4) Based on the accurate identification model and system of early Parkinson's disease, the software can be integrated into a variety of portable devices such as smart phones to support the needs of various application scenarios. It has high portability, low economic cost and good practicality, which is easy to promote and popularize, and is suitable for large-scale screening of early Parkinson's disease.

Brief Description of the Drawings

[0036] Figure 1 This is a flow chart of a method for constructing an accurate identification model for early Parkinson's disease according to the present invention;

[0037] Figure 2 This is a screening model architecture diagram of a method for constructing an accurate identification model for early Parkinson's disease according to the present invention;

[0038] Figure 3 A schematic diagram of the DenseNet dense block DenseBlock structure of a method for building an early Parkinson's disease accurate identification model of the present invention;

[0039] Figure 4 A schematic diagram of the experimental data collection and processing flow based on computer vision and target tracking of a method for constructing an accurate identification model for early Parkinson's disease of the present invention;

[0040] Figure 5 This is a time-mapped chromatogram of a method for constructing an accurate identification model for early Parkinson's disease according to the present invention. [Specific implementation method]

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0042] See also Figure 1-Figure 5 The present invention provides a method for constructing an accurate identification model for early Parkinson's disease, which comprises the following specific steps:

[0043] S1, collects motion data of humans or animals, and draws motion trajectory diagrams and thermal maps;

[0044] Specifically, collecting human data means collecting data of healthy people and Parkinson's disease patients, which is implemented in the following steps:

[0045] S11, delineate the experimental area of ​​5 m × 5 m;

[0046] S12, using the calibrated camera to record the daily activities of healthy people and Parkinson's disease patients in the experimental area for 5 minutes, where the camera can be a high-resolution, high-frame-rate camera such as Logitech C920, and using a single target tracking algorithm to track the daily activities of healthy people and Parkinson's disease patients in real time, where the single target tracking algorithms such as CSRT, KCF, MIL, GOTURN and DaSiamRPN of the OpenCV library can be used, and the position of the target in the next frame can be predicted by analyzing the changes of the target in consecutive frames, so as to achieve real-time single target tracking;

[0047] S13, based on the conversion matrix obtained after camera calibration, uses Python program to convert the pixel coordinates of the video captured by the camera into actual physical coordinates. The camera calibration and spatial position mapping are both based on OpenCV.

[0048] In this embodiment, the camera calibration is specifically as follows: fix the camera at one corner of the experimental area so that the camera overlooks the entire experimental area, ensure that the entire experimental area is within the camera range, and place calibration objects at the four corners of the experimental area for camera calibration. Calibration of the camera can achieve accurate position conversion;

[0049] In this embodiment, data of healthy people and Parkinson's disease patients are collected. The location information can be collected through GPS or Beidou positioning, and then the location data can be extracted using Python. The location information can also be obtained through WiFi or Bluetooth positioning and the three-side positioning algorithm.

[0050] S14, converting the timestamp into a time period, sorting the location information according to the timestamp to form a motion trajectory map, and counting the stay time at different locations to generate a heat map.

[0051] Specifically, data from healthy mice and early Parkinson's mice were collected by the following steps:

[0052] Mice with early Parkinson's disease induced by injection of 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP) and healthy mice were placed in an experimental box for 5 minutes. The activities of the mice in the experimental box were automatically recorded using animal behavior analysis software to obtain the movement trajectory diagrams and corresponding thermal maps of the mice with early Parkinson's disease and healthy mice. Specifically, behavioral analysis equipment such as 0.5m×0.5m experimental boxes from Coulbourn Instruments or San Diego Instruments, animal behavior analysis software such as ANY-maze developed by Stoelting or EthoVision XT developed by Noldus Information Technology, high-definition cameras and computers were used for data collection. The activities of the mice in the experimental box were captured by the camera, and the corresponding thermal maps were generated to facilitate the analysis of the dynamic behavior of the target.

[0053] S2, preprocess the motion trajectory map and thermal map collected by S1 to obtain a standardized data set, and then divide the standardized data set into a training set and a validation set according to the ratio of 6:1:3, and divide the data collected earlier into a test set;

[0054] Specifically, the preprocessing of the motion trajectory map includes: cropping, scaling and denoising. The specific process is to first construct a target pixel set containing specific RGB values, where the specific RGB values ​​represent the color features of interest. Then, each pixel in the motion trajectory map is screened to check whether its color matches any one of the target pixel sets. According to the screening results, the pixels that do not belong to the color features of interest are replaced with interpolated pixels through bilinear interpolation to achieve denoising. Finally, the bounding box of the retained area is calculated and these areas are cropped out from the original motion trajectory map.

[0055] The preprocessing of the heat map includes denoising, cropping, scaling and time chromatogram mapping. The cropping of the heat map needs to be set before scaling and time chromatogram mapping.

[0056] In this embodiment, the denoising method for the heat map is to use Gaussian smoothing, and the specific process is as follows:

[0057] For each pixel value I(x, y) on the original image, after smoothing, the new pixel value I'(x, y) is calculated by the following formula:

[0058]

[0059] Among them, G(i,j) is the value of the Gaussian filter at position (i,j), k is half the size of the filter, σ is the standard deviation of the Gaussian function, and determines the degree of filtering. I(xi,yj)·G(i,j) calculates the product of the pixel value at a certain position in the original image and the corresponding Gaussian kernel. I'(x,y) is the sum of all these products in the sliding window, which achieves the purpose of fusing the neighborhood pixel values ​​according to the weight of the Gaussian function.

[0060] The scaling and cropping process of the heat map is to crop the heat map to remove redundant areas and scale it to a size of 224×224 pixels to meet the input requirements of the DenseNet network.

[0061] Specifically, the heat map normalization method based on Manhattan distance is used for the time chromatogram mapping of the heat map. The heat map normalization is achieved through time chromatogram mapping to reduce the influence of deviations and outliers in the original data of the heat map on the overall. The steps are as follows:

[0062] Step 1, calculate the t value and s value of each pixel in the thermal map, t is the local residence time represented by each pixel, and s is the sequence number of the standardized pixel in the time mapping chromatogram;

[0063] Step 2, according to the s value, obtain the RGB value closest to the calculated RGB value from the time mapping color spectrum;

[0064] Step 3, overwrite the pixel value in the original heat map with the minimum RGB value of the Manhattan distance;

[0065] distance=|R1-R2|+|G1-G2|+|B1-B2| Formula 1-1

[0066] The Manhattan distance between two pixels in the three-dimensional RGB color space is calculated by formula 1-1. The smaller the color distance, the more similar the colors of the two pixels are. R1 and R2 represent the red channel values ​​of the first pixel and the second pixel respectively, G1 and G2 represent the green channel values ​​of the first pixel and the second pixel respectively, B1 and B2 represent the blue channel values ​​of the first pixel and the second pixel respectively, and the value range of R1, R2, G1, G2, B1 and B2 is 0 to 255 (8-bit integer);

[0067]

[0068] Formula 1-2 is used to calculate the local residence time t of each pixel, where T represents the maximum local residence time, i represents the serial number of each pixel in the time mapping color spectrum, and N represents the total number of RGB values ​​in the time mapping color spectrum;

[0069]

[0070] Formula 1-3 is used to map the local residence time to the standardized value of the heat map. s represents the serial number of the standardized pixel in the time mapping color spectrum, which is used for the pixel representation of the heat map. N is the total number of RGB values ​​in the time mapping color spectrum. T total It is the longest residence time of all the measured parts.

[0071] S3, builds a feature extractor based on the twin DenseNet neural network, and builds a classifier based on the multi-layer perceptron (MLP) and Softmax modules. The feature extractor is used to extract key features from the motion trajectory map and the heat map respectively, and then the extracted key features are fused with multi-source features through the fusion module. Then, the classifier is used to perform accurate classification based on the fused features, thereby building a preliminary model for accurate identification of early Parkinson's disease;

[0072] Specifically, Figure 2 , Figure 3 The feature extractor shown in the figure includes two parallel connected DenseNet networks with the same structure. The DenseNet network starts with an initial convolutional layer (Y[i, j] = ∑ m ∑ nX[im,jn]·K[m,n]), with 64 7×7 filters with a stride of 2, followed by 4 dense blocks and 3 transition layers. Each dense block consists of multiple dense layers, each of which includes a 1×1 convolution (bottleneck layer Y[i,j,k]=∑ l F[i, j, l] · W[l, k]) and a 3×3 convolution, the dense block configuration of DenseNet is (6, 12, 24, 16). The transition layer includes a 1×1 convolution layer and a 2×2 average pooling layer It is used to reduce the number of channels by half. After the last dense block, there is a batch normalization layer for batch normalization ( μ and σ 2 are the mean and variance of the input X, γ and β are learnable scaling and offset parameters, and ∈ is a small constant). Each convolutional layer is followed by a batch normalization operation. A global average pooling layer follows the last dense block, and finally the output layer is used for classification.

[0073] The feature extractor structure uses two parallel DenseNet networks, making full use of the dense connection characteristics of DenseNet to capture richer features and patterns related to Parkinson's disease.

[0074] Specifically, the classifier includes a linear layer and a Softmax activation function. The linear layer maps the output of the feedforward neural network to the category space, and the Softmax activation function is applied to obtain the probability distribution of each category. y i is the i-th element Softmax(y) output by the linear layer i is the probability of the ith category.

[0075] Specifically, the fusion module uses a feature concatenation method to achieve multi-source feature fusion. The feature vectors generated by the twin networks are connected end to end along the feature dimension to form a larger comprehensive feature vector that contains the feature information of the two networks.

[0076] Two feature vectors output by the twin DenseNet network and Where d1 and d2 represent the dimensions of the two feature vectors. A new feature vector is obtained by feature concatenation Its definition is as follows: concat = [F1, F2], where [F1, F2] means connecting the feature vectors F1 and F2 end to end. This provides a richer feature representation for subsequent classification or regression tasks while retaining the information in the original feature space.

[0077] S4, use the training set and validation set in S2 to train and tune the preliminary model, complete the early Parkinson's recognition model test, and calculate the model performance.

[0078] Specifically, the specific steps for training the preliminary model are as follows:

[0079] The predicted output and the actual classification label generated by the model are passed to the loss function, and then the model parameters are adjusted through back propagation according to the value of the loss function. The calculation expression of the loss function is Where y is the true label, is the predicted output of the model. On this basis, the loss function with the L2 regularization term added can be expressed as: in, is the total loss function with L2 regularization. W represents the weight parameters of the model, excluding the class weight parameters. λ is the regularization coefficient. n is the number of weight parameters. is the square of the L2 norm of the i-th weight parameter. The Adam optimizer is used during training, and the learning rate is dynamically adjusted using exponential decay.

[0080] In this embodiment, the update rule of Adam is to calculate the first and second order moments of the gradient: t =β1·m t-1 +(1-β1)·g t , m t and v t are estimates of the first and second moments of the gradient, g t is the current gradient, β1 and β2 are hyperparameters. Calculate the deviation of the corrected first-order moment and second-order moment: Update parameters: Among them, α is the learning rate, ∈ is a small constant, and the learning rate is dynamically adjusted using exponential decay: α t =α0·e -0.95t , where α t is the learning rate at time step t, α0 is the initial learning rate, set to 0.0001, and 0.95 is the decay rate.

[0081] Experimental process:

[0082] In this embodiment, based on the data of 1,600 healthy mice and Parkinson's model mice, six network models, DenseNet, VGG, ResNet, MobileNet, Inception, and AlexNet, were trained using motion trajectory graphs and heat maps, respectively, and a trajectory-heat map fusion model was built based on the DenseNet with better results.

[0083] In the experiment, the accuracy (ACC), sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and F1-score evaluation indicators are used to evaluate the model performance. The value range of AUC is between 0 and 1. The closer the AUC value is to 1, the better the performance of the classification model. The closer the AUC value is to 0.5, the performance of the classification model is equivalent to random guessing. The AUC value is lower than 0.5, which means that the performance of the classification model is poor. The calculation formulas for each indicator are as follows:

[0084] TP (True Positive): True positive number

[0085] TN (True Negative): True negative number

[0086] FP (False Positive): number of false positives

[0087] FN (False Negative): False negative number

[0088]

[0089] in:

[0090] The model was trained on the data collected from animal experiments and evaluated on the test set. In the experiment based on motion trajectory map training, the test set AUCs of the six models of DenseNet, VGG, ResNet, MobileNet, Inception, and AlexNet were 0.9296, 0.9185, 0.8674, 0.9178, 0.8809, and 0.9177, respectively. In the experiment based on heat map training, the AUCs of the six models were 0.9656, 0.9467, 0.9642, 0.9565, 0.9572, and 0.9589, respectively. The experimental results of multi-source feature fusion based on DenseNet are shown in Table 1. The experimental results show that accurate identification of early Parkinson's disease can be achieved based on motion trajectory map, heat map and their fusion.

[0091] Table 1. Performance of the DenseNet model with multi-source data fusion of movement trajectory diagram and heat map of early Parkinson's model mice

[0092]

[0093] The present invention also provides a system for building an accurate identification model for early Parkinson's disease. The system is built based on a method for building an accurate identification model for early Parkinson's disease and includes the following modules: a data acquisition module for collecting human or animal data; a data conversion module for converting the collected human or animal data into motion trajectory maps and heat maps; a data processing module for preprocessing the motion trajectory maps and heat maps to obtain standardized data sets; an extractor module for extracting features from the motion trajectory maps and heat maps; a fusion module for fusing multi-source features of the extracted features; and a classifier module for classifying the fused features.

[0094] The present invention is based on a variety of real-time positioning technologies and can be integrated into a variety of portable devices such as smart phones after being made into software, so as to support the needs of various application scenarios. It has good portability, low economic cost, good practicality and is easy to promote and popularize. It is suitable for large-scale screening of early Parkinson's disease.

Claims

1. A method for constructing an accurate identification model for early Parkinson's disease, characterized in that: The following steps are involved: S1, collects motion data of humans or animals to draw motion trajectory diagrams and heat maps; S2, preprocessing the motion trajectory map and the heat map in S1 to obtain a standardized data set, and then dividing the first collected data into a training set and a validation set according to a ratio of 6:1:3, and dividing the later collected data into a test set; S3, constructing a feature extractor based on the twin DenseNet neural network, and constructing a classifier based on the multi-layer perceptron and Softmax, using the feature extractor to extract key features from the motion trajectory map and the heat map respectively, and then fusing the extracted key features through a fusion module to perform multi-source feature fusion, and then using the fused features as the basis to perform classification using the classifier, thereby constructing a preliminary model for accurate identification of early Parkinson's disease; S4, using the training set and validation set in S2 to train and tune the preliminary model to obtain an accurate identification model for early Parkinson's disease; S5, use the test set data to test the model, complete the early Parkinson's recognition model test, and calculate the model performance.

2. The method for constructing an accurate identification model for early Parkinson's disease according to claim 1, characterized in that: The human motion data collected in S1 is the data collected from healthy people and Parkinson's disease patients, which is implemented in the following steps: S11, delineate the experimental area of ​​5 m × 5 m; S12, using the calibrated camera to record the daily activities of healthy people and Parkinson's disease patients in the experimental area for 5 minutes, and using a single target tracking algorithm to track the daily activities of the healthy people and Parkinson's disease patients in real time; S13, based on the conversion matrix obtained after the camera calibration, using a Python program to convert the pixel coordinates of the video captured by the camera into actual physical coordinates; S14, sorting the physical coordinates according to the time sequence of the video to form a motion trajectory diagram, and counting the stay time at different locations to generate a heat map.

3. The method for constructing an accurate identification model for early Parkinson's disease according to claim 2, characterized in that: The animal data collected in S1 are the data of healthy mice and early Parkinson's mice. The specific process is: the early Parkinson's mice induced by injection of drugs and healthy mice are placed in an experimental box for 5 minutes, and the animal behavior analysis software is used to automatically record the activities of the mice in the experimental box, and the movement trajectory diagrams of the early Parkinson's mice and healthy mice and the corresponding heat maps are obtained.

4. The method for constructing an accurate identification model for early Parkinson's disease according to claim 3, characterized in that: The preprocessing of the motion trajectory map in S2 includes: cropping, scaling and denoising. The specific process is to first construct a target pixel set containing specific RGB values, where the specific RGB values ​​represent the color features of interest, then screen each pixel in the motion trajectory map to check whether its color matches any one of the target pixel set, and replace the pixels that do not belong to the color features of interest with interpolated pixels through bilinear interpolation according to the screening results to achieve denoising, and finally calculate the bounding box of the retained area and crop these areas from the original motion trajectory map; The preprocessing of the heat map in S2 includes denoising, cropping, scaling and time chromatogram mapping, and the cropping of the heat map needs to be set before scaling and time chromatogram mapping; Gaussian smoothing is used to denoise the heat map. The specific process is as follows: for each pixel value I(x, y) on the original image, after smoothing, the new pixel value I′(x, y) is calculated by the following formula: Among them, G(i, j) is the value of the Gaussian filter at position (i, j), that is, k is half the size of the filter, σ is the standard deviation of the Gaussian function, which determines the degree of filtering. I(xi, yj)·G(i, j) calculates the product of the pixel value at a certain position in the original image and the corresponding Gaussian kernel. I′(x, y) is the sum of all these products in the sliding window, which is used to realize the fusion of neighborhood pixel values ​​according to the weight of the Gaussian function. The specific process of cropping and scaling the heat map in S2 is to crop the heat map to remove redundant areas and scale it to a size of 224×224 pixels.

5. The method for constructing an accurate identification model for early Parkinson's disease according to claim 4, characterized in that: In S2, the time-chromatographic mapping of the heat map uses a heat map normalization method based on Manhattan distance. The steps are as follows: Step 1, calculate the t value and s value of each pixel in the thermal map, t is the local residence time represented by each pixel, and s is the sequence number of the standardized pixel in the time mapping chromatogram; Step 2, obtaining the RGB value closest to the RGB value from the time mapping color spectrum according to the s value; Step 3, overwrite the pixel value in the original heat map with the minimum RGB value of the Manhattan distance; distance=|R1-R2|+|G1-G2|+|B1-B2| Formula 1-1 The Manhattan distance between two pixels in the three-dimensional RGB color space is calculated by formula 1-1. The smaller the color distance, the more similar the colors of the two pixels are, wherein R1 and R2 represent the red channel values ​​of the first pixel and the second pixel respectively, G1 and G2 represent the green channel values ​​of the first pixel and the second pixel respectively, B1 and B2 represent the blue channel values ​​of the first pixel and the second pixel respectively, and the value ranges of R1, R2, G1, G2, B1 and B2 are all 0 to 255; Formula 1-2 is used to calculate the local residence time of each pixel, where T represents the maximum local residence time, i represents the serial number of each pixel in the time mapping color spectrum, and N represents the total number of RGB values ​​in the time mapping color spectrum; Formula 1-3 is used to map the local residence time to the normalized value of the heat map. T total It is the longest residence time of all the measured parts.

6. The method for constructing an accurate identification model for early Parkinson's disease according to claim 5, characterized in that: The feature extractor in S3 includes two parallel connected DenseNet networks with the same structure. Each layer of the DenseNet network is directly connected to all subsequent layers to ensure maximum information flow. The DenseNet network starts with an initial convolution layer, followed by 4 dense blocks and 3 transition layers. The initial convolution layer has 64 7×7 filters with a step size of 2. The dense block consists of multiple dense layers, each of which includes a 1×1 convolution and a 3×3 convolution. The dense blocks are configured as (6, 12, 24, 16). The transition layer is used to reduce the depth of the feature map, which includes a 1×1 convolution layer and a 2×2 average pooling layer. After the last dense block, there is a batch normalization layer for batch normalization processing. There is a batch normalization operation after each convolution layer. The global average pooling layer is after the last dense block, followed by the output layer for classification.

7. The method for constructing an accurate identification model for early Parkinson's disease according to claim 6, characterized in that: The classifier in S3 includes: a linear layer and a Softmax activation function, wherein the linear layer is used to map the output of the feedforward neural network to the category space, and the Softmax activation function is applied to obtain the probability distribution of each category.

8. The method for constructing an accurate identification model for early Parkinson's disease according to claim 7, characterized in that: The fusion module described in S3 uses the feature splicing method to achieve multi-source feature fusion. The process is that the two feature vectors output by the twin DenseNet network and Where d1 and d2 represent the dimensions of two feature vectors respectively, and a new feature vector is obtained by feature concatenation Its definition is as follows: concat =[F1, F2], where [F1, F2] means connecting the feature vectors F1 and F2 end to end.

9. The method for constructing an accurate identification model for early Parkinson's disease according to claim 8, characterized in that: The specific step of S4 is: passing the predicted output generated by the preliminary model for accurate identification of early Parkinson's disease and the actual classification label to the loss function, and then adjusting the parameters of the preliminary model for accurate identification of early Parkinson's disease through back propagation according to the loss function value to obtain the precise identification model for early Parkinson's disease.

10. A system for constructing an accurate identification model for early Parkinson's disease, which is used to implement the method for constructing an accurate identification model for early Parkinson's disease according to any one of claims 1 to 9, characterized in that: The invention comprises the following modules: a data acquisition module for collecting human or animal data; a data conversion module for converting the collected human or animal data into motion trajectory diagrams and thermal maps; a data processing module for preprocessing the motion trajectory diagrams and thermal maps to obtain standardized data sets; an extractor module for extracting features from the motion trajectory diagrams and thermal maps; a fusion module for fusing multi-source features of the extracted features; and a classifier module for classifying the fused features.

Images