Neurosurgery tumor patient health monitoring system based on machine learning

Abstract

The invention belongs to the technical field of assembly calibration, and particularly discloses a neurosurgery tumor patient health monitoring system based on machine learning, which comprises a patient state acquisition module, a two-dimensional medical image feature extraction module, a three-dimensional rehabilitation video feature extraction module and a state monitoring module. According to the scheme, the grey wolf optimization algorithm is adopted to screen the feature set of the two-dimensional medical image, the gating circulation unit is adopted to screen the features from the three-dimensional medical video according to the matching points between the three-dimensional medical video and the two-dimensional medical image and the key points related to the human body motion, and the training efficiency of the classifier is improved; according to the method, the range of vital sign data used for evaluating the health state of the tumor patient is expanded, vital sign data used for providing nursing decision support is decomposed into three dimensions, a feature extraction mode is refined, the aims of checking the nursing effect and dynamically monitoring the health state of the patient are achieved, and the nursing quality of the brain tumor patient is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health monitoring, and specifically refers to a health monitoring system for neurosurgical tumor patients based on machine learning. Background Art

[0002] A health monitoring system for neurosurgical tumor patients based on machine learning refers to a system that uses machine learning methods to monitor the health of tumor patients.

[0003] In existing similar solutions, for example, CN118507074B is an intelligent nursing system and method for tumor patients during radiotherapy and chemotherapy. This solution addresses the technical problems of subjectivity and easy missed diagnosis in the traditional manual nursing method for tumor patients during radiotherapy and chemotherapy. It uses an embedded encoder based on a fully connected layer to extract the global semantic and prominent semantic features of vital sign data, fuses the two features, and uses a classifier to analyze the feature vector obtained after fusion to determine whether an adverse reaction occurs, achieving the technical effect of automatically monitoring adverse reactions and assisting medical staff in adjusting the corresponding radiotherapy and chemotherapy plan. However, it has the technical problems of a large scale of patient state features and a large amount of feature data extracted by the encoder, difficult processing by the classifier, and long training time.

[0004] Traditional medical image analysis methods can provide nursing support for brain tumor patients. CN118657756B is an intelligent auxiliary decision-making system and method for the nursing of brain tumor patients. This solution addresses the problem of difficult to ensure the accuracy of segmentation for complex brain tumor images. It uses brain tumor image data such as MRI (Magnetic Resonance Imaging), CT (Computed Tomography), and PET (Positron Emission Tomography) as the basis for judging the patient's recovery status to provide decision-making support for nursing, and uses the gray-level co-occurrence matrix to evaluate the quality of the segmentation result, achieving the technical effect of increasing the accuracy of image segmentation. However, it has the technical problem of only using two-dimensional medical images such as MRI, CT, and PET as the basis for judging the patient's recovery status. Summary of the Invention

[0005] In view of the above situation, to overcome the defects of the prior art, the present invention provides a health monitoring system for neurosurgical oncology patients based on machine learning; aiming at the technical problems that the scale and data volume of the patient state features extracted by the encoder are large, the classifier is difficult to process, and the training time is long, this solution uses the Grey Wolf Optimization Algorithm to screen the feature set of two-dimensional medical images, and according to the matching points between the three-dimensional medical video and the two-dimensional medical images and the key points related to human movement, uses a gated recurrent unit to screen features from the three-dimensional medical video, reducing the number and dimension of features, shortening the training time, and improving the training efficiency of the classifier; aiming at the technical problem of using only two-dimensional medical images such as MRI, CT, and PET as the basis for judging the patient's rehabilitation status, this solution expands the scope of vital sign data for evaluating the health status of oncology patients, decomposes the vital sign data for providing nursing decision support into three dimensions, refines the feature extraction method, extracts body posture features from the rehabilitation exercise video of oncology patients, and realizes the goals of testing the nursing effect and dynamically monitoring the patient's health status by evaluating the patient's limb movement state, thus improving the nursing quality of brain tumor patients.

[0006] The technical solution adopted by the present invention is as follows: The present invention provides a health monitoring system for neurosurgical oncology patients based on machine learning, and the health monitoring system for neurosurgical oncology patients based on machine learning includes a patient state acquisition module, a two-dimensional medical image feature extraction module, a three-dimensional rehabilitation video feature extraction module, and a state monitoring module;

[0007] The patient state acquisition module is used to monitor and obtain the health status data of oncology patients, and the health status data includes one-dimensional data, two-dimensional medical images, and three-dimensional rehabilitation videos;

[0008] The two-dimensional medical image feature extraction module uses a two-dimensional image analysis method to preprocess two-dimensional medical images, extract an image spatial feature set, and screen the image spatial feature set;

[0009] The three-dimensional rehabilitation video feature extraction module uses a three-dimensional video analysis method to extract the video body posture feature set of oncology patients and screen the video body posture feature set;

[0010] The state monitoring module uses a Transformers deep learning model as a classifier to capture the relationship between one-dimensional data, the image spatial feature set, and the video body posture feature set, classify the health status of oncology patients, and obtain a health monitoring result, and the health monitoring result is used to indicate whether adverse reactions occur in oncology patients during the nursing stage.

[0011] The two-dimensional medical image feature extraction module uses a two-dimensional image analysis method, and the two-dimensional image analysis method specifically includes the following steps:

[0012] Step B1: Preprocessing and pre-training. Specifically, unify the size of two-dimensional medical images, perform denoising on two-dimensional medical images using wavelet transform, remove interfering objects in two-dimensional medical images through morphological operations, apply the k-means clustering algorithm to segment the lesion area in two-dimensional medical images, and apply self-supervised learning on the lesion area to generate preliminary features;

[0013] Step B2: Extract the image spatial feature set;

[0014] Step B3: Two-dimensional feature screening, which is used to reduce the number and dimension of features. Specifically, use the grey wolf optimization algorithm to perform feature screening on the image spatial feature set.

[0015] Furthermore, in Step B2, the extraction of the image spatial feature set specifically includes the following steps:

[0016] Step B21: Calculate the grey level co-occurrence matrix, and the formula used is as follows: ;

[0017] In the formula, represents the grey level co-occurrence matrix, represents the pixel pairs with the same grey value in the two-dimensional medical image, represents the pixel pair appearing times in the distance and direction ; represents the total number of pixel pairs with any grey value

[0018] existing ;

[0019] In the formula, represents the correlation, represents the summation of all pixel pairs with the same grey value in the two-dimensional medical image ; represents the grey level co-occurrence matrix, represents the difference between the grey value of pixel and the grey mean value, represents the difference between the grey value of pixel and respectively represent the standard deviations of pixel and pixel ;

[0020] Step B23: Calculate the contrast, which is used to quantify the local variation of pixels. Specifically, calculate the weighted sum of the squared differences of gray values in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0021] In the formula, represents the contrast, represents the summation of all pixel pairs with the same gray value in the two-dimensional medical image for summation, represents the gray-level co-occurrence matrix, represents the pixel and the pixel is the difference in gray values;

[0022] Step B24: Calculate the homogeneity, which is used to reflect the similarity degree of adjacent pixels in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0023] In the formula, represents the homogeneity, represents the summation of all pixel pairs with the same gray value in the two-dimensional medical image for summation, represents the gray-level co-occurrence matrix, represents the pixel and the pixel is the difference in gray values;

[0024] Step B25: Calculate the angular second moment, which is used to measure the uniformity of pixel pairs in the two-dimensional medical image. Specifically, obtain the angular second moment by calculating the sum of the squares of each element in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0025] In the formula, represents the angular second moment, represents the summation of all pixel pairs with the same gray value in the two-dimensional medical image for summation, represents the gray-level co-occurrence matrix;

[0026] Step B26: Calculate the entropy, which is used to quantify the randomness of pixels in the two-dimensional medical image. The formula used is as follows: ;

[0027] In the formula, represents the entropy, represents the summation of all pixel pairs with the same gray value in the two-dimensional medical image for summation, represents the gray-level co-occurrence matrix, Represents a positive constant;

[0028] Step B27: Binary encoding, used to reflect the neighborhood relationship of a two-dimensional medical image. Specifically, the neighborhood binary method is adopted to obtain a histogram;

[0029] Step B28: Enrich the feature space. Specifically, the preliminary features, correlation, contrast, homogeneity, angular second moment, entropy, and histogram are combined and denoted as the image space feature set.

[0030] Further, in Step B27, the neighborhood binary method specifically includes the following steps:

[0031] Step B271: Obtain the neighborhood binary encoding. Specifically, with any pixel in the two-dimensional medical image as the center, in the circular sampling area with a radius of, collect sampling points, calculate the neighborhood binary operator, arrange the neighborhood binary operators of all sampling points in clockwise order, and represent them in binary to obtain the neighborhood binary encoding. The formula for the used neighborhood binary encoding is as follows: ;

[0032] In the formula, represents the neighborhood binary operator, represents the radius of the circular sampling area, represents the total number of sampling points in the circular sampling area, represents the cumulative sum of all sampling points, represents the pixel value of the center point of the circular sampling area, represents the th sampling point, represents the sampling point subscript, represents the step function, represents the neighborhood binary operator;

[0033] For the step function, when is greater than or equal to 0, the function value is 1, and when is less than 0, the function value is 0;

[0034] Step B272: Update the neighborhood binary encoding to enhance robustness. Specifically, perform circular shifting on the binary neighborhood binary encoding, and the number of shift times is , select the neighborhood binary encoding with the smallest value after circular shifting as the new neighborhood binary encoding, and the take values in sequence as ;

[0035] Step B273: Construct a histogram to reflect neighborhood features. Specifically, convert the neighborhood binary encoding into decimal. Use the neighborhood binary encoding as the abscissa and the number of times the neighborhood binary encoding appears in the two-dimensional medical image as the ordinate to construct the histogram.

[0036] The three-dimensional rehabilitation video feature extraction module adopts a three-dimensional video analysis method. The three-dimensional video analysis method specifically includes the following steps:

[0037] Step C1: Human key point tracking. Specifically, use the AlphaPose pose estimation technology. According to the three-dimensional rehabilitation video, predict the positions of human key points and the confidence scores of the key point positions in the three-dimensional rehabilitation video. Generate a corresponding heat map matrix for each key point. Determine the key point coordinates by finding the position of the maximum value in the heat map matrix.

[0038] The key points are specifically the joints and body parts of the human body.

[0039] Step C2: Key point weight assignment, which is used to assign weights to different body parts of the human body. Specifically, according to the significance degree of each key point's influence on the body posture characteristics, assign a significance weight to each key point.

[0040] Step C3: Multilevel smoothed body posture signal, which is used to capture and smooth the long-term and short-term changes of key points. Specifically, use the fast Fourier transform to analyze the frequency components of the key point change trajectory. Adopt a smoothing window to capture the key point position changes in a long time period, smooth the key point trajectory change trend, calculate the weighted exponential moving average, and use the weighted exponential moving average to smooth the key point coordinates in a short time period. The calculation formula of the weighted exponential moving average used is as follows: ;

[0041] In the formula, represents the weighted exponential moving average at time t, represents the significance weight, represents the smoothing factor, represents the key point coordinates at time t, represents the weighted exponential moving average at time t-1;

[0042] Initially, that is, when , ;

[0043] Step C4: Semantic matching, which is used to mine the semantic relationship between the two-dimensional medical image and the three-dimensional rehabilitation video to obtain matching points.

[0044] Step C5: Three-dimensional feature screening. Specifically, use a gated recurrent unit to screen the video body posture feature set from the three-dimensional rehabilitation video according to the key points and the matching points.

[0045] Further, in step C4, the semantic matching includes the following steps:

[0046] Step C41: Convolution extraction, specifically, using a traditional convolutional neural network to mine features in two-dimensional medical images and three-dimensional rehabilitation videos;

[0047] Step C42: Intermediate feature extraction, specifically, using a residual network to extract intermediate features from the convolutional layer of the traditional convolutional neural network, and using bilinear interpolation on the intermediate features to unify the spatial dimensions of the intermediate features;

[0048] Step C43: Feature slicing, specifically, slicing the intermediate features along the channel dimension into n groups of sub-features, calculating the cosine similarity of each group of sub-features to generate a single-channel correlation map, and stacking the single-channel correlation maps of all groups of sub-features along the channel dimension to obtain a multi-layer correlation map;

[0049] Step C44: Optimize the multi-layer correlation map, specifically, using a point convolution method to perform channel aggregation operation on the multi-layer correlation map to obtain a final correlation matrix;

[0050] Step C45: Matching point mapping, specifically, mapping the two-dimensional medical image and the three-dimensional rehabilitation video according to the final correlation matrix to obtain matching points between the two-dimensional medical image and the three-dimensional rehabilitation video.

[0051] The beneficial effects achieved by the present invention using the above solution are as follows:

[0052] (1) Aiming at the technical problems that the scale and amount of feature data of the patient's state features extracted by the encoder are large, the classifier is difficult to process, and the training time is long, this solution uses the grey wolf optimization algorithm to screen the feature set of two-dimensional medical images, and according to the matching points between the three-dimensional medical video and the two-dimensional medical image and the key points related to human movement, uses a gated recurrent unit to screen features from the three-dimensional medical video, reducing the number and dimension of features, shortening the training time, and improving the training efficiency of the classifier;

[0053] (2) Aiming at the technical problem of using only two-dimensional medical images such as MRI, CT, and PET as the basis for judging the patient's rehabilitation status, this solution expands the range of vital sign data used to evaluate the health status of tumor patients, decomposes the vital sign data used to provide nursing decision support into three dimensions, refines the feature extraction method, extracts body posture features from the rehabilitation exercise videos of tumor patients, and realizes the goal of testing the nursing effect and dynamically monitoring the patient's health status by evaluating the patient's limb movement status, improving the nursing quality of brain tumor patients. Description of the Drawings

[0054] Figure 1Module connection diagram of the health monitoring system for neurosurgical tumor patients based on machine learning provided by the present invention;

[0055] Figure 2 Schematic flow diagram of the two-dimensional image analysis method;

[0056] Figure 3 Schematic flow diagram of step B2;

[0057] Figure 4 Schematic flow diagram of step B27;

[0058] Figure 5 Schematic flow diagram of the three-dimensional video analysis method;

[0059] Figure 6 Schematic flow diagram of step C4.

[0060] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0062] Embodiment 1: Refer to Figures 1 to 6 , this embodiment provides a health monitoring system for neurosurgical tumor patients based on machine learning. The health monitoring system for neurosurgical tumor patients based on machine learning includes a patient status acquisition module, a two-dimensional medical image feature extraction module, a three-dimensional rehabilitation video feature extraction module, and a status monitoring module;

[0063] The patient status acquisition module is used to monitor and obtain the health status data of tumor patients. The health status data includes one-dimensional data, two-dimensional medical images, and three-dimensional rehabilitation videos;

[0064] The two-dimensional medical image feature extraction module uses a two-dimensional image analysis method to preprocess two-dimensional medical images, extract an image spatial feature set, and screen the image spatial feature set;

[0065] The three-dimensional rehabilitation video feature extraction module uses a three-dimensional video analysis method to extract the video body posture feature set of tumor patients and screen the video body posture feature set;

[0066] The state monitoring module uses a Transformers deep learning model as a classifier to capture the relationships between one-dimensional data, image spatial feature sets, and video body posture feature sets, classify the health status of tumor patients, and obtain health monitoring results, which are used to indicate whether adverse reactions occur in tumor patients during the nursing stage.

[0067] Example 2: Refer to Figures 1 to 3 , based on the above example, the two-dimensional medical image feature extraction module adopts a two-dimensional image analysis method, and the two-dimensional image analysis method specifically includes the following steps:

[0068] Step B1: Preprocessing and pre-training, specifically, unify the size of two-dimensional medical images, use wavelet transform to denoise two-dimensional medical images, remove interference objects in two-dimensional medical images through morphological operations, apply the k-means clustering algorithm to segment the lesion area in two-dimensional medical images, and apply self-supervised learning on the lesion area to generate preliminary features;

[0069] Step B2: Extract the image spatial feature set;

[0070] Step B3: Two-dimensional feature screening, used to reduce the number and dimension of features, specifically, use the grey wolf optimization algorithm to perform feature screening on the image spatial feature set.

[0071] Example 3: Refer to Figures 1 to 4 , based on the above example, in step B2, the extraction of the image spatial feature set specifically includes the following steps:

[0072] Step B21: Calculate the grey level co-occurrence matrix, and the formula used is as follows: ;

[0073] In the formula, represents the grey level co-occurrence matrix, represents the pixel pairs with the same grey level value in the two-dimensional medical image, represents the pixel pair at a distance and direction appearing times, represents at a distance and direction the total number of pixel pairs with any grey level value existing;

[0074] Step B22: Calculate the correlation, used to quantify the linear relationship between pixels, specifically, use the product of covariance divided by standard deviation to calculate the correlation, and the formula used is as follows: ;

[0075] In the formula, Indicates correlation Indicates summing over all pairs of pixels with the same gray value in a two-dimensional medical image For summation Indicates the gray-level co-occurrence matrix Indicates a pixel The difference between the gray value of the pixel and the mean gray value Indicates a pixel The difference between the gray value of the pixel and the mean gray value And Respectively indicate the standard deviations of pixel And pixel ;

[0076] Step B23: Calculate the contrast, which is used to quantify the local variation of pixels. Specifically, calculate the weighted sum of the squared differences of gray values in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0077] In the formula, Indicates the contrast Indicates summing over all pairs of pixels with the same gray value in a two-dimensional medical image For summation Indicates the gray-level co-occurrence matrix Indicates a pixel And pixel The difference in gray values;

[0078] Step B24: Calculate the homogeneity, which is used to reflect the similarity degree of adjacent pixels in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0079] In the formula, Indicates the homogeneity Indicates summing over all pairs of pixels with the same gray value in a two-dimensional medical image For summation Indicates the gray-level co-occurrence matrix Indicates a pixel And pixel The difference in gray values;

[0080] Step B25: Calculate the angular second moment, which is used to measure the uniformity of pixel pairs in a two-dimensional medical image. Specifically, obtain the angular second moment by calculating the sum of the squares of each element in the gray-level co-occurrence matrix. The formula used is as follows: ;

[0081] In the formula, Indicates the angular second moment Denote the sum of all pixel pairs with the same gray value in the two-dimensional medical image as the gray-level co-occurrence matrix;

[0082] Step B26: Calculate the entropy, which is used to quantify the randomness of pixels in the two-dimensional medical image. The formula used is as follows: ;

[0083] In the formula, denotes the entropy, denotes the sum of all pixel pairs with the same gray value in the two-dimensional medical image as the gray-level co-occurrence matrix, denotes a positive constant;

[0084] Step B27: Binary coding, which is used to reflect the neighborhood relationship of the two-dimensional medical image. Specifically, the neighborhood binary method is adopted to obtain a histogram;

[0085] Step B28: Enrich the feature space. Specifically, the preliminary features, correlation, contrast, homogeneity, angular second moment, entropy, and histogram are combined and denoted as the image space feature set.

[0086] Example 4: Refer to Figures 1 to 4 , this example is based on the above example. In step B27, the neighborhood binary method specifically includes the following steps:

[0087] Step B271: Obtain the neighborhood binary coding. Specifically, with any pixel in the two-dimensional medical image as the center, as the radius of the circular sampling area, collect sampling points, calculate the neighborhood binary operator, arrange the neighborhood binary operators of all sampling points in clockwise order, and represent them in binary to obtain the neighborhood binary coding. The formula for the neighborhood binary coding used is as follows: ;

[0088] In the formula, denotes the neighborhood binary operator, denotes the radius of the circular sampling area, denotes the total number of sampling points in the circular sampling area, denotes the cumulative sum of all sampling points, denotes the pixel value of the center point of the circular sampling area, denotes the th sampling point's pixel value, denotes the sampling point subscript, denotes the step function, denotes the neighborhood binary operator;

[0089] The step function, when greater than or equal to 0, the function value is 1, and when less than 0, the function value is 0;

[0090] Step B272: Update the neighborhood binary encoding for enhancing robustness. Specifically, perform circular shifting on the binary neighborhood binary encoding, and the number of displacement times is , select the neighborhood binary encoding with the smallest value after circular shifting as the new neighborhood binary encoding, and the take values in sequence as ;

[0091] Step B273: Construct a histogram to reflect neighborhood features. Specifically, convert the neighborhood binary encoding to decimal, use the neighborhood binary encoding as the abscissa, and use the number of times the neighborhood binary encoding appears in the two-dimensional medical image as the ordinate to construct a histogram.

[0092] Example 5: Refer to Figures 1 to 6 , this example is based on the above example. The three-dimensional rehabilitation video feature extraction module adopts a three-dimensional video analysis method, and the three-dimensional video analysis method specifically includes the following steps:

[0093] Step C1: Human key point tracking. Specifically, adopt the AlphaPose pose estimation technology, predict the positions of human key points and the confidence scores of key point positions in the three-dimensional rehabilitation video according to the three-dimensional rehabilitation video, generate a corresponding heat map matrix for each key point, and determine the key point coordinates by finding the position of the maximum value in the heat map matrix;

[0094] The key points are specifically the joints and body parts of the human body;

[0095] Step C2: Key point weight assignment, which is used to assign weights to different body parts of the human body. Specifically, assign significance weights to each key point according to the significance degree of each key point's influence on the body posture characteristics;

[0096] Step C3: Multilevel smoothed body posture signals, which are used to capture and smooth the long-term and short-term changes of key points. Specifically, use the fast Fourier transform to analyze the frequency components of the key point change trajectory, use a smoothing window to capture the key point position changes in a long time period, smooth the trajectory change trend of the key points, calculate the weighted exponential moving average, and use the weighted exponential moving average to smooth the key point coordinates in a short time period. The formula for the weighted exponential moving average used is as follows: ;

[0097] In the formula, represents the weighted exponential moving average at time t, represents the said significance weight, represents the smoothing factor, represents the key point coordinates at time t, represents the weighted exponentially moving average at time t - 1;

[0098] Initially, i.e., when time, ;

[0099] Step C4: Semantic matching, which is used to mine the semantic relationship between the two-dimensional medical image and the three-dimensional rehabilitation video to obtain matching points;

[0100] Step C5: Three-dimensional feature screening. Specifically, a gated recurrent unit is used to screen the video body posture feature set from the three-dimensional rehabilitation video according to the key points and the matching points.

[0101] Example Six: Refer to Figures 1 to 6 , this example is based on the above example. In step C4, the semantic matching includes the following steps:

[0102] Step C41: Convolution extraction. Specifically, a traditional convolutional neural network is used to mine the features in the two-dimensional medical image and the three-dimensional rehabilitation video;

[0103] Step C42: Intermediate feature extraction. Specifically, a residual network is used to extract intermediate features from the convolutional layer of the traditional convolutional neural network, and the bilinear interpolation method is used for the intermediate features to unify the spatial dimensions of the intermediate features;

[0104] Step C43: Feature slicing. Specifically, the intermediate features are sliced into n groups of sub-features along the channel dimension, the cosine similarity of each group of sub-features is calculated to generate a single-channel correlation map, and the single-channel correlation maps of all groups of sub-features are stacked along the channel dimension to obtain a multi-layer correlation map;

[0105] Step C44: Optimize the multi-layer correlation map. Specifically, a point convolution method is used to perform channel aggregation operation on the multi-layer correlation map to obtain a final correlation matrix;

[0106] Step C45: Matching point mapping. Specifically, the two-dimensional medical image and the three-dimensional rehabilitation video are mapped according to the final correlation matrix to obtain the matching points between the two-dimensional medical image and the three-dimensional rehabilitation video.

[0107] Example Seven: This example is based on the above example. In step B1, the interfering object specifically refers to all hairs.

[0108] Example Eight: This example is based on the above example. In step B26, .

[0109] Embodiment Nine: This embodiment is based on the above embodiment. In step C1, the human key points specifically refer to the tip of the nose, the left and right earlobes, the left and right eyebrows, the left and right eyes, the left and right corners of the mouth, the neck, the left and right shoulders, the left and right elbows, the left and right wrists, the left and right hips, the left and right knees, and the left and right ankles.

[0110] Embodiment Ten: This embodiment is based on the above embodiment. In step C3, the value range of the smoothing factor is between 0 and 1.

[0111] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0112] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

[0113] The above describes the present invention and its embodiments. Such a description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural modes and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A health monitoring system for neurosurgery tumor patients based on machine learning, characterized by: It includes a patient status acquisition module, a two-dimensional medical image feature extraction module, a three-dimensional rehabilitation video feature extraction module and a status monitoring module; The patient status acquisition module is used to monitor and obtain health status data of tumor patients, and the health status data includes one-dimensional data, two-dimensional medical images, and three-dimensional rehabilitation videos; The two-dimensional medical image feature extraction module uses a two-dimensional image analysis method to pre-process the two-dimensional medical image and extract and screen the image space feature set; The three-dimensional rehabilitation video feature extraction module uses a three-dimensional video analysis method to extract and filter a video posture feature set; The state monitoring module uses the Transformers deep learning model as a classifier to capture the relationship between one-dimensional data, image space feature sets and video posture feature sets to obtain health monitoring results.

2. The machine learning-based health monitoring system for neurosurgery tumor patients according to claim 1 is characterized in that: The two-dimensional medical image feature extraction module adopts a two-dimensional image analysis method, and the two-dimensional image analysis method specifically includes the following steps: Step B1: preprocessing and pretraining the two-dimensional medical image to generate preliminary features; Step B2: extracting image spatial feature sets; Step B3: Two-dimensional feature screening, specifically, using the Grey Wolf Optimization Algorithm to perform feature screening on the image space feature set.

3. The machine learning-based health monitoring system for neurosurgery tumor patients according to claim 2 is characterized in that: In step B2, the extraction of the image spatial feature set specifically includes the following steps: Step B21: Calculate the gray level co-occurrence matrix; Step B22: Calculate the correlation; Step B23: Calculate contrast; Step B24: Calculate homogeneity; Step B25: Calculate the angular second moment; Step B26: Calculate entropy; Step B27: binary encoding, used to reflect the neighborhood relationship of the two-dimensional medical image, specifically, using a neighborhood binary method to obtain a histogram; Step B28: enriching the feature space, specifically, combining the preliminary features, correlation, contrast, homogeneity, angular second moment, entropy and histogram, and recording them as the image space feature set.

4. The neurosurgery tumor patient health monitoring system based on machine learning according to claim 3 is characterized in that: In step B27, the neighborhood binary method specifically includes the following steps: Step B271: Obtaining a neighborhood binary code, specifically, taking any pixel in the two-dimensional medical image as the center of a circle, In the circular sampling area with radius Sampling points, calculate the neighborhood binary operator, arrange the neighborhood binary operators of all sampling points in clockwise order, use binary representation to obtain the neighborhood binary code; Step B272: updating the neighborhood binary code, specifically, cyclically shifting the binary neighborhood binary code, and selecting the neighborhood binary code with the smallest value after cyclic shift as the new neighborhood binary code; Step B273: construct a histogram, specifically, convert the neighborhood binary code into decimal, use the neighborhood binary code as the horizontal axis, and use the number of times the neighborhood binary code appears in the two-dimensional medical image as the vertical axis to construct a histogram.

5. The neurosurgery tumor patient health monitoring system based on machine learning according to claim 1 is characterized in that: The three-dimensional rehabilitation video feature extraction module adopts a three-dimensional video analysis method, and the three-dimensional video analysis method specifically includes the following steps: Step C1: human key point tracking, specifically, using AlphaPose posture estimation technology, based on the 3D rehabilitation video, predicting the human key point positions and the confidence scores of the key point positions in the 3D rehabilitation video, generating a corresponding heat map matrix for each key point, and determining the key point coordinates by finding the position of the maximum value in the heat map matrix; Step C2: Key point weight assignment, assigning significance weights to each key point; Step C3: multi-level smoothing of posture signals, specifically, using fast Fourier transform to analyze the frequency components of the key point change trajectory, using a smoothing window to capture the key point position changes over a long period of time, smoothing the trajectory change trend of the key points, calculating the weighted exponential moving average, and using the weighted exponential moving average to smooth the key point coordinates over a short period of time; Step C4: semantic matching, obtaining matching points; Step C5: three-dimensional feature screening, specifically, using a gated recurrent unit to screen a video posture feature set from the three-dimensional rehabilitation video according to key points and matching points.

6. The neurosurgery tumor patient health monitoring system based on machine learning according to claim 5 is characterized in that: In step C4, the semantic matching specifically includes the following steps: Step C41: convolution extraction, specifically, using a traditional convolutional neural network to mine features in two-dimensional medical images and three-dimensional rehabilitation videos; Step C42: extracting intermediate features, specifically, using a residual network to extract intermediate features from the convolutional layer of the traditional convolutional neural network, and using a bilinear interpolation method on the intermediate features to unify the spatial dimensions of the intermediate features; Step C43: feature slicing, specifically, dividing the intermediate feature into n groups of sub-features along the channel dimension, calculating the cosine similarity of each group of sub-features, generating a single-channel correlation graph, and stacking the single-channel correlation graphs of all groups of sub-features along the channel dimension to obtain a multi-layer correlation graph; Step C44: Optimizing the multi-layer correlation graph, specifically, performing a channel aggregation operation on the multi-layer correlation graph using a point convolution method to obtain a final correlation matrix; Step C45: Matching point mapping, specifically, mapping the two-dimensional medical image and the three-dimensional rehabilitation video according to the final correlation matrix to obtain matching points between the two-dimensional medical image and the three-dimensional rehabilitation video.

Images