A cobb angle measurement method based on AiteekEngine algorithm model

Abstract

The present invention relates to a cobb angle measurement method based on the AiteekEngine algorithm model. Step 1: acquiring characteristic data and color images of a plurality of human backs, and using machine learning to identify X-ray images and color images; Step 2: the machine identifies the color images and calculates the cobb angle; Step 3: generating a visualization report. The present invention provides a cobb angle measurement method based on the AiteekEngine algorithm model, which has high measurement efficiency and low economic cost.

Description

Technical Field

[0001] The invention belongs to the technical field of scoliosis assessment, and in particular relates to a cobb angle measurement method based on an AiteekEngine algorithm model. Background Art

[0002] Existing scoliosis detection technologies cannot support large-scale comprehensive screening. Current scoliosis screening still uses manual assessment methods, which have the disadvantages of low efficiency, strong subjectivity and high cost. Professional scoliosis assessment can be performed in hospitals or professional institutions by taking X-rays and calculating the Cobb angle, which has the disadvantages of expensive equipment and complicated procedures.

[0003] The severity of scoliosis, progression assessment and surgical indications all require the measurement of the Cobb angle of scoliosis. Commonly used Cobb angle measurement methods in clinical practice include protractors, computer image archiving and communication systems, and smartphone software, all of which require specific measurement tools, have limited usage conditions, and have poor repeatability. Therefore, using the shooting function of electronic devices such as mobile phones and mobile terminals to measure the Cobb angle of various scoliosis and evaluate the degree of scoliosis has high clinical value.

[0004] Therefore, it is of great significance to invent a simple, fast, efficient, non-invasive, radiation-free, non-contact, highly accurate and low-cost screening method by utilizing the shooting functions of electronic devices such as mobile phones, mobile terminals, all-in-one machines, robots, etc. Summary of the invention

[0005] In order to solve the existing scoliosis assessment method, a manual assessment method is usually adopted, which has the disadvantages of low efficiency, strong subjectivity and high cost. The present invention provides a cobb angle measurement method based on the AiteekEngine algorithm model, which has high measurement efficiency and low economic cost.

[0006] The technical solution of the present invention is as follows:

[0007] A cobb angle measurement method based on the AiteekEngine algorithm model, the method is as follows:

[0008] Step 1: Obtain feature data and color images of several human backs, and use machine learning to identify X-ray images and color images;

[0009] Step 2: The machine recognizes the color image and calculates the cobb angle;

[0010] Step 3: Generate a visualization report.

[0011] Preferably, the characteristic data described in step 1 includes the ATR value of the lumbar segment.

[0012] Preferably, the color images described in step 1 include X-ray spine images, thoracic segment images, thoracolumbar segment images, full-body images of the back of the human body, half-body images of the back, forward flexion images, and backward flexion images.

[0013] Preferably, machine learning is used to identify X-rays and color images, and the AiteekEngine algorithm model is trained, specifically: by comparing X-ray spine images, the full-body image of the back, the half-body image of the back, the flexion image and the flexion image of the human body are annotated, and multiple convolution feature maps are extracted from the color image; multiple convolution feature maps corresponding to the color image are cascaded to obtain the cascaded convolution feature map; the cascaded convolution feature map is used as input, and the human skeleton key point group data is used as output to train the prediction model, and the human skeleton key point prediction model is obtained; the key point group data is the coordinates of 13 human skeleton key points.

[0014] Preferably, the coordinates of the 13 key points of the human skeleton are: upper angle of left scapula, upper angle of right scapula, lower angle of left scapula, lower angle of right scapula, lowest point of left costal arch, lowest point of right costal arch, highest point of left iliac crest, highest point of right iliac crest, seventh cervical vertebra, seventh thoracic vertebra, twelfth thoracic vertebra, fourth lumbar vertebra and median sacral crest.

[0015] Preferably, step 2 is specifically as follows: the machine identifies the coordinates of the human skeleton key points of the color image according to the human skeleton key point prediction model, and then calculates 7 indicators based on the 13 human skeleton key point coordinates, namely: the shoulder inclination angle, that is, the angle between the scapular line and the horizontal line, the iliac crest inclination angle, the lumbar concave angle, the spinal deviation angle from the midline, the ATR value of the thoracic segment, the ATR value of the thoracolumbar segment and the ATR value of the lumbar segment, and each X-ray spine image is used to calculate the cobb angle, the ATR value of the thoracic segment, the ATR value of the thoracolumbar segment, the ATR value of the lumbar segment and the color image to establish a linear relationship, and at the same time establish a database of 7 indicators and the cobb angle, calculate the offset of the scoliosis cobb angle and the 7 indicators, and complete the measurement and evaluation of the cobb angle.

[0016] Preferably, step 3 is specifically as follows: performing a multi-dimensional evaluation based on the AiteekEngine algorithm model, and finally generating a screening report containing results, data and suggestions, generating corresponding test results by AI analysis, and generating personalized suggestions and evaluations based on the test results.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] The present invention provides an independently developed AlteekEngine core algorithm, which is based on the recognition of key bone points on the back of the Transforme model, integrating the advantages of FasterViT and YOLOv8, establishing the AlteekEngine core algorithm model, and using the AI ​​model to transform the image classification problem into a sequence modeling problem, thereby improving the calculation efficiency and accuracy;

[0019] Combining artificial intelligence with the Internet of Things and big data, it can dynamically identify and reconstruct scoliosis for evaluation, and can quickly and contactlessly screen for scoliosis on a large scale. The AlteekEngine core algorithm is applied to human posture recognition, capturing and identifying key bone points on the back and sides, establishing a matrix of four horizontal and one vertical key bone points to compare key movements, and combining the Adam test to establish seven-dimensional scoliosis assessment data for AI recognition and evaluation.

[0020] Compared with other products, AIteekEngine, built around AI vision technology, combines artificial intelligence with the Internet of Things and big data, significantly improving screening efficiency; AIteekEngnie's core algorithm can accurately collect more than 960 joint movement information per second in a single thread, and has an extremely high motion posture recognition rate; at the same time, AIteekEngnie's core algorithm is mainly based on cloud computing, supplemented by front-end computing, and the evaluation results are highly consistent with the Cobb angle of X-ray photography.

[0021] The above solution solves the problems of low overall efficiency and high subjectivity of traditional screening, as well as the radiation of imaging equipment, high equipment requirements, and the need for professional doctors. The AlteekEngine core algorithm ensures accuracy while improving the speed of detection, comprehensively solving the problems of existing screening technologies. It is very suitable for large-scale detection work and provides a very good method to assist in the screening of adolescent scoliosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart for machine learning;

[0023] Figure 2 is the mean square error matrix of the two teams from W1 to W7;

[0024] Figure 3 This is a schematic diagram of the scoliosis detection report (partial);

[0025] Figure 4 This is a schematic diagram of the scoliosis detection report (partial);

[0026] Figure 5 This is a schematic diagram of the scoliosis detection report (partial);

[0027] Figure 6Schematic diagram of 13 human skeleton points. DETAILED DESCRIPTION

[0028] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] A cobb angle measurement method based on the AiteekEngine algorithm model, the method is as follows:

[0030] Step 1: Obtain feature data and color images of several human backs, and use machine learning to identify X-ray images and color images;

[0031] Step 2: The machine recognizes the color image and calculates the cobb angle;

[0032] Step 3: Generate a visualization report.

[0033] The AiteekEngine algorithm model is a new and efficient CNN and Transformer hybrid visual basic model with a mixture of high and low frequency features, and an improved application of Transformer. The algorithm introduces a new tokenmixer called RepMixer, which uses structural reparameterization technology to reduce memory access costs by removing shortcuts in the network, especially at higher resolutions. Further architectural changes are also proposed to improve the performance of ImageNet classification tasks and other downstream tasks. Under the premise of the same accuracy on mobile devices and ImageNet datasets, the AiteekEngine algorithm model is 3.5 times faster than FastVit. Under similar latency, FastViT's Top-1 accuracy on ImageNet is 4.2% higher than MobileOne, making it a highly competitive hybrid architecture model.

[0034] FastViT introduces large kernel convolution in two locations, namely the PatchEmbedding layer and FFN. This method enhances the expressiveness of the model by mixing CNN and Vit, enabling it to effectively capture local and global information. At the same time, deployment-friendly mechanisms such as convolution multi-group reparameterization, lightweight multi-head self-attention, and convolution fusion blocks are introduced to further improve the performance of the model and reduce computational overhead. With the increase of receptive field, large kernel convolution also helps to improve the robustness of the model.

[0035] The AiteekEngine algorithm model is based on the recognition of key bone points on the back of the Transformer, integrating the advantages of FasterViT and YOLOv8 to establish the AiteekEngine algorithm model. The AI ​​model is used to transform the image classification problem into a sequence modeling problem, which improves the computational efficiency and accuracy. Compared with the convolutional neural network (CNN), the Vision Transformer (ViT) has achieved stable performance on multiple benchmarks such as human image segmentation, detection and recognition by relying on its excellent modeling ability; at the same time, it only requires fewer computing resources to train, achieving more efficient and accurate recognition of key bone points. Transformer is a sequence modeling method based on the self-attention mechanism. It transforms the image classification problem into a sequence modeling problem. By introducing a self-attention mechanism based on a sliding window and combining the local receptive field, it improves the computational efficiency and accuracy. Compared with the convolutional neural network (CNN), the Vision Transformer (ViT) has achieved very excellent performance on multiple benchmarks such as human image segmentation, detection and recognition by relying on its excellent modeling ability, and at the same time requires fewer computing resources to train.

[0036] The present invention will be further described below in conjunction with a specific embodiment:

[0037] S1: Machine learning to identify X-ray images and color images;

[0038] S11. Recruit 5,000 volunteers with poor posture, mild, moderate and severe scoliosis. For each volunteer, attach magnets to five key bone points on the body, namely, the inferior angles of the left and right scapulae, the seventh cervical vertebra, the seventh thoracic vertebra, and the fourth lumbar vertebra, to locate the bone points, then perform X-ray scanning to take X-ray spinal images, use a mobile phone to take four color images of the back, including the whole body, half of the back, flexion and backward flexion, and use a scoliosis measuring ruler to measure the thoracic segment ATR value, thoracolumbar segment ATR value, and lumbar segment ATR value.

[0039] S12. Annotate 20,000 images of 5,000 volunteers with human skeleton points. Collect RGB images of the whole body from the back, half body from the back, back-bent forward RGB images, and front-bent forward RGB images of people of different genders and different degrees of scoliosis, and save them as image datasets of corresponding body positions. At the same time, take a standing X-ray of the entire spine in the same posture. When shooting, attach magnets to the five key skeleton points of the tester's body, namely, the lower angles of the left and right scapulae, the seventh cervical vertebra, the seventh thoracic vertebra, and the fourth lumbar vertebra, to locate the skeleton points. Pair the collected X-rays with the image dataset.

[0040] S13. Use snipaste software to call out the X-ray film, use Labelme annotation software to manually select the human back image area in the photo, fit the X-ray film with the human image, compare them, adjust the transparency of the X-ray film for data correction, and further manually annotate 13 human bone points, namely, the upper angle of the left scapula, the upper angle of the right scapula, the lower angle of the left scapula, the lower angle of the right scapula, the lowest point of the left costal arch, the lowest point of the right costal arch, the highest point of the left iliac crest, the highest point of the right iliac crest, the seventh cervical vertebra, the seventh thoracic vertebra, the twelfth thoracic vertebra, the fourth lumbar vertebra and the median sacral crest.

[0041] S14. Convert the labeled data set into MS COCO format, and then classify the data set into 12,000 training data sets, 4,000 test data sets, and 4,000 validation data sets, which are allocated in a ratio of 60%, 20%, and 20%.

[0042] S15. Using the MS COCO dataset, the ViTAE-G model with an accuracy of 81.1AP and 1 billion parameters is trained as the basis. In addition, considering the problem of insufficient pixel and resolution accuracy of mobile phone cameras, tuning is performed during the training process to further optimize the impact of factors such as model scale, input resolution, and feature map size on the vision transformer in the task of human bone point recognition. Through training, an autonomously trained aiteek-G model that can accurately recognize 13 human bone points is obtained. This model structure serves as an encoder and a small number of deconvolution layers as a decoder, which can achieve good performance in the task of human bone point recognition and has good scalability and flexibility.

[0043] S16, the four color images of the back of the human body, the back half of the human body, the front bending, and the back bending are represented by s1, s2, s3, and s4 respectively. The specific machine learning process is as follows: Figure 1 As shown:

[0044] The shoulder inclinations of the two points p1 and p2 and the horizontal line are calculated by the upper scapular angles on both sides of s1. The p3 and p4 of the first image s1 are calculated by the same method and recorded as s1p1p2. The p1 and p2 of the second image s2 are recorded as s2p1p2. The shoulder inclinations of s1 include s1p1p2 and s1p3p4. The shoulder inclinations of s2 include s2p1p2 and s2p3p4. The maximum shoulder inclination x1 is obtained by the activation function softMax. Use the same method to calculate x2 to x7 respectively, and calculate the cobb angle through the linear relationship, that is, cobb=ax+b, a is the coefficient, b is the offset, where x=w1x1+w2x2+w3x3+w4x4+w5x5+w6x6+w7x7, where w1 is the learning weight of x1, w2 to w7 are the learning weights of x2 to x7 respectively, and the estimated values ​​of w1 to w7 are calculated respectively through the mean square error matrix of the two teams.

[0045] A 7-dimensional data set is formed by w1 to w7, such as Figure 2 As shown, the calculation steps of the covariance matrix are: ;

[0046] The estimated values ​​of w1 to w7 are calculated separately through the second-order mean square error matrix, and the gradient descent algorithm of linear algebra is used to implement it. When processing large-scale data sets, it is also necessary to use the properties of linear algebra to perform matrix decomposition, eigenvalue calculation and other operations to improve the efficiency of the algorithm.

[0047] Through the above training, we obtain our own aiteek-G human body key bone point recognition model and the linear equation for calculating the cobb angle by MSE: cobb=ax, coefficient a is 0.38, offset b is 0, where x= x1+x2+x3+3.2x4+x5+x6+x7.

[0048] S2, the machine recognizes the color image and calculates the cobb angle;

[0049] S21. The tester stands about 1.5-2 meters in front of the measuring area. The tester is topless (male) or wears a sports bra (female). The tester remains relaxed, with both hands open, palms facing forward, and stands horizontally.

[0050] S22. The tester takes the photo as required and uploads the first full-body photo of the back for detection. The 13 key bone points of the human body are detected and identified according to the trained aiteek-G recognition model. After passing, the second half-body photo of the back is collected. After passing the same detection, the third forward bending and the fourth backward bending color images are completed in sequence. After all the tests are passed, the saved data is obtained. The coordinates of the 13 key bone points P1 to P13 of the 4 human back photos are identified. Figure 6As shown, P1 to P13 correspond to the left superior angle of scapula, the right superior angle of scapula, the left inferior angle of scapula, the right inferior angle of scapula, the lowest point of left costal arch, the lowest point of right costal arch, the highest point of left iliac crest, the highest point of right iliac crest, the seventh cervical vertebra, the seventh thoracic vertebra, the twelfth thoracic vertebra, the fourth lumbar vertebra and the median sacral crest, respectively.

[0051] S23. Calculate the inclination angle α between P1 and P2 and the horizontal line according to the coordinate points P1 and P2 on the first image s1, and record the inclination angle α as s1p1p2. Calculate the inclination angle β between P3 and P4 and the horizontal line according to the coordinate points P3 and P4 on the first image s1, and record the inclination angle β as s1p3p4. Calculate the inclination angle γ between P1 and P2 and the horizontal line in the second image s2 by the same method, and record the inclination angle γ as s2p1p2. Calculate the inclination angle d between P3 and P4 and the horizontal line in the second image s2, and record the inclination angle d as s2p3p4. The shoulder inclination of s1 is expressed as s1p1p2 and s1p3p4, and the shoulder inclination of s2 is expressed as s2p1p2 and s2p3p4. Obtain the maximum shoulder inclination x1 through the activation function softMax.

[0052] That is, x1=SoftMax(s1p1p2, s1p3p4, s2p1p2, s2p3p4)

[0053] Calculate x2 to x7 using the same method as above;

[0054] x2=SoftMax(s1p5p6,s2p5p6)

[0055] x3=SoftMax(s1p7p8,s2p7p8)

[0056] x4=SoftMax(s1p9p10, s1p10p11, s2p11p12, s2p12p13)

[0057] x5=SoftMax(s3p1p2, s3p3p4, s4p1p2, s4p3p4)

[0058] x6=SoftMax(s3p5p6,s4p5p6)

[0059] x7=SoftMax(s3p7p8,s4p7p8)

[0060] S24. After calculating x1 to x7, calculate the value of x according to x= x1+x2+x3+3.2x4+x5+x6+x7. After obtaining the x value, calculate the cobb angle value according to the linear relationship cobb=0.38x.

[0061] S3, based on the AiteekEngine algorithm model, conducts a multi-dimensional assessment, and finally generates a screening report containing results, data, and suggestions. It generates corresponding test results through AI analysis, and generates personalized suggestions and assessments based on the test results, such as Figure 3-5 shown.

[0062] Scoliosis is divided into 5 levels according to the size of the Cobb angle: 1. Cobb angle <10 degrees: normal, no special treatment is recommended, only follow-up observation is required.

[0063] 2. Cobb angle 10~20 degrees: poor posture. According to the specific age, set exercise prescription and carry out corresponding functional exercises.

[0064] 3. Cobb angle 20~30 degrees: mild scoliosis, choose brace correction, pay attention to go to the hospital for follow-up in time, and adjust the brace in time according to height changes.

[0065] 4. Cobb angle 30~40 degrees: moderate scoliosis. If the patient is at the peak stage of growth and development, surgical treatment is recommended; if the patient is not growing or at the end of the growth stage, surgical treatment is not recommended.

[0066] 5. Cobb angle > 40 degrees: Severe scoliosis is generally treated with surgery, but considering the high risk of surgery, conservative treatment is continued.

[0067] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A cobb angle measurement method based on the AiteekEngine algorithm model, characterized in that: The method is as follows: Step 1: Obtain feature data and color images of several human backs, and use machine learning to identify X-ray images and color images; Machine learning recognizes X-rays and color images and trains the AiteekEngine algorithm model. Specifically, by comparing X-ray spine images, the back full-body image, back half-body image, flexion image and back flexion image of the human body are annotated, and multiple convolution feature maps are extracted from the color image; multiple convolution feature maps corresponding to the color image are cascaded to obtain the cascaded convolution feature map; the cascaded convolution feature map is used as input, and the human skeleton key point group data is used as output to train the prediction model to obtain the human skeleton key point prediction model; the key point group data is the coordinates of 13 human skeleton key points; Step 2: The machine recognizes the color image and calculates the cobb angle; Step 3: Generate a visualization report; The color images in step 2 include four color images of the whole back, half back, forward bending, and backward bending taken by a mobile device; The machine recognizes color images and calculates the cobb angle, including: Based on the four color images, the coordinates of 13 key bone points on the back of the human body are identified, including the upper angle of the left scapula, the upper angle of the right scapula, the lower angle of the left scapula, the lower angle of the right scapula, the lowest point of the left costal arch, the lowest point of the right costal arch, the highest point of the left iliac crest, the highest point of the right iliac crest, the seventh cervical vertebra, the seventh thoracic vertebra, the twelfth thoracic vertebra, the fourth lumbar vertebra and the median sacral crest; Based on the coordinates of the skeleton key points, combined with the activation function softMax, the maximum shoulder inclination x1, the maximum iliac crest inclination x2, the maximum lumbar concave angle x3, the maximum spinal deviation angle x4, the maximum thoracic ATR x5, the maximum thoracolumbar ATR x6, and the maximum lumbar ATR x7 are obtained; According to x1 to x7, an x ​​value is calculated by weighted summation, and the x value is linearly related to the cobb angle.

2. The cobb angle measurement method based on the AiteekEngine algorithm model according to claim 1 is characterized in that: The characteristic data described in step 1 include the ATR value of the lumbar segment.

3. The cobb angle measurement method based on the AiteekEngine algorithm model according to claim 1 is characterized in that: The color images described in step 1 include thoracic segment images, thoracolumbar segment images, full-body back images, half-body back images, forward flexion images, and backward flexion images.

4. The cobb angle measurement method based on the AiteekEngine algorithm model according to claim 1 is characterized in that: Step 3 is as follows: Conduct a multi-dimensional assessment based on the AiteekEngine algorithm model, and finally generate a screening report containing results, data, and suggestions. Generate corresponding test results through AI analysis, and generate personalized suggestions and evaluations based on the test results.

Images