Intelligent scoliosis diagnosis typing system based on medical images

Abstract

The invention discloses a scoliosis intelligent diagnosis typing system based on medical images. The scoliosis intelligent diagnosis typing system comprises a medical image obtaining and preprocessing module, an automatic Cobb angle measuring and calculating module, a scoliosis intelligent typing module and a personalized digital rehabilitation training module. The medical image obtaining and preprocessing module is used for collecting and storing a spinal orthotopic image X-ray film of an upright position of a patient and preprocessing the X-ray film. The invention belongs to the technical field of medical diagnosis, and aims at solving the problems that in the prior art, a manual measurement method is prone to generating obvious individual differences and errors, objectivity and repeatability of results are difficult to guarantee, and meanwhile scoliosis typing is not fully brought into conservative treatment requirements. The method has the technical effects that errors or delay caused by multi-link manual operation can be effectively reduced, and meanwhile, conservative treatment and fine typing are combined.

Description

Technical Field

[0001] This application belongs to the technical field of medical diagnosis, and specifically relates to an intelligent diagnosis and classification system for scoliosis based on medical images. Background Art

[0002] Scoliosis is a complex three-dimensional spinal deformity, especially adolescent idiopathic scoliosis (AIS) is the most common, affecting 2% - 3% of adolescents globally. Scoliosis not only affects the patient's physical appearance, but may also lead to impaired cardiopulmonary function and mental health problems. Clinically, doctors usually measure the Cobb angle on X-ray films to accurately diagnose and evaluate the severity of scoliosis. Based on the evaluation results of the Cobb angle, doctors will decide whether to perform surgical treatment or conservative rehabilitation treatment on the patient. In terms of surgical treatment, the classification methods of scoliosis, such as Lenke classification, King classification, etc., are mainly referred to to determine the most appropriate surgical plan. And conservative rehabilitation treatment focuses on specific scoliosis training (PSSE) to help patients improve their conditions and promote recovery.

[0003] The measurement of the Cobb angle includes determining the end vertebrae (the vertebrae with the largest inclination degree towards the concave side of the spinal curvature in scoliosis). Draw a horizontal line on the upper edge of the vertebral body of the upper end vertebra and the lower edge of the vertebral body of the lower end vertebra respectively, and then draw a perpendicular line to each of these two horizontal lines. The included angle between these two perpendicular lines is the Cobb angle. When the Cobb angle < 10°: it is usually considered to be within the normal range and does not require special treatment; when the Cobb angle ≥ 10°: it indicates entering a more definite scoliosis range, and further subdivision is required according to the number of curves and the position of the apical vertebra, and regular follow-up is needed, especially during the adolescent development stage because of the risk of progression. Therefore, accurate measurement of the Cobb angle is crucial for the diagnosis and treatment of scoliosis patients.

[0004] However, the current clinical measurement of Cobb angle mainly relies on professionals for manual calibration and tangent drawing. The operation process is cumbersome, time-consuming, and highly dependent on doctors' experience and subjective judgment. This measurement method is prone to obvious individual differences and errors, making it difficult to ensure the objectivity and repeatability of the results. At the same time, the existing scoliosis classifications (such as Lenke, King, etc.) are more designed for surgical patients and do not fully incorporate the needs of conservative treatment. There is a lack of clear guiding principles for how to select or adjust rehabilitation training, resulting in random implementation, high difficulty, and difficult-to-quantify effects in clinical conservative treatment. In addition, although specific scoliosis training (PSSE) is often applied in conservative treatment, most patients lack professional supervision and personalized feedback when practicing at home. Due to unstandardized movements, insufficient frequency or intensity, the rehabilitation effect is greatly reduced, and in severe cases, the condition may even deteriorate. Due to problems such as limited medical resources, high costs for patients to travel to and from the hospital, and insufficient treatment time in the traditional outpatient rehabilitation model, it is also difficult to implement efficiently in the long term. Therefore, there is an urgent need for a scoliosis intelligent diagnosis and classification system that integrates automated image diagnosis, intelligent and precise classification, and remote digital rehabilitation training to solve the above problems. Summary of the Invention

[0005] This application provides a scoliosis intelligent diagnosis and classification system based on medical images, aiming to solve the problems that the existing manual measurement method is prone to obvious individual differences and errors, making it difficult to ensure the objectivity and repeatability of the results. At the same time, the scoliosis classification does not fully incorporate the needs of conservative treatment, resulting in random implementation and high difficulty in clinical conservative treatment.

[0006] A scoliosis intelligent diagnosis and classification system based on medical images includes a medical image acquisition and preprocessing module, an automated Cobb angle measurement and calculation module, a scoliosis intelligent classification module, and a personalized digital rehabilitation training module;

[0007] The medical image acquisition and preprocessing module is used to collect and store the anterior-posterior X-ray images of the patient's upright spine and preprocess the X-ray images;

[0008] The automated Cobb angle measurement and calculation module is used to calculate the Cobb angle of the preprocessed X-ray image parameters through a deep learning model and mark the upper and lower end vertebrae and the tangent;

[0009] The scoliosis intelligent classification module is used to comprehensively determine and classify scoliosis by combining the number of scoliosis curves and the vertebral segment where the apex vertebra is located;

[0010] The personalized digital rehabilitation training module can match the corresponding rehabilitation training plan according to the classification category.

[0011] Further, the medical image acquisition and preprocessing module includes full-spine anteroposterior X-ray shooting and image preprocessing;

[0012] For the medical image acquisition, the patient is required to stand upright facing the X-ray machine, with feet separated shoulder-width apart, keeping the pelvis basically horizontal, avoiding posture skew and rotation. Subsequently, the X-ray machine shoots and stores to obtain an X-ray film;

[0013] The image preprocessing performs preprocessing on the X-ray film in sequence, including denoising, histogram equalization enhancement, and highlighting of asymmetric features.

[0014] Further, the denoising includes median filtering and edge protection.

[0015] Further, the automated Cobb angle measurement and calculation module includes deep learning model construction, curve fitting and end vertebra determination, and Cobb angle calculation:

[0016] The deep learning model is constructed based on the CNN module and the ViT module, and the deep learning model is trained and optimized;

[0017] Based on the deep learning model, key vertebral body anchor points are identified and curve fitting is performed.

[0018] Further, the deep learning model's identification of key vertebral body anchor points and curve fitting specifically includes identification of key vertebral body anchor points, cubic spline curve fitting, and upper and lower end vertebra localization and scoliosis section identification.

[0019] Further, the Cobb angle calculation is used to calculate the tangent slopes of the upper and lower edges of the end vertebra;

[0020] After identifying the positions of the upper and lower end vertebrae, directly call the coordinates of the "upper and lower edge corner points" of the vertebral body output by the deep learning model for the upper and lower edge points (x u1 , y u1 ), (x u2 , y u2 ) for linear fitting to obtain the tangent slope k1 of the upper end vertebra; for the upper and lower edge points of the lower end vertebra (x d1 , y d1 ), (x d2 , y d2 ) for linear fitting to obtain the tangent slope k2 of the lower end vertebra;

[0021] Cobb angle calculation formula:

[0022]

[0023] where k1 and k2 respectively represent the tangent slopes of the upper end vertebra and the lower end vertebra; then convert the calculation result of the radian value of the Cobb angle to an angular value;

[0024] The conversion formula is as follows:

[0025]

[0026] Furthermore, the scoliosis intelligent classification module needs to preset classification principles, and then pre-classify scoliosis into the following four major types and several subtypes based on the classification principles, and input them into the deep learning model as the subsequent automatic classification discrimination rules, and output the classification results after determination.

[0027] Furthermore, the personalized digital rehabilitation training module includes a digital rehabilitation platform.

[0028] Furthermore, the digital rehabilitation platform is used to construct a PSSE training library, and generate a personalized training plan according to the classification result and the PSSE training library, monitor and adjust dynamically in real time, and conduct phased imaging reexamination and plan update.

[0029] Compared with the prior art, the present application has at least the following beneficial effects:

[0030] Based on further analysis and research of the problems in the prior art, the present application constructs a complete "diagnosis-rehabilitation-follow-up" closed-loop management system by deeply integrating the automatic measurement of Cobb angle, intelligent classification and digital remote rehabilitation of medical images; among them, from image acquisition, denoising and enhancement to Cobb angle calculation, and then to classification and rehabilitation plan push, can be completed within the same system, reducing errors or delays caused by multi-link manual operations, and improving the efficiency of diagnosis and rehabilitation training. Combining conservative treatment with fine classification, for non-surgical scoliosis patients, more attention is paid to the refined division of the chest, thoracolumbar, lumbar and other regions during classification, which can better guide the targeted distribution of rehabilitation movements. At the same time, with the help of the digital platform, long-term and effective remote supervision and personalized guidance for patients can be realized, reducing treatment interruption or incomplete rehabilitation caused by geographical and time limitations. This mode helps to solve problems such as insufficient medical resources and inconvenience for patients to travel back and forth, and is especially suitable for large-scale clinical screening and home rehabilitation scenarios, and has important clinical application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the overall process of a scoliosis intelligent diagnosis and classification system based on medical images provided by an embodiment of the present application;

[0032] Figure 2 It is a schematic diagram of the working process of a scoliosis intelligent diagnosis and classification system based on medical images provided by an embodiment of the present application; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0034] In one embodiment, as Figure 1 and Figure 2 shown, an intelligent diagnosis and classification system for scoliosis based on medical images provided by the present application includes a medical image acquisition and preprocessing module, an automated Cobb angle measurement and calculation module, an intelligent classification module for scoliosis, and a personalized digital rehabilitation training module; wherein, the medical image acquisition and preprocessing module includes full-spine frontal X-ray photography and image preprocessing;

[0035] The medical image acquisition and preprocessing module is used to collect and store the full-spine frontal X-ray image of the patient in the upright position, and preprocess the X-ray image.

[0036] The medical image acquisition and preprocessing module includes full-spine frontal X-ray photography and image preprocessing; the specific content of the full-spine frontal X-ray photography is as follows:

[0037] The patient is required to stand upright facing the X-ray machine, with the feet separated shoulder-width apart, keeping the pelvis basically horizontal, avoiding posture skew and rotation. Subsequently, the X-ray machine takes and stores the image to obtain the X-ray image. According to the patient's height and the length of the spine, an appropriate focus-film distance (FFD) is selected, and the distance is between 170 cm and 185 cm, which can not only cover the entire area from the cervical vertebrae to the sacrum, but also reduce geometric distortion.

[0038] If the equipment permits, the long film photography mode or the full-spine stitching (Stitching X-ray) technology can be used to obtain the complete image of the cervical, thoracic, lumbar, and sacral segments at one time, reducing the measurement deviation caused by stitching errors.

[0039] The X-ray machine is usually adjusted within the range of 70-90 kV and 5-10 mA, and can be fine-tuned according to the patient's body type, bone density, etc.; a grid is used to weaken scattered rays and improve the image clarity; at the same time, the image is stored in DICOM format, completely retaining the patient information and shooting parameters.

[0040] The specific content of the image preprocessing is as follows:

[0041] The X-ray image is preprocessed successively by denoising, histogram equalization enhancement, and highlighting of asymmetric features.

[0042] The denoising process includes median filtering and edge protection.

[0043] Median filtering: In this embodiment, a 3×3 median filter is selected to remove the possible random speckle noise (such as salt-and-pepper noise) in the X-ray image;

[0044] In the 3×3 neighborhood window, the pixel values are sorted by size, and the median value is taken to replace the pixel value at the center of the window:

[0045] Idenoise (x, y) = median{N 3×3 (x, y)}.

[0046] Median filtering can better preserve edge information while removing noise, avoiding over-smoothing of the vertebral contour.

[0047] Edge protection: For high-contrast noise regions at the spine edge, bilateral filtering or weighted Gaussian filtering can be combined for local processing to ensure the vertebral contour.

[0048] Due to the influence of X-ray imaging conditions, patient body type, and bone density, there are often problems of uneven brightness or insufficient contrast. Histogram Equalization can effectively enhance the recognizability of the vertebral edge.

[0049] Histogram Equalization requires statistical calculation of the gray-level distribution function p(i) of the gray-scale image, where i is the gray level and p(i) is the probability of the occurrence of this gray level;

[0050] Calculate the cumulative distribution function s based on p(i). The calculation formula of s is:

[0051]

[0052] Map the original gray level i to the new gray level s to achieve stretching and equalization of the global dynamic range, thereby enhancing the contrast between the spine and surrounding soft tissues.

[0053] The finally generated enhanced image is denoted as I equalized (x, y), improving the recognizability of key structures such as the spinal cortical bone and vertebral edge.

[0054] To further emphasize the left-right asymmetry of the spine, this embodiment introduces a mirror flip superposition algorithm. The specific process is as follows:

[0055] a) Central axis determination

[0056] First, it is necessary to initially identify the approximate central position of the spine in the image. For example, based on the centroid of the vertebral body area or the overall contour in the anterior-posterior view of the spine, an approximate central axis passing through C7 to the sacrum is obtained.

[0057] Since this central axis is used as a reference for horizontal mirror flipping, its accuracy is very important for subsequent processing. The spinal contour can be initially obtained through edge detection or Hough Transform, and then the position of the spinal midline can be estimated from this.

[0058] b) Mirror flip

[0059] The original image I after denoising and enhancement equalized(x, y) is horizontally mirror - flipped along the central axis of the spine to obtain the mirror image I mirror (x, y).

[0060] c) Superposition operation

[0061] The pixel values of the original enhanced image and the mirror image are averaged pixel - by - pixel;

[0062] The calculation formula for pixel - by - pixel averaging is:

[0063]

[0064] In the area where the spine significantly bulges to one side, the superimposed pixel values will show certain abnormal differences, making it easier to identify in the overall image. The finally obtained I enhanced (x, y) not only retains the global contrast enhancement effect of the original image but also further amplifies the asymmetry between the left and right structures of the spine, providing clearer feature information for the subsequent deep - learning model to automatically identify and locate the scoliosis area.

[0065] The automated Cobb angle measurement and calculation module is used to calculate the Cobb angle of the pre - processed X - ray image parameters through a deep - learning model, mark the upper and lower end vertebrae and the tangent lines, and superimpose and display the Cobb angle value.

[0066] The specific content is as follows:

[0067] a) Construct a deep - learning model based on the CNN module and the ViT module:

[0068] The CNN module uses a typical backbone convolutional network (such as ResNet or VGG variants) to extract local features of the spine image, including microscopic information such as vertebral body edges, cortical textures, and intervertebral spaces. The convolutional layer, pooling layer, and batch normalization (BatchNormalization) layer are used in combination to extract multi - scale feature maps (Feature Maps).

[0069] The ViT module serializes the feature maps output by the CNN to generate a series of feature vectors (Patch Embeddings), which are input into the Vision Transformer. At the same time, the ViT module models the global relationships and spatial context information among the vertebrae through the multi-head self-attention mechanism; the global modeling ability helps to accurately locate the most obvious curve segment of scoliosis and determine the position of the apical vertebra. Since the main task of this model is vertebral key point detection and coordinate regression, a key point regression branch can be set at the end of the network to output the (x, y) positions of each vertebral anchor point in the image coordinate system; at the same time, an auxiliary classification branch can be set to determine whether there are multiple curvatures in the image, etc., providing reference information for intelligent classification.

[0070] b) Train and optimize the deep learning model;

[0071] Loss function: Input the loss function in the deep learning model so that it can perform vertebral key point detection. The L1 or L2 distance loss (Smooth L1 Loss / MSE Loss) can be used to measure the gap between the predicted coordinates and the annotated coordinates. The calculation formula is:

[0072]

[0073] where and p i are the predicted coordinates and the true annotated coordinates of the i-th key point respectively, and N is the number of key points.

[0074] Optimization algorithm: Use the adaptive moment estimation (Adam) or AdamW optimization algorithm for training. The initial learning rate can be set to 1×10 -4 or 2×10 -4 , and an annealing (Learning Rate Decay) strategy is carried out during the training process. The number of training epochs depends on the data scale and is usually between 100 and 300. After each Epoch, the validation set is evaluated, and the learning rate is adjusted dynamically.

[0075] Evaluation metrics

[0076] Multidimensional metrics such as the mean key point localization error (Mean Euclidean Distance, MED) and the peak signal-to-noise ratio (PSNR) are used to evaluate the performance of the model in terms of localization accuracy and visual output quality; at the same time, the differences between the Cobb angle values measured manually and the values measured automatically by the model (such as the mean absolute difference MAD, the correlation coefficient R value, etc.) can be compared to test the clinical usability of the model.

[0077] c) Identification of key vertebral anchor points and curve fitting based on a deep learning model:

[0078] Identification of key vertebral anchor points: The optimized deep learning model outputs a vertebral coordinate sequence in the form of {(x1, y1), (x2, y2),..., (x n , y n )}, where n is the number of vertebrae. If there are multiple types of key points such as "upper and lower edge corner points" and "vertebral center points", they need to be classified and stored separately for subsequent processing. If a very small number of points deviate from the spinal region or the coordinates are missing in the output of the deep learning model, a method based on neighborhood interpolation or the nearest legal point can be set for correction; noise points can also be removed according to the geometric constraints of the upper and lower vertebral positions (for example, adjacent vertebrae cannot overlap excessively).

[0079] Cubic spline curve fitting: After obtaining the vertebral center or midline coordinates, cubic spline interpolation is used to smoothly fit the overall course of the spine:

[0080] S i (x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3 (i = 1, 2,..., m)

[0081] where m is the number of spline segments, and the coefficients a i , b i , c i , d i are solved by the least squares method.

[0082] The entire spine can be divided into several interpolation intervals, ensuring continuity between each segment. Finally, a smooth and continuous main curve of the spine is obtained, providing an accurate geometric basis for automatically determining the apex and end vertebrae of scoliosis.

[0083] Location of upper and lower end vertebrae and identification of scoliosis segments

[0084] By calculating the offset (or curvature) of the curve in the longitudinal direction, the most significant region of the scoliosis curve and the possible location of the apex vertebra can be quickly located. If there are multiple curves, multiple apex vertebrae can be identified. For the identified scoliosis curve segment, the vertebra at the maximum bending amplitude is selected as the apex vertebra; the vertebrae at the start and end of the curve segment (where the curvature changes rapidly or the curve turns) are used as the upper and lower end vertebrae, facilitating the use of their upper and lower edges as reference tangents when calculating the Cobb angle subsequently.

[0085] d) Cobb angle calculation:

[0086] Calculation of the tangent slope of the upper and lower edges of the end vertebra

[0087] After identifying the positions of the upper and lower end vertebrae, directly call the coordinates of the "upper and lower edge corner points" of the vertebra output by the deep learning model, or refine the positioning in combination with the size / shape of the vertebra.

[0088] If the "corner points" cannot be accurately obtained, the least squares line fitting method can also be used to estimate the tangent direction of the end vertebra by taking several points on the upper and lower edges of the end vertebra.

[0089] For the upper and lower edge points (x u1 , y u1 ), (x u2 , y u2 ), perform line fitting to obtain the tangent slope k1 of the upper end vertebra;

[0090] For the upper and lower edge points of the lower end vertebra (x d1 , y d2 ), (x d2 , y d2 ), perform line fitting to obtain the tangent slope k2 of the lower end vertebra. If the tangents are approximately perpendicular, the case where the slope tends to infinity can be specially processed.

[0091] Cobb angle calculation formula:

[0092]

[0093] Among them, k1 and k2 respectively represent the tangent slopes of the upper end vertebra and the lower end vertebra.

[0094] The above calculation results are in radian values. For the convenience of clinical application, the Cobb angle radian value should be converted to the Cobb angle value (unit: degree), and the conversion formula is:

[0095]

[0096] Visually display the upper and lower end vertebrae and tangents marked on the image, and superimpose and display the Cobb angle value; at the same time, output a measurement report (Cobb angle value, apex vertebra position, measurement time, etc.) for subsequent classification module calls.

[0097] The scoliosis intelligent classification module is used to comprehensively determine and classify scoliosis by combining the number of scoliosis curves and the vertebral segment where the apex vertebra is located. The specific content is as follows:

[0098] a) Severity determination based on Cobb angle

[0099] When the Cobb angle < 10°, most clinical views believe that it does not belong to typical scoliosis and can be regarded as a minor curve or a mild variation in spinal morphology.

[0100] When the Cobb angle ≥ 10°, it enters a more definite scoliosis range and needs to be further subdivided according to the number of curves and the position of the apical vertebra.

[0101] Macro-classification based on the number of curves

[0102] When there is only a single main curve in the spine, it is defined as the single curve type (Type I);

[0103] When there are two main curves, it is defined as the double curve type (Type II);

[0104] When three curves appear simultaneously, it is defined as the triple curve type (Type III).

[0105] When the deep learning model determines the number of curves, the most significant Cobb angle radian is taken as the main curve; if the Cobb angle value of another / other two curves exceeds a certain threshold (usually 70% - 80% or higher of the main curve), it is regarded as "double curve" or "triple curve", otherwise it may be defined as "compensatory curve", in combination with the clinical auxiliary diagnosis suggestions.

[0106] Refined classification based on the position of the apical vertebra

[0107] Taking the vertebral segment number as a reference, the apical vertebra falling in the thoracic segment (T2 - T11), the thoracolumbar junction (T12 - L1) or the lumbar segment (L2 and below) is used as the main distinguishing point to identify the main plane position where the curve occurs.

[0108] For the double curve type or triple curve type, the upper and lower (or upper, middle and lower) apical vertebra positions are recorded respectively, and the detailed subtypes are finally established.

[0109] b) Based on the classification principle, scoliosis is preliminarily divided into the following four major types and several subtypes, and input into the deep learning model as the subsequent automatic classification discriminant rules;

[0110] a) Type 0: Minor curve, Cobb angle < 10°. Clinically, generally no special intervention is required, only regular follow-up observation or simple postural training is carried out.

[0111] b) Type I: Single curve type, Cobb angle ≥ 10°, and only 1 main radian is identified. It is divided into three subtypes according to the segment where the apical vertebra is located:

[0112] I(a): Thoracic curve (the apical vertebra is located at T2 - T11).

[0113] I(b): Thoracolumbar curve (the apical vertebra is located at T12 - L1).

[0114] I(c): Lumbar curve (the apical vertebra is located at L2 and below).

[0115] Type II: Double-curved type, Cobb angle ≥ 10°, two main curves are identified. For the combination of the apical vertebral segments of the upper and lower curves, the system is subdivided as follows:

[0116] II(a): Double thoracic curve (e.g., the upper thoracic curve is from T2 to T6, and the lower thoracic curve is from T7 to T11).

[0117] II(b): Thoracic curve with thoracolumbar or lumbar curve.

[0118] For example, the apical vertebra of the upper thoracic segment is between T2 and T11, and the apical vertebra of the lower curve is at T12 to L1 or L2 and below.

[0119] II(c): Upper thoracic curve with thoracolumbar or lumbar curve.

[0120] For example, the apical vertebra of the upper curve is at a higher position from T1 to T4, and there is a thoracolumbar or lumbar curve at the same time.

[0121] Type III: Triple-curved type, Cobb angle ≥ 10°, three obvious scoliosis curves exist simultaneously, usually manifested as the coexistence of "upper thoracic curve + main thoracic curve + thoracolumbar or lumbar curve", and multiple apical vertebrae are located in the corresponding segments respectively.

[0122] c) Determine and output the classification result

[0123] Automatic classification determination

[0124] Input the Cobb angle, number of curves, and apical vertebral coordinate information calculated and detected in the above steps, and automatically discriminate according to the rules set above; if the Cobb angle is abnormal (such as exceeding the reasonable range) or the number of curves has a fuzzy boundary, prompt for manual review.

[0125] Multi-curve priority judgment

[0126] When the system detects multiple curves, it will mark the curve with the largest amplitude as the "main curve" and give priority to the main curve in the classification structure.

[0127] Apical vertebral position; if other curves reach a certain proportion, they will be marked as "secondary curve" or "third curve".

[0128] If there are two almost identical main curves (such as the Cobb angles of the thoracic and lumbar segments differ by only 1°), the system will give priority to the curve with the higher apical vertebra or the curve following the clinical statistical law as the main curve, but also prompt the doctor to pay attention to the actual condition.

[0129] Visualization and reporting

[0130] Mark the main curved segments, apical vertebral positions, and Cobb angles on the spinal image, and output a "scoliosis classification" report.

[0131] The Cobb angle is calculated by fitting the upper and lower edge points of the end vertebrae to obtain the tangent lines of the upper and lower end vertebrae. The Cobb angle value is obtained through calculation and is visually marked on the image.

[0132] The intelligent scoliosis classification module takes the Cobb angle size (<10° or ≥10°), the number of curves (single curve, double curve, triple curve), and the position of the apical vertebra (thoracic segment, thoracolumbar segment, lumbar segment) as the core parameters.

[0133] When the Cobb angle <10°, it is defined as type 0 (minor curve), mainly for observation or mild postural training; when the Cobb angle ≥10°, it is subdivided into type I (single curve), type II (double curve), and type III (triple curve) according to the number of curves and the position of the apical vertebra, and subtypes such as thoracic curve, thoracolumbar curve, and lumbar curve are further specifically distinguished under each type. The system automatically determines the patient's image data according to the classification discrimination rules. If there are boundary situations or multiple curves that are difficult to distinguish, it will prompt for manual review. Finally, a classification report is output, and the classification information (such as the position of the main curve, Cobb angle value, etc.) is transmitted to the digital rehabilitation platform to provide a basis for personalized training.

[0134] The personalized digital rehabilitation training module includes a digital rehabilitation platform that can match the corresponding rehabilitation training plan according to the classification category. The specific content is as follows:

[0135] a) Construct a PSSE training library (scoliosis-specific exercise)

[0136] According to the classification subdivisions such as I(a), I(b), I(c), II(a), II(b), II(c), and type III, different training focuses are defined: for example, patients with thoracic curves focus on upper torso symmetry and respiratory correction, and patients with lumbar curves strengthen the stability of the pelvis and lumbar back muscles; each training action is accompanied by metadata such as the execution frequency, action key points, and video demonstrations in the library.

[0137] According to the Cobb angle stratification, for patients with a larger Cobb angle, higher-intensity or more complex overcorrection actions can be added. For patients with a smaller Cobb angle or in the early stage, postural maintenance and flexible training are the main focuses to avoid unnecessary overtraining.

[0138] b) Generate a personalized training plan

[0139] According to the patient's classification and comprehensive evaluation (such as muscle strength, flexibility), the system automatically generates a daily / weekly training plan and pushes it to the patient's mobile terminal or Web terminal; the action demonstration video contains information such as correct breathing coordination, action amplitude, and correction of common mistakes.

[0140] c) Real-time supervision and interaction

[0141] Use a motion sensor (wearable device) to detect the patient's movement posture, duration, and frequency; if significant deviations or insufficient training are detected, the platform will automatically remind and, if necessary, notify the rehabilitation therapist to intervene remotely. And data recording and feedback, the data of each training (accuracy, completion, etc.) are uploaded in real time, and a statistical report is automatically generated. Patients can view the training effect and whether they have achieved their daily goals through the interface, enhancing their enthusiasm for self-management.

[0142] d) Real-time monitoring and dynamic adjustment, and perform periodic imaging reexamination and program update.

[0143] Use a mobile terminal or wearable device sensor to record information such as the patient's training time, action completion, and heart rate changes to complete the collection of training data; the platform automatically generates a visual report for patients and rehabilitation therapists to view.

[0144] Take an X-ray again during follow-up and automatically measure the Cobb angle. If the degree of scoliosis changes, dynamically adjust the training actions and intensity of the plan. If the degree of scoliosis does not change, maintain the current plan to achieve closed-loop management of rehabilitation.

[0145] In the above intelligent diagnosis and classification system for scoliosis based on medical images, the automatic measurement of the Cobb angle, intelligent classification, and digital remote rehabilitation of medical images are deeply integrated to construct a complete "diagnosis-rehabilitation-follow-up" closed-loop management system. Among them, from image acquisition, denoising and enhancement to Cobb angle calculation, and then to classification and rehabilitation plan push, can be completed within the same system, reducing errors or delays caused by multi-link manual operations and improving the efficiency of diagnosis and rehabilitation training. Combine conservative treatment with fine classification. For non-surgical scoliosis patients, more attention is paid to the refined division of the chest, thoracolumbar, lumbar and other regions in the classification, which can better guide the targeted allocation of rehabilitation actions. At the same time, with the help of a digital platform, long-term and effective remote supervision and personalized guidance for patients can be realized, reducing treatment interruptions or incomplete rehabilitation caused by geographical and time limitations. This model helps to solve problems such as insufficient medical resources and inconvenience for patients to travel back and forth, and is especially suitable for large-scale clinical screening and home rehabilitation scenarios, with important clinical application and promotion value.

[0146] The personalized digital rehabilitation training module can automatically match the corresponding specific scoliosis training (PSSE) action combinations and intensities in the database according to different classifications and Cobb angle ranges. If the Cobb angle is large or the curvature progresses rapidly, the training intensity can be appropriately increased or combined with brace correction.

[0147] The digital rehabilitation platform can push personalized training plans, video demonstrations, and precautions to patients through the network or mobile applications. If insufficient training frequency or movement accuracy is detected, the platform will automatically remind the patient and notify the therapist for remote intervention. Data such as the patient's training time, movement completion, and sensor feedback are recorded to form a visual report. During follow-up, X-ray films are taken again, the Cobb angle is automatically measured again and the classification is updated. If the condition changes, the training plan is dynamically adjusted to achieve closed-loop management.

[0148] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

Claims

1. An intelligent diagnosis and classification system for scoliosis based on medical images, characterized in that, It includes a medical image acquisition and preprocessing module, an automated Cobb angle measurement and calculation module, a scoliosis intelligent classification module, and a personalized digital rehabilitation training module; The medical image acquisition and preprocessing module is used to collect and store the anteroposterior X-ray images of the patient's spine in the upright position, and preprocess the X-ray images; The automated Cobb angle measurement and calculation module is used to calculate the Cobb angle of the preprocessed X-ray image parameters through a deep learning model, and mark the upper and lower end vertebrae and tangents; The scoliosis intelligent classification module is used to comprehensively determine and classify scoliosis by combining the number of scoliosis curves and the vertebral segment where the apex vertebra is located; The personalized digital rehabilitation training module can match the corresponding rehabilitation training plan according to the classification category.

2. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 1, wherein The medical image acquisition and preprocessing module includes full-spine anteroposterior X-ray photography and image preprocessing; For medical image acquisition, the patient is required to stand upright facing the X-ray machine, with feet shoulder-width apart, keeping the pelvis basically horizontal, avoiding posture deviation and rotation. Subsequently, the X-ray machine takes and stores the image to obtain the X-ray film; Image preprocessing performs preprocessing on the X-ray film in sequence, including denoising, histogram equalization enhancement, and highlighting of asymmetric features.

3. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 2, characterized in that The denoising process includes median filtering and edge protection.

4. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 1, wherein The automated Cobb angle measurement and calculation module includes deep learning model construction, curve fitting and end vertebra determination, and Cobb angle calculation: The deep learning model is constructed based on the CNN module and the ViT module, and the deep learning model is trained and optimized; Based on the deep learning model, key vertebral anchor points are identified and curves are fitted.

5. An intelligent diagnosis and classification system for scoliosis based on medical images according to claim 4, characterized in that, The identification and curve fitting of key vertebral anchor points by the deep learning model specifically include identification of key vertebral anchor points, cubic spline curve fitting, and positioning of upper and lower end vertebrae and identification of scoliosis segments.

6. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 4, characterized in that, The Cobb angle calculation is used to calculate the tangent slopes of the upper and lower edges of the end vertebrae; After identifying the upper and lower end vertebra positions, directly call the coordinates of the "upper and lower edge corner points" of the vertebra output by the deep learning model to perform linear fitting on the upper and lower edge points (x u1 , y u1 ), (x u2 , y u2 ) to obtain the tangent slope k1 of the upper end vertebra; for the upper and lower edge points (x d1 , y d1 ), (x d2 , y d2 ) of the lower end vertebra, perform linear fitting to obtain the tangent slope k2 of the lower end vertebra; Cobb angle calculation formula: Among them, k1 and k2 respectively represent the tangent slopes of the upper end vertebra and the lower end vertebra; then the calculation result of the radian value of the Cobb angle is converted into an angle value; The conversion formula is:

7. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 1, characterized in that, The scoliosis intelligent classification module needs to preset classification principles, and then pre-classify scoliosis into the following four major types and several subtypes based on the classification principles, and input them into the deep learning model as subsequent automatic classification discriminant rules, and output the classification result after determination.

8. An intelligent diagnosis and classification system for scoliosis based on medical images according to claim 1, characterized in that, The personalized digital rehabilitation training module includes a digital rehabilitation platform.

9. The intelligent diagnosis and classification system for scoliosis based on medical images according to claim 8, wherein, The digital rehabilitation platform is used to construct a PSSE training library, match and generate a personalized training plan according to the classification result and the PSSE training library, monitor in real time and adjust dynamically, and conduct phased imaging reexamination and plan update.

Images