X-ray-based sacroiliac arthritis determination method and system

Abstract

The invention relates to the technical field of medical detection, in particular to a sacroiliac arthritis judgment method and system based on X-rays, and the method comprises the steps of multi-modal data synchronous collection, image data real-time optimization and control, cross-modal causal reasoning and three-dimensional reconstruction, and AI-based lesion preliminary analysis and dynamic grading and interpretability report generation. According to the method, a real-time image processing technology is integrated in a data acquisition stage, an original image is directly and dynamically optimized at an imaging end, and the image resolution and the detail definition are improved through an advanced enhancement algorithm; the problems of delay and detail loss caused by relying on post-processing in a traditional method are solved, so that a doctor can obtain a high-quality diagnosis image in real time; by means of the optimization, the requirement for repeated shooting is reduced, and the availability and diagnosis efficiency of images are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of medical detection technology, and in particular to an X-ray-based sacroiliitis determination method and system. Background Art

[0002] Sacroiliitis is a common type of chronic inflammatory joint disease. Early diagnosis and accurate assessment are crucial for slowing disease progression and improving patient outcomes. Currently, plain radiographs remain the preferred imaging modality in clinical practice due to their ease of use and low cost. However, conventional radiographic imaging techniques have significant drawbacks. Firstly, their spatial resolution is limited by the two-dimensional projection principle and quantum noise interference, making it difficult to clearly visualize early lesions (such as micro-erosions of the cortical bone and fractures of the trabecular texture), resulting in a high rate of missed diagnosis of grade I-II lesions of 25%-30%. Secondly, fixed protocols for imaging parameters (such as kVp and mAs) cannot adapt to individual patient differences (e.g., obese patients require higher penetration due to differences in tissue thickness, and pediatric patients require lower radiation doses), often leading to unstable image quality or radiation dose overload. The cumulative radiation risk is particularly important in the management of chronic diseases requiring multiple follow-up visits. In addition, in the existing diagnostic process, the collection and analysis of imaging data and biomechanical parameters (such as plantar pressure distribution during the gait cycle and the dynamic motion trajectory of the pelvis) have been separated for a long time. Clinicians can only rely on experience to infer the relationship between mechanical abnormalities and joint lesions. There is a lack of systematic evidence support, resulting in a lack of precise basis for the formulation of early intervention strategies (such as customized orthotic devices and rehabilitation training).

[0003] In recent years, although three-dimensional imaging technologies (such as CT and MRI) can provide three-dimensional anatomical information about joints, their high equipment costs, complex operating procedures, and high radiation exposure from CT (a single dose can reach 3-5 mSv) have limited their widespread adoption in primary healthcare settings. Furthermore, existing AI-based medical image analysis models often focus on single-modality data processing. While this has improved lesion detection sensitivity, limitations in their algorithmic architecture hinder the real-time simultaneous integration of biomechanical data and imaging features, further hindering causal explanation of the underlying mechanisms linking mechanical load and bone degeneration. For example, some studies have attempted to combine pressure sensor data with X-ray images. However, the lack of temporal alignment (e.g., misalignment between gait phase and image exposure) has resulted in significant correlation analysis errors, making it impossible to accurately reveal the spatiotemporal correspondence between stress concentration areas and bone microlesions. Furthermore, existing three-dimensional reconstruction techniques often rely on multi-view X-ray or CT data. Single-view reconstruction suffers from deformation distortion and poor anatomical structure reproduction, making it difficult to meet clinical needs for three-dimensional quantitative assessment of minimal lesions.

[0004] Therefore, there is an urgent need for a comprehensive diagnostic solution that can achieve high-resolution imaging at low radiation doses, synchronously integrate biomechanical and imaging data, and support three-dimensional visual causal analysis, so as to break through the technical bottlenecks of traditional methods in early lesion detection rate, radiation safety, multimodal data fusion and etiology tracing, and provide reliable technical support for the precise diagnosis and treatment of sacroiliitis. Summary of the Invention

[0005] In order to overcome the problems raised in the above background technology, the present invention proposes an X-ray-based sacroiliitis determination method and system.

[0006] The technical solution of the present invention is: a method for determining sacroiliitis based on X-ray, comprising the following steps: S11: Multimodal data is collected synchronously through data acquisition equipment. The collected multimodal data includes X-ray images, biomechanical data and clinical texts. The biomechanical data includes obtaining the plantar pressure distribution during the gait cycle through the pressure sensing floor and measuring the pelvic tilt angle and trunk rotation angle using the IMU sensor. By collecting multimodal data such as X-ray images, biomechanical data and clinical texts, information related to the sacroiliac joint can be obtained from different angles. X-ray images provide anatomical structure information of the joint, biomechanical data reflects the mechanical state of the joint during movement, and clinical texts contain important information such as the patient's symptoms and medical history. The combination of multimodal data provides a rich and comprehensive information basis for the subsequent accurate determination of sacroiliitis. S12: Real-time optimization and control of image data. While collecting X-ray images, edge computing technology is used to optimize the collected X-ray images in real time and to adjust the X-ray parameters in real time. Using edge computing technology for real-time optimization while collecting X-ray images enables timely image noise reduction and enhancement, effectively removing artifacts and interference factors in the images, improving image clarity and contrast, and making the structure and pathological characteristics of the sacroiliac joint more clearly displayed, thereby improving diagnostic accuracy. S13: Cross-modal causal reasoning and 3D reconstruction: Using input biomechanical data to build a multimodal Transformer model, and constructing a 3D joint model based on X-ray images. This multimodal Transformer model, built using input biomechanical data, can mine potential causal relationships and correlation patterns between biomechanical data and X-ray images. Through this cross-modal causal reasoning, valuable information can be extracted from data of different modalities, revealing pathological characteristics and patterns that are difficult to detect with single-modal data, providing a more in-depth basis for the diagnosis of sacroiliitis. S14: Preliminary AI-based lesion analysis: Analyze the constructed multimodal Transformer model, and then use the analysis results to analyze and identify lesions in the 3D joint model constructed based on X-ray images. The constructed multimodal Transformer model uses artificial intelligence algorithms to quickly process large amounts of multimodal data, automatically extract features and patterns in the data, and make a preliminary judgment on the presence and type of sacroiliac joint lesions. This AI-based analysis method greatly improves diagnostic efficiency and reduces the subjectivity and errors of manual analysis. S15: Dynamic grading and interpretable report generation combines the results of cross-modal causal analysis and 3D joint model analysis to perform dynamic grading, and uses causal inference algorithms to generate visual reports. Dynamic grading, combining the results of cross-modal causal analysis and 3D joint model analysis, allows for personalized assessment and grading of sacroiliitis based on the patient's specific condition, taking into account multiple factors. This dynamic grading helps doctors develop more precise treatment plans, track disease progression, and adjust treatment measures in a timely manner.

[0007] As a preference, when performing synchronous multimodal data acquisition, the specific process is as follows: S21: Structured clinical text entry: The patient's voice interview data is recorded and transcribed in real time using a speech recognition engine. Key entities, including the temporal characteristics, location, and nature of the symptoms, are extracted using the BERT-Med model. Extracting key entities, such as the temporal characteristics, location, and nature of the symptoms, based on the BERT-Med model, can accurately filter out key content valuable for sacroiliitis diagnosis from a large amount of text information. This helps doctors quickly understand the patient's core condition, providing an important basis for subsequent diagnosis and treatment, while also facilitating the structured storage and analysis of text information. S22: Biomechanical data acquisition: The patient stands on an intelligent biomechanical acquisition platform. The platform's built-in high-density piezoelectric pressure sensor array captures the dynamic distribution of plantar pressure during the gait cycle in real time at a sampling rate of 100Hz. The IMU sensor also collects data on the patient's pelvic tilt angle and trunk rotation angle. This data, combined with plantar pressure data, allows for a more comprehensive assessment of the biomechanical function of the sacroiliac joint, providing richer information for diagnosing sacroiliitis. S23: Synchronous X-ray imaging, using an X-ray imaging system, triggers capture at key phases of the gait cycle, including heel strike. This captures the anatomical structure of the sacroiliac joint under specific motion conditions. Capturing at these critical phases allows for the identification of joint morphological changes under varying loads, helping to identify lesions that are difficult to detect in static conditions and improving diagnostic accuracy. S24: Multimodal data fusion and synchronization verification uses the IEEE1588 precision time protocol to provide a unified time reference for all devices through a master clock. It also runs an anomaly detection algorithm in real time to detect data collection loss. This allows for timely detection of problems that arise during the data collection process. Once data loss is detected, remedial measures can be taken promptly to ensure data integrity and reliability, improving the stability and accuracy of the entire collection system. S25: Quality Verification and Abnormal Data: This system verifies outliers and missing values ​​in real time, automatically re-collecting and recalibrating data based on pre-set logic. This system verifies outliers and missing values ​​in real time, enabling timely detection of errors and irregularities in the data. Automatic re-collection and recalibration based on pre-set logic effectively improves data quality and ensures the accuracy of subsequent analysis.

[0008] Preferably, when collecting biomechanical data, the following steps are specifically included: S31: Equipment initialization and data acquisition: Initialize and calibrate the high-density piezoelectric pressure sensor array built into the intelligent biomechanical acquisition platform. Then, clean the skin on the patient's bilateral iliac crests and sacral midline. Use medical double-sided tape to secure the IMU sensor to ensure anatomical alignment with bony landmarks before data acquisition. Initializing and calibrating the high-density piezoelectric pressure sensor array built into the intelligent biomechanical acquisition platform ensures that the sensor is in an accurate working state before data acquisition, reducing data deviations caused by sensor errors, thereby improving the accuracy of the collected plantar pressure data. S32: Data preprocessing: Applying a sliding average filter to denoise the collected data. Using a bicubic interpolation algorithm, the discrete sensor data is converted into a continuous pressure distribution matrix. The data is then normalized to generate a dynamic heat map. Applying a sliding average filter to denoise the collected data effectively removes random noise and interference signals, making the data smoother and more stable, thereby improving data quality. This facilitates more accurate analysis of plantar pressure distribution and pelvic motion characteristics. S33: IMU data fusion and posture solution, using the Madgwick filter algorithm to update the quaternion, calculate the real-time pelvic posture during the gait cycle, and then convert the IMU raw data to the pelvic anatomical coordinate system. The principle formula for calculating the real-time pelvic posture during the gait cycle is: ; in, is the fusion coefficient, is the angular velocity vector at the current moment, is the posture representation at the current moment, is the posture at the previous moment, is the sampling time interval, is the quaternion multiplication operator, is the gradient vector, is the norm of the gradient vector; the Madgwick filter algorithm is used to update the quaternion and calculate the real-time pelvic posture during the gait cycle. This algorithm can integrate acceleration, angular velocity, and other data collected by the IMU sensor to accurately estimate the posture changes of the pelvis in three-dimensional space. By calculating the pelvic posture in real time, we can gain a deeper understanding of the mechanical state of the sacroiliac joint during movement, providing an important basis for analyzing the relationship between joint pathology and movement. S34: Extraction of key features, including plantar pressure features such as pressure center trajectory, pressure asymmetry index, and peak pressure area positioning, as well as pelvic kinematic features including coronal plane tilt angle dynamic range, trunk rotation angular velocity, and joint load symmetry, extracts plantar pressure features such as pressure center trajectory, pressure asymmetry index, and peak pressure area positioning. The pressure center trajectory can reflect the movement of plantar pressure during walking, the pressure asymmetry index can quantify the difference in plantar pressure between the left and right feet, and the peak pressure area positioning can determine the area of ​​the plantar that bears the greatest pressure. These features can comprehensively evaluate the plantar pressure distribution and provide important information for determining whether sacroiliitis affects the patient's gait and plantar mechanical characteristics. Among them, through the formula: and , calculate the coordinates of the pressure weighted center of each frame and generate the pressure center trajectory moving path curve, where, and are the horizontal and vertical coordinates of the plantar pressure center in the plane coordinate system, is the pressure value of the i-th pressure sensor point, and are the horizontal and vertical coordinates of the i-th pressure sensor point in the platform coordinate system, and n is the total number of pressure sensors; And through the formula: , calculate the pressure asymmetry index, where is the pressure asymmetry index, which reflects the percentage of the difference in plantar pressure between the left and right feet. is the average total pressure on the left sole during the stance phase of gait, is the average total pressure on the right sole during the stance phase of gait.

[0009] Preferably, when performing real-time optimization and control of image data, the following steps are specifically included: S41: DICOM data reception and parsing. After the X-ray imaging system is exposed, the original DICOM file containing pixel data, exposure parameters, and patient position labels is received and parsed to extract key information. After the X-ray imaging system is exposed, the original DICOM file containing pixel data, exposure parameters, and patient position labels is received and parsed to fully obtain various key information of the image. Pixel data is the basis of the image, exposure parameters reflect the conditions at the time of shooting, and patient position labels help to accurately understand the shooting angle and position of the image. This information is crucial for subsequent image optimization and analysis, ensuring a comprehensive understanding and processing of the image. S42: Non-local means denoising: traverses the image with an 11×11 pixel search window and calculates similarity weights for each 5×5 pixel block, eliminating salt and pepper noise caused by photon count fluctuations while preserving the sacroiliac joint edges. This denoising method considers the similarity between pixels in the image and smoothes noise through weighted averaging, while avoiding the image blurring that can occur with traditional denoising methods. S43: Contrast-limited adaptive histogram equalization divides the image into 8×8 local blocks, clips the grayscale histogram of each block to a maximum slope of 3.0, and uses bilinear interpolation to eliminate block boundary effects. This method adaptively enhances image contrast, making the different tissues and structures of the sacroiliac joint more clearly visible in the image, helping doctors more accurately observe the details of the joint. S44: Lightweight SRGAN super-resolution reconstruction uses a lightweight SRGAN model to perform super-resolution reconstruction of images. Using a lightweight SRGAN model to perform super-resolution reconstruction of images can convert low-resolution X-ray images into high-resolution images. High-resolution images provide richer details, allowing doctors to more clearly observe the subtle structures and pathological features of the sacroiliac joint, helping to improve diagnostic accuracy. S45: Tenengrad gradient evaluation and quality judgment, including applying the Sobel operator to the lower third of the sacroiliac joint area, calculating the Tenengrad value, and judging the image quality based on the Tenengrad value. The Tenengrad value calculation formula is: , , ; Among them, ROI is the region of interest, that is, the specific area of ​​the edge of the sacroiliac joint, N is the total number of pixels in the region of interest, is the horizontal gradient value, is the vertical gradient value, The Tenengrad value quantifies the image edge sharpness. Applying the Sobel operator to the lower third of the sacroiliac joint region calculates the Tenengrad value, which quantifies the image edge sharpness. As an objective image quality evaluation metric, the Tenengrad value provides a scientific basis for judging image quality and helps screen out poor-quality images for further processing or reshooting. S46: Adaptive exposure control and parameter resetting, using a random forest regression model to optimize and predict exposure parameters and resetting the parameters of the X-ray imaging system. Using a random forest regression model to optimize and predict exposure parameters automatically adjusts exposure parameters based on individual patient differences and the characteristics of the imaging site. This helps achieve a more appropriate exposure, avoids overexposure or underexposure, and improves image quality and diagnostic value. S47: Simulates high-dose imaging with multi-view comparison, outputting and comparing the original image, optimized original image, and simulated image after parameter resetting. This output and comparison visually demonstrates the impact of different parameter settings and optimization methods on image quality. This helps physicians and researchers better understand the effects of parameter adjustments and further optimize exposure parameters and image processing methods.

[0010] Preferably, when the lightweight SRGAN model is used to perform super-resolution reconstruction on an image, the structure of the lightweight SRGAN model is: A11: Generator structure, with input as 512×512 denoised and enhanced image and output as 1024×1024 high-definition image, including: The residual block is ×6, the number of channels of the residual block is 64, and it uses a 3×3 convolution kernel and a PReLU activation function; Sub-pixel convolution layer ×2, using 4x upsampling; A12: Discriminator, using a 5×5 convolution kernel, outputs a 30×30 true / false matrix, and the loss function of the discriminator is: ; Where L is the total loss, To combat losses, is the content loss, that is, the feature matching loss based on VGG-19.

[0011] Preferably, when performing cross-modal causal reasoning and 3D reconstruction, the following steps are specifically included: S51: X-ray image feature extraction, using a deep neural network model to extract features from X-ray images in layers; using a deep neural network model to extract features from X-ray images in layers, it can automatically learn feature information at different levels and scales within the image. Deep neural networks have powerful feature extraction capabilities and can capture subtle structures and pathological features in X-ray images that are difficult to detect with the naked eye, providing richer information for subsequent analysis; S52: Attention mechanism matching, associating biomechanical data with image features; Associating biomechanical data with image features, the attention mechanism can automatically focus on the key correlations between biomechanical data and image features. By assigning different attention weights, it highlights features that have a significant impact on causal reasoning and 3D reconstruction, making the association between multimodal data more accurate and effective. S53: Causal verification: Validating the reliability of an association through statistical methods can eliminate accidental factors and false associations, ensuring that the association between biomechanical data and imaging features is truly causal. This helps improve the accuracy of causal reasoning and provides a more reliable basis for subsequent diagnosis and treatment. S54: 3D reconstruction and feature annotation. Based on a CT database of the sacroiliac joints of healthy individuals, a standard anatomical template is constructed. Using neural radiation field technology, the 3D morphology of the joint is restored from a single X-ray. This provides physicians with an intuitive and realistic 3D sacroiliac joint model. The 3D model demonstrates the spatial structure of the joint, the relative position and morphology of its components, and helps physicians gain a more comprehensive understanding of joint pathology.

[0012] Preferably, when using a deep neural network model to extract features from X-ray images in layers, the method includes: A21: First layer, identifying the overall gap width of the sacroiliac joint; A22: The second layer analyzes the sharpness of the bone cortical edge; A23: The third layer detects the roughness or fracture of the trabecular texture; A24: The fourth layer, locates tiny sclerosis foci.

[0013] Preferably, when conducting preliminary analysis of lesions based on AI, the following steps are specifically included: S61: Biomechanically driven disease pre-screening: Based on the extracted key biomechanical features, a random forest-based pre-classifier is used to output preliminary disease hypotheses. Based on the extracted key biomechanical features, a random forest-based pre-classifier is used to output preliminary disease hypotheses. Biomechanical features can reflect the mechanical state of the sacroiliac joint during movement and are closely related to joint pathology. The pre-classifier can quickly make preliminary judgments on possible diseases, improve diagnostic efficiency, and provide direction for subsequent analysis. S62: Targeted lesion localization: Based on the preliminary symptom hypothesis output by the random forest-based pre-classifier, the lesion of the preliminary symptom hypothesis is located to obtain the region of interest, and the image features of the region of interest are enhanced; Based on the preliminary symptom hypothesis, the lesion is located to obtain the region of interest. This helps to focus subsequent analysis on the area where the lesion may exist, improving the accuracy of diagnosis. By enhancing the image features of the region of interest, the characteristics of the lesion can be more clearly displayed, making it easier for doctors to observe and judge; S63: Cascade classification and lesion grading: A cascade classification network is used to analyze X-ray images for disease classification and lesion grading. The cascade classification network can gradually extract image features and make comprehensive judgments on features at different levels, thereby more accurately identifying different disease types. This helps distinguish sacroiliitis from other similar diseases and improves diagnostic specificity. S64: 3D pathology quantification and risk modeling. Based on the marching cubes algorithm, the surface mesh of the eroded area is extracted from the 3D joint model, and the volume percentage and maximum erosion depth are quantified. Based on the marching cubes algorithm, the surface mesh of the eroded area is extracted from the 3D joint model, and the volume percentage and maximum erosion depth are quantified. 3D pathology quantification can more intuitively reflect the extent and scope of the lesion, providing physicians with more comprehensive lesion information. Indicators such as volume percentage and maximum erosion depth can serve as important references for assessing lesion severity. S65: Verification and Decision-Making: Verify the consistency of biomechanical predictions with imaging and 3D model results. If a discrepancy is detected, a cross-modal alignment verification process is initiated, feature weights are re-evaluated, and manual intervention is prompted. Finally, the grading, localization, and quantification data are integrated to generate a structured report that clearly annotates the lesion grade, spatial distribution, and clinical intervention recommendations. Verify the consistency of biomechanical predictions with imaging and 3D model results to ensure that the analysis results of different modalities are mutually confirmed. If a discrepancy is detected, a cross-modal alignment verification process is initiated, feature weights are re-evaluated, and manual intervention is prompted. This helps improve the reliability and accuracy of diagnostic results and avoid misdiagnosis due to inconsistent data.

[0014] Preferably, when using a cascade classification network to analyze X-ray images for disease classification and lesion grading, the dual-branch neural network used includes: A31: Input layer, used to input the joint embedding vector of biomechanical key features and X-ray image features; A32: The first-level random forest feature filter is used to select key discriminant features from the multimodal fusion features based on preliminary disease hypotheses, reduce dimensionality and remove noise; A33: The second-level dual-branch neural network is used to simultaneously perform lesion grade classification and spatial localization tasks, including: a. A modified New York grading network for global analysis of X-ray images, classifying lesions from grade 0 to grade IV based on selected features; b. Pixel-level lesion localization network, used to analyze the region of interest, generate high-resolution heat maps, and accurately mark the spatial distribution of bone erosion and sclerosis areas.

[0015] An X-ray-based sacroiliitis determination system comprises: Multimodal data acquisition module, used to synchronously collect X-ray, biomechanical and clinical text data, and ensure the time alignment and integrity of multimodal information; Image optimization and control module, used to improve X-ray image quality and dynamically optimize shooting parameters to ensure high definition and low radiation required for diagnosis; A cross-modal causal reasoning module is used to establish the causal relationship between biomechanical abnormalities and imaging lesions, revealing the dynamic association between mechanics and pathology; 3D reconstruction and feature annotation module, used to generate 3D joint models from single-view X-rays and annotate the features of X-ray images in the 3D joint models; Intelligent lesion analysis module for automatic grading and submillimeter lesion detection through cascade classification and spatial positioning; The explainability report generation module is used to integrate multimodal evidence chains and output structured diagnostic reports and interactive visual evidence.

[0016] Beneficial effects of the present invention: 1. Real-time Image Optimization and Resolution Enhancement: This invention integrates real-time image processing technology during the data acquisition phase, dynamically optimizing raw images directly at the imaging end. Advanced enhancement algorithms improve image resolution and detail clarity. This eliminates the delays and detail loss associated with traditional post-processing methods, enabling doctors to instantly obtain high-quality diagnostic images. This optimization not only reduces the need for repeated imaging but also significantly improves image usability and diagnostic efficiency. 2. Adaptive parameter adjustment reduces radiation dose: Traditional X-ray imaging uses fixed exposure parameters, which are difficult to adapt to individual patient differences and can easily lead to radiation waste or unstable imaging quality. This invention uses an intelligent algorithm to dynamically adjust imaging parameters, optimizing exposure conditions based on patient size, target area characteristics, and real-time quality feedback. While ensuring image clarity, the radiation dose is significantly reduced, ensuring patient safety and improving the sustainability of examinations. 3. Biomechanical feature screening and precise analysis: The system efficiently screens key features from biomechanical data and combines them with imaging information to precisely locate lesions. While traditional methods require processing massive amounts of redundant data, this invention significantly reduces computational complexity through intelligent feature selection. This improves lesion identification accuracy while reducing hardware resource consumption, making the analysis process more efficient and targeted. 4. 3D Model Intuitive Visual Screening: A high-precision 3D model is generated based on single-view images, allowing physicians to intuitively assess joint structure through multi-angle observation and interactive operation. This model breaks through the limitations of traditional 2D images and clearly displays the spatial distribution and three-dimensional morphology of bone erosion and sclerosis, significantly improving screening efficiency and diagnostic confidence for complex lesions. 5. Synchronous biomechanical and imaging analysis to deduce etiology: By synchronously analyzing biomechanical abnormalities and imaging findings, the system reveals the dynamic relationship between mechanical load and pathological changes. This overcomes the drawback of traditional methods that view the two types of data in isolation. Doctors can intuitively understand the etiology (e.g., how gait abnormalities lead to localized bone damage), thereby developing personalized treatment plans and improving the accuracy and scientific nature of diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shown is a flow chart of the X-ray-based sacroiliitis determination method of the present invention; Figure 2 Shown is a schematic diagram of the structure of the X-ray-based sacroiliitis determination system of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] See also Figure 1-2 The present invention provides an embodiment: a method for determining sacroiliitis based on X-ray, comprising the following steps: A method for determining sacroiliitis based on X-rays comprises the following steps: Step 1: Multimodal data is collected synchronously through data acquisition equipment. The collected multimodal data includes X-ray images, biomechanical data, and clinical text. The biomechanical data includes obtaining the plantar pressure distribution during the gait cycle through the pressure sensing floor and measuring the pelvic tilt angle and trunk rotation angle using the IMU sensor. The specific process is as follows: Clinical text structured input: record the patient's voice data during consultation, transcribe it in real time through the speech recognition engine, and extract key entities based on the BERT-Med model, including the temporal characteristics of the symptoms, the location of the symptoms, and the nature of the symptoms; Biomechanical data collection: The patient stands on the intelligent biomechanical acquisition platform. The platform uses the built-in high-density piezoelectric pressure sensor array to capture the dynamic distribution of plantar pressure during the gait cycle in real time at a sampling rate of 100Hz. The IMU sensor also collects the patient's pelvic tilt angle and trunk rotation angle data. Specifically: ① Equipment initialization and data acquisition: initialize and calibrate the high-density piezoelectric pressure sensor array built into the intelligent biomechanical acquisition platform, then clean the patient's bilateral iliac crests and sacral midline skin, use medical double-sided tape to fix the IMU sensor, ensure anatomical alignment with bony landmarks, and then collect data; ② Data preprocessing: apply sliding average filtering to denoise the collected data, use bicubic interpolation algorithm to convert discrete sensor data into a continuous pressure distribution matrix, and then normalize the data to generate a dynamic heat map; ③ S33: IMU data fusion and posture solution: use Madgwick filtering algorithm to update quaternions, calculate the real-time pelvic posture during the gait cycle, and then convert the IMU raw data to the pelvic anatomical coordinate system. The principle formula for calculating the real-time pelvic posture during the gait cycle is: ;in, is the fusion coefficient, is the angular velocity vector at the current moment, is the posture representation at the current moment, is the posture at the previous moment, is the sampling time interval, is the quaternion multiplication operator, is the gradient vector, is the norm of the gradient vector; ④ key feature extraction, extracting plantar pressure features including pressure center trajectory, pressure asymmetry index and peak pressure area positioning, as well as pelvic kinematic features including dynamic range of coronal plane tilt angle, trunk rotation angular velocity and joint load symmetry, wherein, through the formula: and , calculate the coordinates of the pressure weighted center of each frame and generate the pressure center trajectory moving path curve, where, and are the horizontal and vertical coordinates of the plantar pressure center in the plane coordinate system, is the pressure value of the i-th pressure sensor point, and are the horizontal and vertical coordinates of the i-th pressure sensor point in the platform coordinate system, n is the total number of pressure sensors; and the formula is: , calculate the pressure asymmetry index, where is the pressure asymmetry index, which reflects the percentage of the difference in plantar pressure between the left and right feet. is the average total pressure on the left sole during the stance phase of gait, is the average total pressure on the right sole during the stance phase of gait; X-ray images are taken synchronously, using an X-ray imaging system to trigger the capture at key phases of the gait cycle, where the key phase of the gait cycle includes the moment when the heel strikes the ground; Multimodal data fusion and synchronization verification uses the IEEE1588 precision time protocol to provide a unified time reference for all devices through a master clock, and runs anomaly detection algorithms in real time to detect data collection loss; Quality inspection and abnormal data, real-time verification of abnormal and missing values ​​in the data, and automatic re-sampling and recalibration according to preset logic.

[0020] Step 2: Real-time optimization and control of image data: While collecting X-ray images, edge computing technology is used to optimize the collected X-ray images in real time and to adjust the X-ray parameters in real time, including: DICOM data reception and parsing: After the X-ray imaging system is exposed, the original DICOM file containing pixel data, exposure parameters, and patient position labels is received and parsed to extract key information; Non-local means denoising, which traverses the image with an 11×11 pixel search window and calculates similarity weights for each 5×5 pixel block, removing salt and pepper noise caused by photon count fluctuations while preserving the sacroiliac joint edges; Contrast-limited adaptive histogram equalization divides the image into 8×8 local blocks. The grayscale histogram of each block is clipped to a maximum slope of 3.0, and bilinear interpolation is used to eliminate block boundary effects. Lightweight SRGAN super-resolution reconstruction uses a lightweight SRGAN model to perform super-resolution reconstruction of images. The structure of the lightweight SRGAN model includes: a generator structure, which inputs a 512×512 denoised and enhanced image and outputs a 1024×1024 high-definition image. It includes: a residual block × 6, the number of channels of the residual block is 64, using a 3×3 convolution kernel and a PReLU activation function; a sub-pixel convolution layer × 2, using 4x upsampling. The discriminator uses a 5×5 convolution kernel and outputs a 30×30 true-false matrix. The discriminator loss function is: ; Where L is the total loss, To combat losses, is the content loss, i.e., the feature matching loss based on VGG-19; Tenengrad gradient evaluation and quality judgment: Apply the Sobel operator to the lower third of the sacroiliac joint area to calculate the Tenengrad value, and judge the image quality based on the Tenengrad value. The calculation formula of the Tenengrad value is: , , ; ROI is the region of interest, which specifies the specific area at the edge of the sacroiliac joint, and N is the total number of pixels in the region of interest. is the horizontal gradient value, is the vertical gradient value, is the Tenengrad value, which quantifies the edge sharpness of the image; Adaptive exposure control and parameter resetting, using random forest regression models to optimize and predict exposure parameters and resetting parameters of X-ray imaging systems; Simulate high-dose imaging multi-view comparison, output original images, optimized original images, and simulated images after parameter reset, and compare them; Step 3: Cross-modal causal reasoning and 3D reconstruction: Using input biomechanical data to build a multimodal Transformer model and constructing a 3D joint model based on X-ray images. Specifically, X-ray image feature extraction uses a deep neural network model to extract features from X-ray images in layers: the first layer identifies the overall gap width of the sacroiliac joint; the second layer analyzes the sharpness of the bone cortical edge; the third layer detects the roughness or fracture of the trabecular texture; and the fourth layer locates tiny sclerosis foci. Attention mechanism matching to associate biomechanical data with image features; Causal verification, which uses statistical methods to verify the reliability of the association; 3D reconstruction and feature annotation, based on the sacroiliac joint CT database of healthy people, build a standard anatomical template, and use neural radiation field technology to restore the 3D shape of the joint from a single X-ray Step 4: The AI-based preliminary lesion analysis analyzes the constructed multimodal Transformer model, and then uses the obtained analysis results to analyze and identify lesions in the 3D joint model built based on X-ray images. Specifically, the following are included: Biomechanically driven disease pre-screening: Based on the extracted biomechanical key features, a random forest-based pre-classifier is used to output preliminary disease hypotheses; Targeted lesion localization: Based on the preliminary disease hypothesis output by the random forest-based pre-classifier, the lesions of the preliminary disease hypothesis are located to obtain the region of interest, and the image features of the region of interest are enhanced; Cascade classification and lesion grading uses a cascade classification network to analyze X-ray images for disease classification and lesion grading. The two-branch neural network includes: an input layer for inputting a joint embedding vector of key biomechanical features and X-ray image features; a first-level random forest feature filter for filtering key discriminant features from multimodal fusion features based on preliminary disease hypotheses, reducing dimensionality and removing noise; a second-level two-branch neural network for simultaneously performing lesion grade classification and spatial localization tasks, including: a. a modified New York standard grading network for global analysis of X-ray images and classification of lesions from grade 0 to grade IV based on selected features; b. a pixel-level lesion localization network for analyzing regions of interest, generating high-resolution heat maps, and accurately marking the spatial distribution of bone erosion and sclerosis areas; Three-dimensional pathology quantification and risk modeling: Based on the marching cubes algorithm, the surface mesh of the eroded area is extracted from the 3D joint model to quantify the volume percentage and maximum erosion depth; Verification and decision-making: Verify the consistency of biomechanical predictions with imaging and 3D model results. If a contradiction is detected, the cross-modal alignment verification process is initiated, feature weights are re-evaluated, and manual intervention is prompted. Finally, the grading, positioning, and quantification data are integrated to generate a structured report that clearly marks the lesion grade, spatial distribution, and clinical intervention recommendations.

[0021] Step 5: Dynamic grading and interpretable report generation combines the analysis results of cross-modal causal analysis and 3D joint models for dynamic grading, and uses causal reasoning algorithms to generate visual reports.

[0022] Example 1 Consider a 32-year-old male professional soccer player who presents with a three-month history of intermittent low back pain, with morning stiffness lasting approximately 30 minutes and no nighttime awakenings. Clinical examination reveals tenderness in the left sacroiliac joint, and laboratory tests reveal a mildly elevated CRP level (8 mg / L) and a negative HLA-B27.

[0023] Specific steps: Multimodal data acquisition: The patient stood on an intelligent biomechanical platform, with IMU sensors fixed to the iliac crest and midline of the sacrum. Gait analysis revealed a peak heel strike pressure of 240 kPa on the left side (180 kPa on the right side), and a fluctuation in the coronal pelvic tilt of 3.2° (normal <2°). X-rays were acquired at mid-gait with a dose of 0.28 mSv. The original images showed mild blurring of the left joint space.

[0024] Image optimization and reconstruction: Edge computing nodes performed real-time denoising and super-resolution to 1024×1024, improving joint space clarity and cortical bone edge sharpness (Tenengrad value) from 45 to 62. NeRF-generated 3D models revealed a rough trabecular texture in the lower third of the left joint (CT value difference >100 HU).

[0025] AI analysis: Random forest analysis identified "excessive left-sided pressure and trabecular fracture" as key features, and a two-branch network classification indicated modified New York Class I (68% probability). 3D quantification revealed erosion volume accounting for 4.2% with a maximum depth of 0.9 mm, which did not meet the Class II threshold (5%).

[0026] Report and Intervention: The system indicated, "Biomechanical abnormalities (left / right pressure ratio 1.33) predated imaging findings. Custom orthotic insoles were recommended, with follow-up in three months." The doctor adjusted the rehabilitation plan based on the report. Six-month follow-up showed symptom relief, with no progression to Grade II.

[0027] Effect: Through biomechanics-imaging correlation analysis, early risks can be identified before X-rays clearly show structural damage, thus avoiding missed diagnosis.

[0028] Example 2 Consider a 58-year-old woman with a two-year history of low back pain and morning stiffness, which worsens at night. Laboratory tests reveal normal CRP and positive HLA-B27. A plain radiograph reveals "bilateral sacroiliac joint blurring," which cannot clearly distinguish between ankylosing spondylitis and degenerative spondylitis.

[0029] Specific steps: Multimodal data acquisition revealed symmetrical gait pressure (PAI = 8%), but IMU measurements indicated a trunk rotation rate of only 18° / s (normal > 30° / s). X-rays revealed bilateral joint space narrowing with irregular cortical bone margins.

[0030] Cross-modal reasoning: A multimodal Transformer found a strong correlation between "restricted trunk rotation" and "broken trabecular bone texture" (r=0.81). 3D reconstruction revealed asymmetrical distribution of bilateral sclerotic lesions (3 on the left and 1 on the right), with the largest diameter being 2.5 mm.

[0031] AI grading and early warning: The cascade classifier determined the patient to be Modified New York Class III (bilateral stenosis >50%). However, due to asymmetric sclerosis and biomechanical compensation, a high-risk alert for ankylosing spondylitis was triggered. The system automatically correlated the positive HLA-B27 result and recommended additional MRI assessment for active inflammation.

[0032] Clinical decision: MRI confirmed left-sided joint bone marrow edema, confirming ankylosing spondylitis. Based on the quantitative results of the 3D model, the physician selected biologic therapy. A 3D follow-up examination three months later showed a 12% reduction in the volume of the sclerotic lesions.

[0033] Effect: Breaking through the bottleneck of traditional X-ray diagnosis, the system clearly distinguishes inflammatory and degenerative lesions through mechanical compensation model and 3D pathological annotation, and guides precise treatment.

[0034] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for determining sacroiliitis based on X-ray, characterized by: The following steps are included: S11: Synchronous multimodal data acquisition: Multimodal data acquisition is performed using a data acquisition device. The acquired multimodal data includes X-ray images, biomechanical data, and clinical text. The biomechanical data includes obtaining plantar pressure distribution during the gait cycle through a pressure sensing floor and measuring the pelvic tilt angle and trunk rotation angle using an IMU sensor. S12: Real-time optimization and control of image data. While collecting X-ray images, edge computing technology is used to optimize the collected X-ray images in real time and to adjust the X-ray parameters in real time. S13: Cross-modal causal reasoning and 3D reconstruction, using input biomechanical data to build a multimodal Transformer model and construct a 3D joint model based on X-ray images; S14: Preliminary AI-based lesion analysis: Analyze the constructed multimodal Transformer model, and then use the obtained analysis results to analyze and identify lesions in the 3D joint model constructed based on X-ray images; S15: Dynamic grading and interpretability report generation, combining the analysis results of cross-modal causal analysis and 3D joint model for dynamic grading, and using causal reasoning algorithms to generate visual reports.

2. The X-ray-based sacroiliitis determination method according to claim 1, characterized in that: When performing synchronous multimodal data acquisition, the specific process is as follows: S21: Structured clinical text input: records the patient's voice data during consultation, transcribes it in real time through the speech recognition engine, and extracts key entities based on the BERT-Med model, including the temporal characteristics of the symptoms, the location of the symptoms, and the nature of the symptoms; S22: Biomechanical data acquisition: The patient stands on the intelligent biomechanical acquisition platform. The platform's built-in high-density piezoelectric pressure sensor array captures the dynamic distribution of plantar pressure during the gait cycle in real time at a sampling rate of 100 Hz. The IMU sensor also collects the patient's pelvic tilt angle and trunk rotation angle data. S23: synchronously capturing X-ray images, using an X-ray imaging system to trigger the capture at a key phase of the gait cycle, wherein the key phase of the gait cycle includes the moment of heel contact; S24: Multimodal data fusion and synchronization verification, using the IEEE1588 precision time protocol, provides a unified time reference for all devices through a master clock, and runs anomaly detection algorithms in real time to detect data collection loss; S25: Quality inspection and abnormal data, real-time verification of abnormal and missing values ​​in the data, and automatic re-sampling and recalibration according to preset logic.

3. The X-ray-based sacroiliitis determination method according to claim 2, characterized in that: When collecting biomechanical data, it specifically includes: S31: Equipment initialization and data acquisition: Initialize and calibrate the high-density piezoelectric pressure sensor array built into the intelligent biomechanical acquisition platform. Clean the skin of the patient's bilateral iliac crests and sacral midline. Use medical double-sided tape to secure the IMU sensor to ensure alignment with bony landmarks before data acquisition. S32: Data preprocessing: applying sliding average filtering to denoise the collected data, using bicubic interpolation algorithm to convert discrete sensor data into a continuous pressure distribution matrix, and then normalizing the data to generate a dynamic thermal map; S33: IMU data fusion and posture solution, using the Madgwick filter algorithm to update the quaternion, calculate the real-time pelvic posture during the gait cycle, and then convert the IMU raw data into the pelvic anatomical coordinate system; S34: Key feature extraction, extracting plantar pressure features including pressure center trajectory, pressure asymmetry index peak pressure area positioning, and pelvic kinematic features including coronal plane tilt angle dynamic range, trunk rotation angular velocity, and joint load symmetry.

4. The X-ray-based sacroiliitis determination method according to claim 3, characterized in that: When performing real-time optimization and control of image data, it specifically includes: S41: DICOM data reception and parsing: After the X-ray imaging system completes exposure, the original DICOM file containing pixel data, exposure parameters, and patient position labels is received and parsed to extract key information; S42: Non-local means denoising, which traverses the image with an 11×11 pixel search window and calculates similarity weights for each 5×5 pixel block to remove salt and pepper noise caused by photon count fluctuations while preserving the sacroiliac joint edges; S43: Contrast-limited adaptive histogram equalization, which divides the image into 8×8 local blocks. The grayscale histogram of each block is clipped to a maximum slope of 3.0, and bilinear interpolation is used to eliminate block boundary effects. S44: Lightweight SRGAN super-resolution reconstruction, using the lightweight SRGAN model to perform super-resolution reconstruction of images; S45: Tenengrad gradient evaluation and quality judgment: Apply the Sobel operator to the lower third of the sacroiliac joint area to calculate the Tenengrad value, and judge the image quality based on the Tenengrad value; S46: Adaptive exposure control and parameter resetting, using random forest regression model to optimize and predict exposure parameters and resetting parameters of X-ray imaging system; S47: Simulate high-dose imaging multi-view comparison, output the original image, the optimized original image, and the simulated image after parameter reset, and compare them.

5. The X-ray-based sacroiliitis determination method according to claim 4, characterized in that: When using the lightweight SRGAN model to perform super-resolution reconstruction on an image, the structure of the lightweight SRGAN model is: A11: Generator structure, with input as 512×512 denoised and enhanced image and output as 1024×1024 high-definition image, including: The residual block is ×6, the number of channels of the residual block is 64, and it uses a 3×3 convolution kernel and a PReLU activation function; Sub-pixel convolution layer ×2, using 4x upsampling; A12: Discriminator, using a 5×5 convolution kernel and outputting a 30×30 true / false matrix.

6. The X-ray-based sacroiliitis determination method according to claim 5, characterized in that: When performing cross-modal causal reasoning and 3D reconstruction, it specifically includes: S51: X-ray image feature extraction, using a deep neural network model to extract features from X-ray images in layers; S52: Attention mechanism matching to associate biomechanical data with image features; S53: Causal verification, verifying the reliability of the association through statistical methods; S54: 3D reconstruction and feature annotation. Based on the sacroiliac joint CT database of healthy people, a standard anatomical template is constructed, and the 3D morphology of the joint is restored from a single X-ray using neural radiation field technology.

7. The X-ray-based sacroiliitis determination method according to claim 6, characterized in that: When using a deep neural network model to extract features from X-ray images in layers, including: A21: First layer, identifying the overall gap width of the sacroiliac joint; A22: The second layer analyzes the sharpness of the bone cortical edge; A23: The third layer detects the roughness or fracture of the trabecular texture; A24: The fourth layer, locates tiny sclerosis foci.

8. The X-ray-based sacroiliitis determination method according to claim 7, characterized in that: When conducting preliminary AI-based lesion analysis, the following are included: S61: Biomechanically driven disease pre-screening, based on the extracted biomechanical key features, using a random forest-based pre-classifier to output preliminary disease hypotheses; S62: Targeted lesion localization: Based on the preliminary disease hypothesis output by the random forest-based pre-classifier, the lesion of the preliminary disease hypothesis is located to obtain the region of interest, and the image features of the region of interest are enhanced; S63: Cascade classification and lesion grading, using a cascade classification network to analyze X-ray images for disease classification and lesion grading; S64: 3D pathology quantification and risk modeling, based on the marching cubes algorithm, extracts the surface mesh of the eroded area from the 3D joint model, quantifies the volume percentage and maximum erosion depth; S65: Verification and decision-making: Verify the consistency of biomechanical predictions with imaging and 3D model results. If a contradiction is detected, initiate the cross-modal alignment verification process, re-evaluate feature weights, and prompt manual intervention; finally, integrate grading, positioning, and quantification data to generate a structured report that clearly marks the lesion grade, spatial distribution, and clinical intervention recommendations.

9. The X-ray-based sacroiliitis determination method according to claim 8, characterized in that: When using a cascade classification network to analyze X-ray images for disease classification and lesion grading, the two-branch neural network used includes: A31: Input layer, used to input the joint embedding vector of biomechanical key features and X-ray image features; A32: The first-level random forest feature filter is used to select key discriminant features from the multimodal fusion features based on preliminary disease hypotheses, reduce dimensionality and remove noise; A33: The second-level dual-branch neural network is used to simultaneously perform lesion grade classification and spatial localization tasks, including: a. A modified New York grading network for global analysis of X-ray images, classifying lesions from grade 0 to grade IV based on selected features; b. Pixel-level lesion localization network, used to analyze the region of interest, generate high-resolution heat maps, and accurately mark the spatial distribution of bone erosion and sclerosis areas.

10. An X-ray-based sacroiliitis determination system, used in an X-ray-based sacroiliitis determination method according to any one of claims 1 to 9, characterized in that: Includes: Multimodal data acquisition module, used to synchronously collect X-ray, biomechanical and clinical text data, and ensure the time alignment and integrity of multimodal information; Image optimization and control module, used to improve X-ray image quality and dynamically optimize shooting parameters to ensure high definition and low radiation required for diagnosis; A cross-modal causal reasoning module is used to establish the causal relationship between biomechanical abnormalities and imaging lesions, revealing the dynamic association between mechanics and pathology; 3D reconstruction and feature annotation module, used to generate 3D joint models from single-view X-rays and annotate the features of X-ray images in the 3D joint models; Intelligent lesion analysis module for automatic grading and submillimeter lesion detection through cascade classification and spatial positioning; The explainability report generation module is used to integrate multimodal evidence chains and output structured diagnostic reports and interactive visual evidence.

Images