System and method for intelligently evaluating three-dimensional posture and gradient of pelvis with single anterior-posterior X-ray film based on deep learning

Abstract

The invention discloses an intelligent evaluation system and method for the three-dimensional posture and gradient of a pelvis of a single anterior-posterior X-ray film based on deep learning, belongs to the crossing field of medical image analysis, artificial intelligence and bone science, and particularly relates to an intelligent evaluation system and method for the three-dimensional posture and gradient of the pelvis of the single anterior-posterior X-ray film based on deep learning. The system is characterized by comprising five modules: preprocessing (more than or equal to 1000 cases of AP sheets, and the signal-to-noise ratio is more than or equal to 30dB), U-Net + + ResNet 50 mark point detection (the error is less than or equal to 1.2 mm), three-dimensional reconstruction (800 cases of CT modeling, and RMSE is less than 2.0 mm), gradient calculation (the precision is less than or equal to 0.5 degree, and CT-MAE is less than 1.5 degree), and visualization (docking PACS). According to the method, end-to-end evaluation can be carried out through a single AP piece, radiation is reduced by 40%, consumed time is smaller than or equal to 2.5 minutes, and CV is equal to 3%-5%. The method is suitable for total hip replacement preoperative planning and the like, provides low-radiation and high-precision evaluation, assists in precise decision making and adapts to basic level and emergency treatment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image analysis, artificial intelligence (AI) and orthopedics, and specifically relates to a method and system for accurately and efficiently evaluating the three-dimensional posture and functional pelvic tilt of a patient from a single standing anteroposterior (AP) X-ray film using deep learning technology. The method and system are suitable for preoperative planning of total hip arthroplasty (THA), evaluation of pelvic fracture reduction, correction of spinal-pelvic deformity, and other scenarios, and are particularly suitable for low-radiation and high-efficiency evaluation requirements of emergency and primary medical institutions. BACKGROUND

[0002] Pelvic tilt (PT) is a crucial biomechanical parameter in orthopedic clinics, especially in preoperative planning of total hip arthroplasty (THA). Accurate evaluation of pelvic tilt helps guide prosthesis selection and optimize prosthesis placement angle, thereby effectively preventing common postoperative complications such as prosthesis impingement, dislocation, leg length discrepancy, and low back pain.

[0003] Traditional methods of measuring pelvic tilt mainly rely on lateral X-ray films. However, the acquisition of lateral X-ray films involves additional radiation exposure, with a radiation dose of 0.15 mSv per lateral X-ray film (0.09 mSv per AP X-ray film), which increases the patient's radiation exposure by 40%. In addition, traditional manual measurement methods, such as measuring through bony landmarks, have large inter-observer and intra-observer differences, with a coefficient of variation (CV) of manual measurement between observers reaching 18% to 25%, resulting in insufficient repeatability and reliability of the evaluation results. This subjectivity limits the accurate application of pelvic tilt parameters by clinicians in preoperative planning and postoperative evaluation.

[0004] With the rapid development of AI technology, especially the breakthrough of deep learning in medical image analysis, new opportunities have been provided to solve the above problems. Existing AI applications in the field of orthopedics have shown great potential, such as in fracture detection, classification, segmentation, and surgical planning assistance. However, there are few studies that apply deep learning to single anteroposterior X-ray films for three-dimensional posture reconstruction and accurate evaluation of functional pelvic tilt to replace or assist traditional lateral X-ray film measurement. Existing CT-based pelvic three-dimensional evaluation schemes have a single examination cost of about 800 yuan (AP X-ray film only 50 yuan), and CT examination takes 15 to 20 minutes, which is not suitable for THA emergency preoperative planning (results are required within 30 minutes). Therefore, developing an innovative AI-based solution to efficiently and accurately obtain pelvic three-dimensional posture and tilt information from a single anteroposterior X-ray film has important clinical significance and technical value. SUMMARY

[0005] The present application aims to overcome the deficiencies of the prior art, and provides a single anteroposterior X-ray film pelvis three-dimensional posture and inclination intelligent evaluation system and method based on deep learning. The present application constructs an efficient deep learning model, uses a single standing anteroposterior X-ray film, realizes accurate identification of key anatomical landmark points of the pelvis, reconstruction of the three-dimensional posture of the pelvis, and on this basis, calculates the functional pelvic inclination, thereby providing a low-radiation, high-precision and high-efficiency evaluation tool for orthopedic clinics.

[0006] The purposes of the present application include: (1) automatically and accurately identifying key anatomical landmark points of the pelvis, such as anterior superior iliac spine (ASIS) and pubic tubercle, from a single standing anteroposterior X-ray film, with an average error of less than 1.5 mm; (2) based on the identified two-dimensional landmark point information, combining the biomechanical model of the pelvis and the powerful feature extraction capability of deep learning, reconstructing the three-dimensional posture of the pelvis, with a root mean square error (RMSE) of less than 2.0 mm; (3) accurately calculating the functional pelvic inclination according to the reconstructed three-dimensional posture of the pelvis, and providing standardized and repeatable measurement results, with an average absolute error of less than 1.5° compared with the CT gold standard; (4) developing a user-friendly software system to integrate the above algorithms, making it convenient for clinicians to operate, and displaying the evaluation results in real time, supporting the interface with the hospital PACS system; (5) reducing the use of lateral X-ray films, reducing the radiation exposure of patients (40% less than traditional lateral films), improving the clinical work efficiency (evaluation time reduced to 2.3 minutes), and providing more accurate preoperative planning basis for total hip arthroplasty and other pelvis-related surgeries.

[0007] The core innovation of the present application lies in: breaking through the technical limitations of traditional pelvis evaluation relying on lateral X-ray films, creating an integrated solution of “single anteroposterior X-ray film + deep learning + three-dimensional reconstruction”, only using a single AP image, combining a three-dimensional statistical shape model (SSM) based on 800 CT data and a U-Net+++ResNet50 landmark detection architecture (positioning error ≤1.2 mm), realizing pelvis three-dimensional posture reconstruction (RMSE <2.0 mm) and functional inclination calculation (accuracy ≤0.5°), reducing 40% radiation exposure, and reducing evaluation time from 10 minutes to 2.3 minutes; at the same time, introducing biomechanical constraints (convergence error <0.3°) and standardized visualization modules (interface with PACS), solving the pain points of inconsistent manual measurement (CV from 18% to 25% to 3% to 5%) and low efficiency, providing low-radiation, high-precision and high-efficiency technical support for orthopedic precise preoperative planning.

[0008] The technical scheme of the present application is centered on the core goal of "single AP X-ray film to realize three-dimensional evaluation of pelvis", and five modular processes are constructed: ① Data preprocessing: collect ≥1000 cases containing labeled AP X-ray film, and process by gray scale normalization, denoising and the like to ensure that the signal-to-noise ratio is ≥30dB and the size is uniform to ≥2048×2048 pixels; ② Landmark point detection: adopt U-Net+++ResNet50 architecture to automatically identify 8 key anatomical landmark points (bilateral ASIS, pubic tubercle, etc.), and the coordinate positioning error is ≤1.2mm; ③ Three-dimensional reconstruction: take two-dimensional landmark points as input, combine the SSM model constructed by 800 CT and the PnP / MLP algorithm, introduce biomechanical constraints (ASIS-pubis joint distance deviation ≤3mm, etc.) for iterative optimization, and realize three-dimensional posture reconstruction (RMSE<2.0mm); ④ Inclination calculation: based on the angle between the "bilateral ASIS-pubic tubercle anatomical plane" and the vertical line of the ground, the functional pelvic inclination is calculated, and the accuracy is ≤0.5°, and the MAE with the CT gold standard is <1.5°; ⑤ Result visualization: develop an interactive interface supporting PACS docking, real-time display of landmark point labeling, three-dimensional model (rotatable and scalable) and quantitative report (PDF / XML export), and the evaluation time of a single sample is ≤2.5 minutes, and the radiation is reduced by 40% compared with the traditional scheme.

[0009] After the implementation of the present application, multiple significant positive effects can be produced: only through a single anteroposterior X-ray film, the evaluation can be completed, the radiation exposure is reduced by 40% compared with the traditional lateral film scheme, and it is more in line with the medical safety principle; the pelvic inclination evaluation time is shortened from 10 minutes of manual measurement to 2.5 minutes, which greatly improves the clinical efficiency, especially adapts to the preoperative planning needs of emergency; the average error of landmark point detection is <1.5mm, the three-dimensional reconstruction RMSE is <2.0mm, and the MAE with the CT gold standard is <1.5°, and the evaluation consistency (CV=3%-5%) is significantly better than manual measurement (18%-25%), which provides accurate parameters for total hip replacement and other orthopedic surgeries, and reduces the risk of postoperative complications; the system supports PACS docking and standardized report output, and the operation is simple and easy to promote, which can improve the pelvis evaluation ability of primary medical institutions, and has clinical safety, efficiency and precision. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 It is a schematic diagram of the principle of three-dimensional posture reconstruction of pelvis.

[0011] Figure 2 It is a schematic diagram of the correlation analysis of key parameters and inclination of pelvis. DETAILED DESCRIPTION

[0012] Example 1

[0013] 1. Data set construction and preprocessing implementation details

[0014] 1.1 Data collection Collect 1200 cases of clinical standing anteroposterior (AP) pelvic X-ray films, covering: 500 healthy controls (age 18-80 years, male to female ratio 1:1); 250 cases of total hip dysplasia (Crowe type I-IV); 200 cases of femoral head necrosis (ARCO stage II-IV); 150 cases of pelvic fracture (Tile type A-C); 100 cases of scoliosis combined with pelvic tilt (Cobb angle 15°-60°). All cases were collected synchronously CT data (slice thickness 0.625 mm, pitch 1.0), CT and X-ray film registration was realized by NiftyReg software (registration error ≤0.5 mm), 8 key landmark three-dimensional coordinates were labeled by 2 doctors with the title of deputy chief physician or above in the department of orthopedics, Kappa value = 0.93, inconsistent cases were determined by a third party chief physician.

[0015] 1.2 Image preprocessing algorithm Gray scale normalization: adaptive histogram equalization (CLAHE) was used, clip limit = 2.0, grid size 8x8, formula:

[0016] Where μ is the mean value of the image, σ is the standard deviation De-noising: combined with Gaussian filter (σ = 1.0) and non-local mean de-noising (search window 7x7, similar window 5x5), the signal-to-noise ratio is improved to 35±2dB.

[0017] Size normalization: bicubic interpolation to 2048x2048 pixels, preserving the original anatomical proportions.

[0018] 2. Deep learning landmark detection model implementation details

[0019] 2.1 Network architecture (U-Net+++ResNet50)

[0020] Encoder: use pre-trained ResNet50 as the backbone, take the output features of conv2_x~conv5_x (channel number 256, 512, 1024, 2048).

[0021] Decoder: 4 layers of transposed convolution (kernel = 3x3, step = 2), fuse the corresponding layer features of the encoder through skip connection.

[0022] Attention mechanism: introduce a channel attention module at each layer of the decoder,

[0023] Formula:

[0024] Where σ is the sigmoid function, and MLP contains one hidden layer (neuron number = channel number / 4).

[0025] Output layer: 8 landmark heatmaps (2048x2048x8) with Gaussian kernel activation (σ=3.0).

[0026] 2.2 Model training parameters

[0027] Loss function: Improved L1 loss + heatmap peak constraint

[0028]

[0029] where α=0.7, p i is the true coordinate, is the predicted coordinate, h pi is the heatmap peak, σ=2.0.

[0030] Optimizer: AdamW (β1=0.9, β2=0.999), weight decay 1e−5.

[0031] Training strategy: Initial learning rate 1e−4, decay 50% every 20 epochs, batch size=8 (mixed precision training), training epochs 120, early stopping threshold 25 epochs.

[0032] Data augmentation: Random rotation (±10°), scaling (0.85-1.15 times), brightness adjustment (±20%), elastic deformation (α=100, σ=10).

[0033] 2.3 Landmark post-processing

[0034] Heatmap peak detection: Non-maximum suppression (NMS window 5x5)

[0035] Sub-pixel level coordinate calculation: Fit a quadratic surface to the 3x3 area around the peak, formula:

[0036] where H is the Hessian matrix, is the gradient vector. Accuracy verification: 80 cases in the test set, average landmark error 1.12±0.35mm, maximum error 1.8mm.

[0037] 3. Pelvic three-dimensional pose reconstruction implementation details

[0038] 3.1 Construction of three-dimensional statistical shape model (SSM)

[0039] Training data: 800 CT three-dimensional reconstruction models (500 healthy + 300 pathological).

[0040] Model construction:

[0041]

[0042] ​(1) Align all models based on Procrustes analysis:

[0043] (2) Principal Component Analysis (PCA):

[0044] (3) Retain 95% shape variation, take the first 30 principal components, model shape error ≤0.8mm

[0045] 3.2 Pose estimation algorithm (improved PnP + MLP fusion)

[0046] PnP algorithm: core formula:

[0047] MLP regression correction: input features: 8 landmark points 2D coordinates (16 dimensions) + patient gender (1 dimension) + BMI (1 dimension) + age (1 dimension), total 19 dimensions network structure: input layer → 128 neurons (ReLU) → 64 neurons (ReLU) → output layer (12 dimensions: 9 parameters of rotation matrix + 3 parameters of translation vector),

[0048] Loss function:

[0049] Biomechanical constraint optimization:

[0050] The midpoint distance between the bilateral ASIS and the pubic symphysis is 95±5mm, the vertical distance from the sacral promontory to the pelvic inlet plane is 35±3mm, and the inclination angle of the acetabular roof is 15°-25°. The Levenberg-Marquardt algorithm is used for iterative optimization, and the convergence condition is that the parameter error of the adjacent iteration attitude is less than 0.3°.

[0051] 3.3 Reconstruction accuracy verification

[0052] ​​​​​​​​​​​​​​​The root mean square error (RMSE) of three-dimensional posture reconstruction of the test set 80 examples was 1.72±0.45 mm, and the average deviation from the CT direct measurement value was 0.85 mm.

[0053] 4. Functional pelvic tilt calculation implementation details

[0054] 4.1 Anatomical plane definition

[0055] Pelvic functional plane: four points fitted by bilateral ASIS (A1, A2) and bilateral pubic tubercle (P1, P2), plane equation:

[0056] ax+by+cz+d=0, where the normal vector n=(a,b,c) is calculated by four-point cross product.

[0057] 4.2 Tilt calculation

[0058] Ground vertical vector: =(0,0,1) (based on X-ray imaging coordinate system)

[0059] Tilt angle calculation formula:

[0060] Accuracy verification: compared with the CT gold standard, the mean absolute error (MAE) was 1.23±0.38°, and the 95% confidence interval was [0.52°, 1.94°].

[0061] 5. System deployment and performance indicators

[0062] Hardware configuration: Intel i7-13700K CPU+NVIDIA RTX 4090 GPU (24GB video memory).

[0063] Software environment: Python 3.10+PyTorch 2.0+VTK 9.2.

[0064] Processing efficiency: total time for a single sample was 2.1±0.3 minutes (preprocessing 25s+landmark detection 48s+three-dimensional reconstruction 32s+tilt calculation 15s+visualization 10s).

[0065] Clinical verification: 100 cases of total hip arthroplasty patients preoperative evaluation, the consistency with intraoperative fluoroscopy measurement value ICC=0.92, significantly better than manual measurement (ICC=0.68).

[0066] This embodiment shows the realizability of the invention through specific algorithm parameters, training strategies and verification data. The model architecture, formula derivation and performance indicators involved are verified by clinical data to ensure the stability and reliability of the technical solution.

[0067] In the drawings of the present application:Figure 1 It is a three-dimensional posture reconstruction principle diagram of pelvis, which shows the three-dimensional posture reconstruction principle of SSM model (shape variation degree ≤5%) based on 800 CT and PnP algorithm (core formula: sx=K[R|t]X) of pelvis, labels 8 key anatomical landmarks, clearly defines the biomechanical constraints such as ASIS-pubic symphysis distance deviation ≤3mm, and the precision indexes such as three-dimensional reconstruction RMSE<2.0mm and average deviation 0.85mm compared with C; Figure 2 It is a tilt calculation diagram, which shows the correlation between key anatomical parameters (such as bilateral ASIS and pubic tubercle position) of pelvis and tilt, contains parameter change curve and abnormal threshold label, and assists in understanding the internal relationship between tilt and anatomical structure, and provides reference for clinical evaluation.

[0068] The protection scope of the application focuses on the core technical system of "single front and back X-ray film pelvis three-dimensional posture and tilt intelligent evaluation based on deep learning", specifically covering: an integrated system for end-to-end evaluation based on single AP X-ray film (containing five modules of data preprocessing, deep learning landmark detection, pelvis three-dimensional posture reconstruction, tilt calculation and user interaction visualization, supporting PACS docking and report export); core method (landmark detection adopts U-Net+++ResNet50 architecture, three-dimensional reconstruction combines 800 CT constructed SSM model and PnP / MLP algorithm and introduces biomechanical constraints, and tilt is calculated based on the angle between ASIS-pubic tubercle plane and ground vertical line); key technology (precision control technology with landmark positioning error ≤1.2mm, three-dimensional reconstruction RMSE<2.0mm, and tilt MAE<1.5° compared with CT gold standard); and application in total hip arthroplasty preoperative planning, pelvic fracture reduction evaluation and other orthopedic scenes, which overall covers system architecture, core algorithm, technical parameters and clinical application scenarios, forming a complete technical protection system.

Claims

1. A deep learning-based intelligent assessment system and method for three-dimensional pelvic pose and tilt on a single anteroposterior X-ray film, characterized in that, include: (1) Data acquisition and preprocessing module, used to collect labeled standing anteroposterior (AP) X-ray images, perform grayscale normalization, image enhancement, denoising and size adjustment on the original X-ray images, and the signal-to-noise ratio of the processed X-ray images is ≥30dB and the size is unified to ≥2048×2048 pixels; (2) Deep learning pelvic landmark detection module, using U-Net+++ResNet50 architecture, outputs heat maps of 8 key landmarks (bilateral anterior superior iliac spines, bilateral pubic tubercles, bilateral acetabular roof, sacral promontory, and midpoint of pubic symphysis), and uses 80% of the maximum value of the heat map as the threshold to determine the coordinates, with a positioning error ≤1.2mm; (3) The pelvic three-dimensional posture reconstruction module takes the two-dimensional coordinates of the landmark points as input, combines the three-dimensional statistical shape model (SSM, shape variability ≤5%) constructed based on 800 CT data, estimates the three-dimensional rotation matrix and translation vector through PnP algorithm or MLP regression algorithm, introduces biomechanical constraints (bilateral ASIS-pubic symphysis distance deviation ≤3mm, sacral promontory angle 15°~25°) for iterative optimization, convergence error <0.3°, reconstruction RMSE <2.0mm; (4) Functional pelvic tilt calculation module, based on "bilateral ASIS The angle between the IS-pubic tubercle anatomical plane and the vertical line passing through the origin of the X-ray imaging coordinate system is calculated to determine the tilt, with an accuracy of ≤0.5°, which is less than 1.5° compared to the CT gold standard MAE; (5) User interaction and result visualization module, providing a graphical interface to support the uploading of APX films (DICOM / JPEG format, ≤10MB), rendering three-dimensional models through the VTK library (supporting ±360° rotation and 10%~200% scaling), and outputting PDF / XML reports containing marker coordinates and three-dimensional parameters, which can be connected to the PACS system through the HL7 protocol.

2. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 1, characterized in that, The dataset size of the data acquisition and preprocessing module is ≥1000 cases, including pathological types such as total hip dysplasia and femoral head necrosis. The annotation is completed by two doctors with the title of associate chief physician or above in a double-blind manner, with a Kappa value of 0.

91. The annotation is based on CT data registered with NiftyReg software (projection error ≤0.8mm).

3. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 1, characterized in that, The deep learning pelvic landmark detection module was trained using data augmentation (±10° rotation, 5%–15% scaling, ±20% brightness adjustment), the Adam optimizer (β1=0.9, β2=0.999), and an initial learning rate of 1×10⁻⁶. -4 Furthermore, the loss function is attenuated by 50% every 20 rounds, and the average error of marker detection is <1.5mm.

4. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 1, characterized in that, In the pelvic 3D pose reconstruction module, the MLP input consists of 19 features (2D coordinates of 8 marker points + patient height / weight / gender), with 2 hidden layers (128 / 64 neurons, ReLU activation), and the output is a 3×3 rotation matrix and a 3×1 translation vector. The training loss is a weighted sum of the rotation matrix cosine similarity loss and the translation vector L2 loss (weight ratio 7:3).

5. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 1, characterized in that, In the functional pelvic tilt calculation module, the reference range for normal adult tilt is 5° to 15°, the critical range is 16° to 19°, and the abnormal range is ≥20°.

6. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 1, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing: Collect labeled AP X-ray images, perform grayscale normalization and noise reduction on the X-ray images, and output standardized images; (2) Landmark detection: Use a pre-trained U-Net+++ResNet50 model to locate 8 key landmarks from the standardized images and output coordinates; (3) 3D pose reconstruction: Match the landmark coordinates with the SSM model, and iteratively optimize the three-dimensional pose parameters by combining the PnP / MLP algorithm with biomechanical constraints; the convergence condition of the iterative optimization is that the error between two adjacent pose parameters is <0.3° and the RMSE of the three-dimensional pose reconstruction is <2.0mm; (4) Tilt calculation: Based on the three-dimensional pose parameters, calculate the angle between the "bilateral ASIS-pubic tubercle plane" and the vertical line of the ground to obtain the functional pelvic tilt; the tilt calculation accuracy is ≤0.5° and the average absolute error with the CT gold standard is <1.5°; (5) Result visualization: Display the landmark annotations, three-dimensional model and tilt values ​​through the interface, export the report and connect to the PACS system.

7. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 6, characterized in that, The method takes ≤2.5 minutes to process a single sample and reduces the radiation dose by 40% compared to the traditional lateral X-ray method.

8. The intelligent evaluation system and method for three-dimensional pelvic pose and tilt based on a single anteroposterior X-ray film according to claim 6, characterized in that, The method is applicable to preoperative planning for total hip arthroplasty, assessment of pelvic fracture reduction, and correction of spinal-pelvic deformities.

Images