Study on the Evaluation of Hip Dysplasia by Measuring the Lateral Center-Edge Angle of Hip Joint on X-Ray Using Deep Learning Algorithm

Introduction

Hip dysplasia is a disease characterized by abnormal hip joint structure, with features such as underdeveloped acetabulum and reduced contact area between the femoral head and acetabulum, which can lead to decreased hip joint stability. Early diagnosis and intervention are recommended in clinical practice.^1,2 The early diagnosis of hip dysplasia typically relies on clinical signs and imaging examinations, with X-ray measurements being one of the important tools.³ The lateral center-edge angle (LCEA) is a critical indicator for assessing acetabular coverage, as its value directly reflects the extent of acetabular coverage of the femoral head, with values below normal indicating acetabular dysplasia.⁴ Although the method for measuring LCEA is not complex, manual measurement is susceptible to subjective factors of the operator, leading to errors and variability in results.⁵ In recent years, deep learning algorithms have rapidly developed and can achieve automated analysis of medical images.⁶ Therefore, this study attempts to evaluate the effectiveness and value of deep learning algorithms in the automated measurement of hip LCEA on X-rays, exploring its potential application in diagnosing hip dysplasia. We hope that this study can provide an efficient and reliable automated measurement method for the imaging evaluation of hip dysplasia, promoting early diagnosis and intervention of the condition.

Materials and Methods

Participants

This retrospective study included adult patients who underwent bilateral anteroposterior (AP) hip radiographs at our institution between February 2023 and February 2024. A total of 287 patients (574 hips) were initially screened. Patients were included if they were aged 18 years or older with closed epiphyseal plates, had normal cognitive and communication abilities, and were willing to participate in the study. Complete clinical data and high-quality, analyzable bilateral AP hip radiographs were required.

Exclusion criteria were as follows: a history of hip or pelvic surgery (n = 10), hip dislocation or fracture (n = 5), femoral head necrosis or hip osteoarthritis with Kellgren-Lawrence grade > 2 (n = 9), and other hip-related structural disorders, including rheumatoid arthritis (n = 9), Paget’s disease (n = 5), and ankylosing spondylitis (n = 3).

After applying these criteria, 254 patients remained eligible. Subsequently, 23 patients (46 hips) were excluded due to poor image quality, as determined by quality control assessment (see Section 1.2). The final dataset consisted of 231 patients (462 hips), including 75 males and 156 females. The participants’ ages ranged from 21 to 74 years, with a mean age of 47.95 ± 10.13 years.

Imaging Acquisition and Quality Control

All radiographs were obtained using a Siemens Axiom Aristos MX digital radiography system. Patients were positioned supine with legs extended and parallel, and feet internally rotated by approximately 15°. The source-to-image distance was set at 100 cm. Exposure parameters included a tube voltage of 70–100 kV and a tube current of 160–200 mA. The central X-ray beam was directed perpendicular to the midpoint between the anterior superior iliac spine (ASIS) line and the upper margin of the pubic symphysis. The exposure field extended from the ASIS (superior border) to the level of the lesser trochanters (inferior border).

To ensure image quality, two senior musculoskeletal radiologists (with 10 and 15 years of experience, respectively) independently reviewed all AP hip radiographs. Images were excluded if they showed pelvic rotation (defined as a deviation of the pubic symphysis from the sacral midline greater than 1 cm; n = 15), poorly defined anatomical landmarks such as the acetabular roof or teardrop (n = 6), or inadequate exposure leading to unclear bony structures (n = 2). In total, 23 patients (46 hips) were excluded due to suboptimal image quality.

Image Annotation and Ground Truth Definition

Image annotations were performed using ITK-SNAP software by two senior musculoskeletal radiologists with 10 and 15 years of experience, respectively. The annotation process included labeling the proximal femoral contour and key acetabular landmarks (the lateral edge of the acetabulum and the inferior margin of the teardrop). Before the formal annotation, both radiologists underwent a 2-hour standardized training session to ensure consistency.

To assess inter-rater reliability, a random subset of 50 radiographs was selected. The intraclass correlation coefficient (ICC) was 0.92 for femoral head center coordinates and 0.89 for acetabular landmarks. Annotations with differences exceeding 2 mm were reviewed and resolved by consensus. The final set of annotations was saved in NIfTI format and used as the ground truth for model training.

Model Development

All images were preprocessed by zero-padding or cropping to a uniform size of 2560×3072 pixels, followed by downsampling to 640×768 pixels. Data augmentation techniques were applied to the training set, including horizontal flipping (50% probability), random rotation (±10°), brightness adjustment (±15%), and random cropping (to 512×512 pixels).

A Dense U-Net architecture was employed for segmentation and key point detection. The encoder consisted of five dense blocks, with output channel sizes of 32, 64, 128, 256, and 512, respectively. Each block included 3×3 convolutions followed by batch normalization and ReLU activation. The decoder path utilized upsampling operations combined with skip connections that fused low-level features from the encoder. The output layer used a sigmoid activation to generate heatmaps for key point detection and segmentation masks for bone contour delineation. The total number of trainable parameters was approximately 2.7 million.

Model training was conducted using Python 3.8, Keras 2.4, and TensorFlow 2.2 on an NVIDIA RTX A6000 GPU (48 GB VRAM) with 128 GB of RAM. The model architecture and implementation adhered to the MINIMAR (Minimum Information for Medical AI Reporting) guideline standards.

Cross-Validation and Testing

To validate the model, a stratified 10-fold cross-validation was performed based on IHDI classification. The full dataset of 462 hips was randomly divided into 10 approximately equal subsets (each containing 46–47 hips). In each fold, 9 subsets were used for training and 1 for validation, ensuring that every sample was used once for validation.

Additionally, an independent test set comprising 93 hips (20% of the total dataset) was set aside before training and was not exposed to the model during training or validation. This test set was used exclusively for final performance evaluation.

LCEA Measurement and Diagnostic Criteria

The model automatically calculated the lateral center-edge angle (LCEA) using the coordinates of the femoral head center and the lateral edge of the acetabulum based on the geometric relationship:

Manual LCEA measurements were independently performed by two radiologists who were blinded to the AI results. Each radiologist measured LCEA twice (with a minimum interval of 30 days), and the average of the two readings was used as the final reference value.

Diagnostic thresholds were defined as follows: Developmental dysplasia of the hip (DDH): LCEA ≤ 20°; Borderline dysplasia: 20° < LCEA ≤ 25°; Normal hip: LCEA > 25°.⁷

Statistical Analysis

Agreement between AI and manual LCEA measurements was assessed using Bland–Altman plots and intraclass correlation coefficients (ICC) with 95% confidence intervals (CI). Prediction error was evaluated using root mean square error (RMSE) and mean absolute error (MAE).

Diagnostic performance was evaluated using receiver operating characteristic (ROC) curve analysis, with manual measurement as the reference standard. The area under the ROC curve (AUC) was used to assess the model’s ability to detect DDH.

All statistical analyses were performed using SPSS version 20.0 (Wilcoxon rank-sum test for group comparison, Pearson correlation analysis), and Python with the scikit-learn package for AI performance metrics.

Results

Angle Comparison

A total of 462 hips were measured, and the consistency of the measurement results between observers (ICC 0.875~0.924) and within observers (ICC 0.902~0.953) was good. There was no statistically significant difference between the manual and automatic measurements of left LCEA and right LCEA (P>0.05), as shown in Table 1. The ICC values for the manual and automatic measurements of left LCEA and right LCEA were 0.936 and 0.902, respectively, and the r values were 0.929 and 0.913, respectively (P<0.05), as shown in Figure 1. The median RMSE for left and right LCEA were 2.27° and 3.29°, and the median MAE were 1.53° and 1.68°, respectively.

Table 1 Comparison of Manual and Automatic Measurements of Bilateral Hip LCEA (n=462, °)

Figure 1 Correlation Scatter Plot of Manual and Automatic Measurements of Hip LCEA.

Hip Development

A total of 44 hips were diagnosed with hip dysplasia, and 56 hips were diagnosed with borderline hip dysplasia. The sensitivity of the automatic measurement for diagnosing hip dysplasia was 88.64% (39/44), with an accuracy of 95.67% (442/462); the sensitivity for diagnosing borderline dysplasia was 66.07% (37/56), with an accuracy of 91.56% (423/462). The results of the ROC curve analysis showed that the automatic measurement of the LCE angle, using a threshold of 25.24°, achieved an AUC of 0.917 (P < 0.05) for the diagnosis of hip dysplasia, as shown in Figure 2.

Figure 2 ROC Curve for Diagnosing Hip Dysplasia Using Automatic LCEA Measurement.

Discussion

Developmental dysplasia of the hip (DDH) is a common orthopedic condition that can lead to chronic pain and early osteoarthritis if not diagnosed and treated early. Accurate measurement of the lateral center-edge angle (LCEA) on anteroposterior (AP) pelvic radiographs is essential for the diagnosis of DDH, borderline dysplasia, and normal hip morphology. Manual measurement, although widely used, is time-consuming and subject to inter- and intra-observer variability, especially when anatomical landmarks are not clearly defined. In recent years, artificial intelligence (AI)-based tools, particularly deep learning algorithms, have shown promise in automating radiological assessments with high accuracy and reproducibility.^8–11

In this study, we validated a deep learning model based on a Dense U-Net architecture to automatically measure the LCEA from bilateral AP hip radiographs. Our model demonstrated high agreement with manual measurements, showing intraclass correlation coefficients (ICCs) of 0.936 and 0.902 for the left and right hips, respectively. These values indicate excellent consistency and are comparable to the interobserver and intraobserver ICCs of manual measurement (range: 0.875–0.953), supporting the reliability of the AI approach. Additionally, the root mean square error (RMSE) and mean absolute error (MAE) for automatic measurements were relatively low (RMSE: 2.27°–3.29°, MAE: 1.53°–1.68°), further confirming the precision of the model.

Compared with previous studies, our findings demonstrate superior or comparable performance. Weinstein et al¹² developed a CNN-based model for automated DDH diagnosis and reported an AUC of 0.89 for LCEA classification, which is slightly lower than the AUC of 0.917 achieved in our model. Similarly, Kamenaga et al¹³ reported ICCs between 0.85 and 0.90 for LCEA measurements using a ResNet architecture, while our model achieved higher ICCs on both sides. This performance improvement may be attributed to the Dense U-Net’s multi-scale feature fusion and skip connections, which enhanced anatomical landmark localization. These comparisons reinforce the effectiveness of our model design and the robustness of our dataset.

Interestingly, the automatic measurements did not significantly differ from manual measurements (P > 0.05), suggesting that the model not only approximates human-level performance but does so consistently across different hip sides. This is especially relevant considering the variability in pelvic positioning and anatomical landmark visibility that often complicates manual assessments. The high correlation coefficients (r = 0.929 and 0.913) also reflect strong linear agreement between AI and manual methods, as visualized in the correlation scatter plots.

Beyond measurement accuracy, the diagnostic utility of the automated LCEA was also evaluated. Our model achieved excellent performance in detecting hip dysplasia, with an area under the ROC curve (AUC) of 0.917 and an accuracy of 95.67%. These results are consistent with previously reported deep learning methods used for similar tasks,^14–17 and further support the potential of AI-assisted tools in improving diagnostic efficiency. Although the sensitivity for borderline dysplasia detection was lower (66.07%), this is not uncommon, as borderline cases often present with subtle anatomical variations and diagnostic ambiguity.^18–21 Future integration of clinical data or multimodal imaging such as MRI may help improve model performance in this subgroup.

Several methodological strengths support the robustness of our findings. First, all radiographs were acquired under standardized imaging protocols and reviewed by experienced musculoskeletal radiologists, ensuring high-quality inputs. Second, landmark annotation was performed by trained experts with good inter-rater reliability, and discrepancies were resolved through consensus, ensuring the reliability of ground truth. Third, the study followed the MINIMAR (Minimum Information for Medical AI Reporting) standards,^22–24 used stratified 10-fold cross-validation, and reserved an independent test set for unbiased evaluation, enhancing the generalizability and credibility of the results.

Nevertheless, this study has several limitations. First, it was a retrospective single-center study, and the model’s generalizability to radiographs obtained from other institutions, or with different imaging protocols, requires further validation. Second, although the model showed high diagnostic accuracy for DDH, its performance in borderline dysplasia cases remains suboptimal and warrants future improvement. Third, all manual measurements used for model validation were based on radiologist consensus and lacked surgical or long-term follow-up confirmation as an objective standard. Fourth, this study did not perform stratified analysis based on patient characteristics such as age, sex, or body mass index (BMI), which may influence hip morphology and LCEA measurement.^25,26 Future research should consider incorporating stratified sub-group analysis to evaluate whether the model maintains consistent performance across different demographic and anatomical groups.

Future directions include multi-center prospective validation, real-world integration with PACS systems, and interface development for clinical use. It may also be valuable to explore the model’s application in longitudinal follow-up of hip development or postoperative outcomes, and to investigate whether incorporating 3D imaging or clinical context improves performance.

In conclusion, we developed a deep learning-based model capable of automatically measuring the LCEA from standard AP hip radiographs with high accuracy and diagnostic utility. The model showed excellent agreement with expert manual measurements and demonstrated strong performance in detecting DDH. These findings support the potential of AI to enhance the efficiency, reproducibility, and objectivity of radiographic assessments in clinical orthopedics, particularly in the early diagnosis and management of hip dysplasia.