Deep Learning-Based Automatic Diagnosis System for Developmental Dysplasia of the Hip

TL;DR

This work addresses the variability and inefficiency of manual DDH diagnosis by introducing an automated end-to-end system that detects eight pelvic keypoints with a Mask-RCNN-based detector (ResNet-$50$ backbone), computes the $CE$, $T ext{önnis}$, and $Sharp$ angles, and integrates them through a data-driven scoring scheme. The system demonstrates strong angle measurement accuracy (ICC$CE=0.957$, ICC$T ext{önnis}=0.942$, ICC$Sharp=0.966$) and DDH diagnostic performance (F1$=0.863$), outperforming a cohort of moderately experienced orthopedists and single-angle criteria. A key contribution is the explicit, data-driven scoring that weights angle information to produce an explainable diagnosis with high specificity ($0.996$) and balanced sensitivity ($0.824$). This approach has potential clinical impact by reducing measurement variability, supporting less-experienced clinicians, and enabling remote or centralized second opinions, though external validation is needed. The work advances automated radiographic DDH assessment by unifying keypoint localization, angle computation, and interpretable diagnostic reasoning.

Abstract

Objective: The clinical diagnosis of developmental dysplasia of the hip (DDH) typically involves manually measuring key radiological angles -- Center-Edge (CE), Tonnis, and Sharp angles -- from pelvic radiographs, a process that is time-consuming and susceptible to variability. This study aims to develop an automated system that integrates these measurements to enhance the accuracy and consistency of DDH diagnosis. Methods and procedures: We developed an end-to-end deep learning model for keypoint detection that accurately identifies eight anatomical keypoints from pelvic radiographs, enabling the automated calculation of CE, Tonnis, and Sharp angles. To support the diagnostic decision, we introduced a novel data-driven scoring system that combines the information from all three angles into a comprehensive and explainable diagnostic output. Results: The system demonstrated superior consistency in angle measurements compared to a cohort of eight moderately experienced orthopedists. The intraclass correlation coefficients for the CE, Tonnis, and Sharp angles were 0.957 (95% CI: 0.952--0.962), 0.942 (95% CI: 0.937--0.947), and 0.966 (95% CI: 0.964--0.968), respectively. The system achieved a diagnostic F1 score of 0.863 (95% CI: 0.851--0.876), significantly outperforming the orthopedist group (0.777, 95% CI: 0.737--0.817, p = 0.005), as well as using clinical diagnostic criteria for each angle individually (p<0.001). Conclusion: The proposed system provides reliable and consistent automated measurements of radiological angles and an explainable diagnostic output for DDH, outperforming moderately experienced clinicians. Clinical impact: This AI-powered solution reduces the variability and potential errors of manual measurements, offering clinicians a more consistent and interpretable tool for DDH diagnosis.

Figures (5)

• Figure 1: Diagnosis generated by the proposed system based on an anteroposterior view pelvic radiograph. The system detects four keypoints on each side of the hip: (A) the inferior boundary of the teardrop point, (B) center of the femoral head, (C) lateral edge of the acetabulum, and (D) medial aspect of the acetabulum. The angle measurements and diagnostic scores are displayed in the bottom text (CE: Center–Edge). Angles that exceed the normal range are highlighted in red in the textual results. The right hip (marked as R on the radiograph) is diagnosed as "DDH present”, as the total score (7) is greater than the diagnostic threshold of 5. The diagnosis for the left hip (marked as L on the radiograph) is "DDH absent”.
• Figure 2: The architecture of the keypoint detection model. The ResNet50 model was used to extract features from the input radiograph. The feature maps were then fed into the region proposal network to generate candidate regions of interest (RoI). The RoIAlign layer converts the feature maps and proposed regions of interest into the same size. Subsequently, two parallel neural network branches are responsible for keypoint detection and bounding box regression, respectively.
• Figure 3: Bland-Altman analysis of the detected and reference measurements of the (a) Center-Edge (CE), (b) Tönnis, and (c) Sharp angles in the Test set.
• Figure 4: The confusion matrix of DDH diagnosis in the Test set using the proposed scoring system and the mean angle measurements across 10-fold cross-validation.
• Figure 5: Relationship between the diagnostic threshold of the scoring system (x-axis) and the diagnostic F1 score (y-axis)performance (F1 score). The solid lines connect the mean F1 score using different threshold values over the 10-fold cross-validation grid search. The error bar and shaded area represent the range of plus-minus 1-time standard deviation across the 10-fold cross-validation.