Background and Purpose
Hip joint disorders are common causes of pain and functional impairment, often requiring early detection and precise diagnosis to guide clinical management. While plain radiographs are typically used as the first-line diagnostic...
Background and Purpose
Hip joint disorders are common causes of pain and functional impairment, often requiring early detection and precise diagnosis to guide clinical management. While plain radiographs are typically used as the first-line diagnostic test in most clinical settings, their diagnostic accuracy is limited, especially in differentiating between various pathologies. While advanced imaging modalities such as CT and MRI offer higher diagnostic accuracy for hip pathologies, their use is often limited in clinical practice due to factors such as cost, time consumption, limited accessibility, and, in the case of CT, radiation exposure. With the advancement of deep learning, there is increasing interest in artificial intelligence (AI)-based tools that can augment radiographic interpretation.
Therefore, the aim of this study was to develop a deep learning–based screening model to detect the presence of advanced hip pathology using plain radiographs, and subsequently to construct a multi-class diagnostic model based on this framework. The study also sought to evaluate the clinical applicability and generalizability of the models through external validation.
Methods
This study was a two-stage retrospective deep learning study conducted using electronic medical records (EMRs) and plain hip radiographs in DICOM format. In the first stage, a binary classification screening model was developed to distinguish between normal and patients with hip pathology using anteroposterior (AP) hip radiographs collected from Seoul National University Hospital and Seoul National University Bundang Hospital between 2004 and 2012. Ground truth labels were assigned based on radiographic interpretation, surgical history, clinical diagnosis, and pathological confirmation. The model was trained using a convolutional neural network (CNN) based on the EfficientNet-B3 architecture, and externally validated using an independent test set. In the second stage, a multi-class diagnostic model was developed to classify hip radiographs into seven categories: primary osteoarthritis (OA), secondary OA, inflammatory arthritis (IA), osteonecrosis of the femoral head (ONFH), hip fracture, bone tumors, and normal. The CNNs based on EfficientNet-B3 and ConvNeXt architectures (Small, Base, and Large variants) were used to construct the model. A feature fusion approach was implemented by embedding age and sex information and integrating it with image-based features. Model performance was evaluated using a 10% internal validation set and an external test set from a separate institution.
Performance metrics included accuracy, precision, recall, F1-score, top-2 accuracy, area under the receiver operating characteristic curve (AUC), confusion matrix, and Grad-CAM visualizations. Among these, AUC and F1-score were adopted as the primary indicators of overall classification performance, while Grad-CAM was used to interpret the model's attention to diagnostic features within the radiographs.
Results
The screening model developed in the first stage demonstrated high performance, achieving an AUC of 0.99 and an F1-score of 0.97 on the internal test set. External validation also confirmed its strong generalizability, with an AUC of 0.98 and an F1-score of 0.94. The model consistently maintained a sensitivity of 93% and a specificity of 96% in distinguishing between normal and abnormal cases.
The diagnostic model also demonstrated robust performance. Among the evaluated architectures, the ConvNeXt-Large model with feature fusion—incorporating age and sex embeddings—achieved the highest performance on the internal test set (F1 score: 0.725; AUC: 0.952). Feature fusion models consistently outperformed image-only models across all architectures. On external validation, the ConvNeXt-Large model maintained strong performance (F1 score: 0.625; AUC: 0.921), while the EfficientNet-B3 fusion model also yielded favorable results (F1 score: 0.665; AUC: 0.920).
Analysis by disease category showed the highest AUC for hip fractures (0.986) and the lowest for tumors (0.878). Grad-CAM visualizations indicated that the model primarily attended to anatomically relevant regions such as the femoral head and joint space. Frequent misclassification patterns included confusion between primary and secondary OA, as well as between OA and ONFH, reflecting radiographic overlap in advanced disease stages.
Conclusion
This two-stage study demonstrates the feasibility and clinical relevance of deep learning models for the radiographic assessment of hip diseases. Starting from a high-performing screening tool and extending to a multi-class diagnostic framework, the models showed strong generalizability across institutions and patient populations. These findings suggest that AI-assisted interpretation of plain hip radiographs, even with minimal clinical information, may serve as a valuable adjunct in improving diagnostic accuracy and supporting timely decision-making in orthopedic care. Furthermore, by providing accurate results with minimal input, the proposed framework holds promise as a scalable solution in primary care settings and lays the groundwork for the broader application of AI in orthopedic diagnostics.