Introduction
Avascular Necrosis of the Femoral Head (AVNFH) is a serious condition characterized by the death of bone tissue due to impaired blood supply, which frequently leads to the collapse of the hip joint. While X-rays are often insufficient for early detection and MRIs are too costly for mass screening, Computed Tomography (CT) scans offer a balanced alternative. However, the medical field has lacked automated deep learning tools specifically designed for CT-based AVNFH diagnosis. Furthermore, existing AI models often function as "black boxes," providing classifications without anatomical context. This paper introduces SRANet, a novel deep learning architecture designed to provide both high-accuracy diagnosis and interpretable localization of necrotic regions in CT imaging.
Methods
The Structure Regularized Attentive Network (SRANet) is designed to incorporate anatomical priors into the learning process. Its architecture is built upon three primary pillars:
- Patch-Based Feature Extraction: To manage the complexity of 3D volumes, the model divides CT images into smaller patches. A 3D Convolutional Neural Network (CNN) then extracts localized features from these segments.
- Gated Self-Attention: Recognizing that necrotic lesions are often localized, the model employs an attention mechanism to weigh the importance of different patches, focusing the network's "attention" on areas most likely to contain pathology.
- Structural Regularization: This core innovation uses traditional image processing (such as adaptive thresholding) to generate bone tissue masks. These masks constrain the attention mechanism, ensuring the model prioritizes bone-related features over irrelevant background noise (e.g., soft tissue).
The final loss function is a hybrid that penalizes both classification errors and instances where the model's attention drifts away from the anatomically relevant bone structures.
Datasets
The study utilized the newly introduced AVNFH-CT dataset, the first large-scale collection of its kind. It contains 1,549 CT images from 1,364 subjects, annotated by three experienced orthopedic surgeons using the ARCO staging system (Stages 0–IV). To ensure high label accuracy, corresponding MRI images were used as a ground-truth reference for many cases. The data was split into training (1,244), validation (155), and testing (150) sets.
Key Results
SRANet demonstrated superior performance compared to standard benchmarks like VGG, MLP-Mixer, and Vision Transformers. The model achieved an Accuracy (ACC) of 0.94 and an Area Under the Curve (AUC) of 0.95.
In comparative tests against human experts, SRANet showed higher specificity (0.97) and overall accuracy, although human experts maintained a slightly higher sensitivity (0.95 vs. 0.91). Critically, the model's "patch attention maps" allow for weakly supervised localization, enabling clinicians to visualize exactly which regions the AI identified as suspicious.
Discussion and Conclusion
The success of SRANet highlights the importance of incorporating anatomical priors into medical AI. By using structural regularization, the model avoids the common pitfall of focusing on non-diagnostic background noise.
Despite these gains, challenges remain in detecting "Stage I" cases, where manifestations are extremely faint. Future work will focus on improving detection for these early-stage patients and developing a fully end-to-end system that integrates bone segmentation and diagnosis more robustly. Ultimately, SRANet suggests that AI-assisted CT scans can serve as a viable, cost-effective screening tool for AVNFH, approaching the diagnostic utility of MRI.
Summary updated