An interpretable deep concatenated architecture for osteoporosis detection using enhanced knee radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedics

Background:

• Osteoporosis diagnosis is challenging due to subtle radiographic features and limited access to advanced tools like DXA.
• There is a critical need for accessible, automated osteoporosis detection systems using conventional X-ray images.

Purpose of the Study:

• To develop a precise and automated osteoporosis detection system utilizing deep learning on knee X-ray images.
• To enhance diagnostic accuracy by combining image preprocessing with a robust deep learning architecture.

Main Methods:

• A deep learning model was developed, integrating Rolling Guidance Filtering (RGF) for image enhancement with a concatenated deep model (MobileNetV2 and NASNetLarge).
• RGF preprocessing preserves structural edges while reducing noise, improving image quality for analysis.
• The model underwent training and validation using data augmentation and K-fold cross-validation on a public dataset.

Main Results:

• The proposed model achieved high classification performance with 96.5% accuracy, 0.97 AUC, and 89.5% F1-score.
• Comparative and ablation studies confirmed that RGF preprocessing and feature concatenation significantly improved classification accuracy over single-model approaches.
• The combined approach effectively extracts fine-grained and high-level features from radiographic images.

Conclusions:

• Edge-preserving image enhancement coupled with deep feature fusion offers an effective strategy for improving osteoporosis diagnostic performance.
• The developed framework demonstrates superiority over traditional methods in extracting detailed radiographic characteristics.
• This automated osteoporosis detection system using knee X-rays holds significant potential for clinical application, particularly in resource-limited settings.