An interpretable deep concatenated architecture for osteoporosis detection using enhanced knee radiographs.
Authors
Affiliations (4)
Affiliations (4)
- Department of Computer Science & Engineering, Chandigarh University, Mohali, Punjab, India.
- Department of Computer Science & Engineering, Graphic Era (Deemed to be University), Dehradun, Uttrakhand, India.
- School of Computing, Gachon University, Seongnam, Republic of Korea.
- Department of AI and Data Science, Sejong University, Seoul, Republic of Korea.
Abstract
Osteoporosis refers to a skeletal disease that is progressive and is marked by low bone mineral density and high likelihood of fractures. The problem of early diagnosis is difficult because of the presence of insensitive radiographic features, as well as lack of access to the sophisticated diagnostic instruments including DXA. Thus, it is important to develop a precise and automatic osteoporosis detection system through the traditional X-ray images. This paper suggests a deeper learning model, which combines Rolling Guidance Filtering (RGF) with a deep concatenated model, to classify osteoporosis based on knee X-ray images. The RGF is used as a preprocessing method to enhance the quality of images by maintaining structural edges, but reducing noise and redundant textures. The suggested model is the sum of two pretrained convolutional neural networks, MobileNetV2 and NASNetLarge, to obtain complementary features. These characteristics are joined together and implemented in fully connected layers to do binary classification. Data augmentation and K-fold cross-validation are used to train and test the model on a publicly available dataset. The performance of the proposed model based on the experimental results proves that the model has better performance with the accuracy of 96.5, the AUC of 0.97 and the F1-score of 89.5 to classify osteoporosis. Comparative and ablation studies affirm that RGF and feature concatenation result in a significant improvement of classification accuracy over single models. The results show that edge-preserving image enhancement plus deep feature fusion is an effective way of enhancing diagnostic performance. The proposed framework is superior to the traditional methods in that it can extract fine-grained and high-level characteristics of radiographic images. This method offers a solid and sound program of automated detection of osteoporosis through knee X-rays. It has high potential of actual clinical application, especially in resource-restricted locations where accessibility to sophisticated imaging modalities is restricted.