Classification of Osteoporosis in Knee X-ray using Transfer Learning and Random Forest

Osteoporosis, a prevalent skeletal disorder characterized by weakened bone strength and integrity, poses a significant health risk, particularly for older adults and postmenopausal women. Early detection is critical to mitigate fracture risks and improve patient outcomes. This research investigates the potential of pretrained convolutional neural networks for automated osteoporosis detection in knee X-ray images and highlighting the impact of image preprocessing techniques on model performance. We evaluate four pretrained models (ResNet-50, ResNet-101, VGG16, and VGG19) for feature extraction, coupled with a Random Forest classifier optimized using Bayesian Optimization. Our framework explores the effectiveness of different preprocessing methods, including Bilateral Filtering and Contrast Limited Adaptive Histogram Equalization, to enhance feature quality and improve classification accuracy. Evaluation on a dataset of knee X-ray images collected from the University of Sharjah Hospital reveals that VGG architectures, particularly VGG16, demonstrate superior performance in detecting knee osteoporosis. VGG16, preprocessed using Bilateral Filtering and CLAHE, achieved a $\mathbf{7 8 . 3 9 \%}$ accuracy, $\mathbf{8 0 . 0 0 \%}$ precision, and $\mathbf{7 6 . 9 2 \%}$ recall. This study underscores the importance of selecting both model architecture and preprocessing techniques for optimal performance. Further optimization and more data could potentially lead to a more robust and accurate model.