Enhancing Bone Fracture Detection: A Feature-Based Transfer Learning Approach Using DenseNet with SVM

Accurate classification of bone fractures is essential for effective treatment, yet traditional methods can be error-prone and time-consuming. Transfer learning, particularly feature-based transfer learning, has shown promise in improving medical image analysis. This study investigates the effectiveness of DenseNet architectures combined with SVM classifiers with default hyperparameter for classifying hairline and pathological fractures. A dataset of 500 bone fracture x-ray images was utilized from open source for this study. Features were extracted using DenseNet121, DenseNet169, and DenseNet201 models. The findings indicated that the combination of DenseNet169 and SVM yielded the highest overall accuracy and generalization capacity, with a notable proficiency in differentiating hairline fractures. The classification performance of DenseNet121-SVM and DenseNet201-SVM was robust, although significantly less successful in identifying hairline fractures. The exceptional performance of the DenseNet169-SVM model highlights its promise as a dependable tool for automated bone fracture classification in medical imaging. This work emphasizes the efficacy of integrating advanced transfer learning techniques with conventional classifiers to enhance diagnostic precision, hence facilitating improvements in medical procedures. Additional investigation is advised to authenticate these discoveries using more extensive, practical clinical datasets.