Investigation of Focal Loss in Deep Learning Models For Femur Fractures Classification

This paper develops an approach based on deep learning for the classifications of a common critical type of bone fractures, namely proximal femur. The performance of the state-of-the-art deep learning architecture, DenseNet, is investigated along with a recently introduced loss function, focal loss, to address the problem of imbalanced classes. Quantitative assessment is carried out on a real dataset consisting of 1347 X-ray images. Results demonstrate that the proposed deep learning approach utilizing focal loss show better performance for the fracture detection case and comparable results for the classification scenarios.