Application of stacked autoencoder for identification of bone fracture.

ABSTRACT

This study presents a stacked autoencoder (SAE)-based assessment method which is one of the unsupervised learning schemes for the investigation of bone fracture. Relatively accurate health monitoring of bone fracture requires considering physical interactions among tissue, muscle, wave propagation and boundary conditions inside the human body. Furthermore, the investigation of fracture, crack and healing process without state-of-the-art medical devices such as CT, X-ray and MRI systems is challenging. To address these issues, this study presents the SAE method that incorporates bilateral symmetry of the human legs and low-frequency transverse vibration. To verify the presented method, several examples are employed with plastic pipes, cadaver legs and human legs. Virtual spectrograms, created by applying a short-time Fourier transform to the differences in vibration responses, are employed for image-based training in SAE. The virtual spectrograms are then classified and the fine-tuning is also carried out to increase the accuracy. Moreover, a confusion matrix is employed to evaluate classification accuracy and training validity.