Deep learning approach to assess damage mechanics of bone tissue.

Abstract

Machine learning methods have the potential to transform imaging techniques and analysis for healthcare applications with automation, making diagnostics and treatment more accurate and efficient, as well as to provide mechanistic insights into tissue deformation and fracture in physiological and pathological conditions. Here we report an exploratory investigation for the classification and prediction of mechanical states of cortical and trabecular bone tissue using convolutional neural networks (CNNs), residual neural networks (ResNet), and transfer learning applied to a novel dataset derived from high-resolution synchrotron-radiation micro-computed tomography (SR-microCT) images acquired in uniaxial continuous compression in situ. We present the systematic optimization of CNN architectures for classification of this dataset, visualization of class-defining features detected by the CNNs using gradient class activation maps (Grad-CAMs), comparison of CNN performance with ResNet and transfer learning models, and perhaps most critically, the challenges that arose from applying machine learning methods to an experimentally-derived dataset for the first time. With optimized CNN architectures, we obtained trained models that classified novel images between failed and pristine classes with over 98% accuracy for cortical bone and over 90% accuracy for trabecular bone. Harnessing a pre-trained ResNet with transfer learning, we further achieved over 98% accuracy on the cortical dataset, and 99% on the trabecular dataset. This demonstrates that powerful classifiers for high-resolution SR-microCT images can be developed even with few unique training samples and invites further development through the inclusion of more data and training methods to move towards novel, fundamental, and machine learning-driven insights into microstructural states and properties of bone.