Towards deep learning based estimation of fracture risk in osteoporosis patients

Abstract

Osteoporosis is a skeletal disorder which leads to bone mass loss and to an increased fracture risk. Recently, physics-based models, employing finite element analysis (FEA), have shown great promise in being able to non-invasively estimate biomechanical quantities of interest in the context of osteoporosis. However, these models have high computational demand, limiting their clinical adoption. In this manuscript, we present a deep learning model based on a convolutional neural network (CNN) for predicting average strain as an alternative to physics-based approaches. The model is trained on a large database of synthetically generated cancellous bone anatomies, where the target values are computed using the physics-based FEA model. The performance of the trained model was assessed by comparing the predictions against physics-based computations on a separate test data set. Correlation between deep learning and physics-based predictions was very good (0.895, p < 0.001), and no systematic bias was found in Bland-Altman analysis. The CNN model also performed better than the previously introduced Support Vector Machine (SVM) model which relied on handcrafted features (correlation 0.847, p < 0.001). Compared to the physics based computation, average execution time was reduced by more than 1000 times, leading to real-time assessment of average strain. Average execution time went down from 32.1 ± 3.0 seconds for the FE model to around 0.03 ± 0.005 seconds for the CNN model on a workstation equipped with 3.0 GHz Intel i7 2-core processor.