Predicting specific modulus of tibia from mixed modelled dataset using artificial neural network

Abstract

Abstract Bone strength assessment is critical for diagnosing osteoporosis. A proposed method is to measure the resonance frequencies of the tibia to predict the bone’s specific modulus. Machine learning algorithms have been shown to estimate strength of structures and diagnose osteoporosis. However, concerns have been raised over dataset sizes and feature selection. In this article, an Artificial Neural Network (ANN) is trained on an augmented dataset produced by simulations of tibiae and beams with simple cross-sections. The training regime explored reducing the number of modal frequencies, adding the model dimensions to the input, and the proportion of augmentation of training and testing datasets. Training the ANN on the tibia dataset gives an average testing error of 0.37% ± 16.96%. Adding the length to the input vector produced a 98.16% reduction in maximum training and testing error. Training the algorithm on the simpler shape datasets and testing with the tibia dataset was less successful with poor prediction accuracy. But including 45 tibiae models into the training dataset achieved better accuracy (0.03%) and precision (±1.88%) with 6 modal frequencies and the length. Thus, a highly accurate and precise prediction of bone strength may be possible from adding dimensions, dataset augmentation, and modal frequencies.