Comparison of artificial neural network configurations to extract micro-architectural properties of cortical bone using ultrasonic attenuation: a 2-D numerical study

Osteoporosis is the most common metabolic bone disorder. It affects both cortical and trabecular bone and is characterized by low bone mass, tissue degradation, deteriorated macroscopic mechanical properties, and altered micro-architecture [1], [2], [3], [4]. Loss in bone mass leads to frequent fracturing, higher mortality rates and reduction in life expectancy [5], [6], [7]. The goal of this study is to move towards better diagnosis of osteoporosis through a data driven approach. Two different configurations of the same artificial neural network were used to take advantage of the relationship between frequency dependent attenuation and micro-structural properties such as porosity, pore density, pore diameter, and standard deviation of pore diameter. Finite difference simulation were conducted using in the 1-8MHz frequency range. The frequency-dependent attenuation data was used to create a 1×8 feature vector that was inputted into the neural network. Then a recursive feature elimination was used to optimize the network and identify the most relevant frequencies. Our results indicate that the most important frequencies are 1, 3, 4, and 8 MHz. A new 1×4 feature vector containing only data from these frequencies was used. Both configurations accurately predict pore diameter, porosity, and standard deviation but were less accurate on pore density. Overall the network using the 1×4 feature vector outperformed the one using the 1×8 when comparing results, computational efficiency, and accuracy.