A Consensus Mechanism to Improve Prediction of Cortical Bone Properties using Ultrafast Ultrasound Acquisition

Human bone microarchitecture is complex and density-based bone assessment modalities cannot fully capture bone strength or health. Ultrasound can be used to assess bone microstructure, but it is hindered by the dense and acoustically diverse nature of cortical bone. This study proposes a methodology for predicting cortical bone thickness and porosity through a novel approach utilizing convolutional neural networks (CNNs), processing multi-frequency radiofrequency (RF) data obtained from ultrafast ultrasound, and implementing a consensus mechanism to enhance reliability. Received ultrasound RF signals are processed using a CNN with a mutual consensus mechanism, which is used to discard received RF data when measurement variation is over a certain threshold. The feasibility of the proposed method is demonstrated through realistic simulations and an ex vivo animal bone study using an ultrafast ultrasound scanner. The preliminary findings of this study demonstrate an enhancement in overall accuracy, with an increase from 92% to 95.6% for thickness and an increase from 73.4% to 88.4% for porosity classification, without and with consensus respectively. The implemented mutual consensus mechanism increases the accuracy of the thickness and porosity estimations both in silico and ex vivo. Ultrafast ultrasound scanners can capture thousands of RF signals within seconds, which results in availability of large datasets for implementation of artificial intelligence and machine learning algorithms. Here, we propose a new approach for ultrafast ultrasound data processing that values data quality over quantity by discarding noisy measurements using a consensus mechanism to improve the final estimation reliability.