Cortical Bone Assessment with Ultrafast Ultrasound Technology

Human bone microarchitecture is complex, and traditional density-based bone assessment methods often fail to capture bone strength or health accurately. While ultrasound imaging is valuable for many medical applications, its effectiveness in imaging bone anatomy is limited due to significant differences in the speed of sound through various tissues. The mismatch between the acoustic impedance of bone and soft tissue causes poor ultrasound wave penetration into bone and substantial reflection, scattering, and refraction at tissue boundaries, making accurate visualization of bone structures challenging. Consequently, traditional ultrasound may lack the spatial resolution and accuracy needed for detailed bone microarchitecture and integrity assessments. In contrast, ultrafast ultrasound imaging, with its high temporal resolution and advanced signal processing techniques, offers a promising approach to overcome these limitations and quantify bone properties. By rapidly capturing thousands of frames using pulse-echo ultrasound and analysing the data with refined algorithms, ultrafast ultrasound allows for the generation of extensive datasets that can be further processed using artificial intelligence (AI) and machine learning (ML) algorithms to quantify bone structure. This thesis aims to bridge a gap by investigating the potential of machine learning-based techniques in cortical bone studies. It seeks to develop methods for characterizing cortical bone health and predicting key factors indicative of bone quality, such as cortical bone thickness and porosity, particularly in osteoporosis and bone density-related diseases. The methodology includes generating in silico models of human bone for ultrasound simulations and applying data processing techniques such as wavelet transforms, CNNs, mutual agreement approaches, and multi-frequency processing to analyse RF datasets. Additionally, ex vivo measurements are conducted to validate the robustness of these methods and assess their potential for clinical application. This thesis demonstrates that multi-frequency acquisitions improve classification accuracy in in vivo scanning by analysing all frequencies together, rather than separately. While scalograms with standard CNNs provided good accuracy, they required longer training and II processing times. In contrast, using raw RF data with CNNs tailored for multi-dimensional inputs achieved faster execution and strong performance. Additionally, a consensus mechanism combining outputs from different channels enhanced data reliability, increasing thickness classification accuracy from 92% to 95.6% and porosity classification accuracy from 73.4% to 88.4%.