It is well documented that intraoperative ultrasound offers improvements to the extent of tumour resected in neurosurgery but currently fails to depict the boundaries of more invasive tumours. Quantitative ultrasound (QUS) is a technique that models ultrasound scattering in tissue mathematically. It can act as a quantitative tool to identify cancerous regions and be used to define features which can train a machine learning (ML) classifier. The use of QUS to differentiate healthy and malignant brain tissue is the objective of this thesis.
This work began with a proof of concept study which saw the effective implementation of QUS with a linear array transducer, at conventional frequencies, on phantom materials. The results were then used to train a K-nearest neighbours (KNN) binary classifier to differentiate between two soft tissues. Insight into the most practical parameters for near real time tissue identification was achieved, as well as the opportunity to produce parametric images for various QUS parameters. The effects of freezing and fixation of tissue on QUS results were also considered.

The experimental design was developed to obtain a higher lateral spatial resolution before applying it to ex vivo human samples of ten healthy and eight high-grade glioma (HGG) tissues. This was accomplished with both a linear array and a single element scanning system, at centre frequencies of 25 and 74 MHz, respectively. The SoS and attenuation were found to be higher, on average, in the tumour samples than in the healthy tissue. The homodyned K-distribution (HK) parameters alone could distinguish between healthy and HGG tissue to 96% accuracy at 74 MHz, suggesting this is a viable solution for residual HGG detection.

To explore the potential of ML with a larger data set, and to extend the study to low grade glioma (LGG) tissue, acoustic impedance maps based on 300 previously recorded microscope histology images of each tissue type were created. The interaction with high frequency (HF) ultrasound was explored using finite element analysis and QUS parameters were obtained. A classification algorithm was able to differentiate healthy and HGG to near perfect accuracy, but a significantly lower accuracy of 79% was found when distinguishing LGG from healthy tissue maps.

This research represents a step forward in the otherwise unexplored landscape of HF QUS in brain tissue which necessitates further work to transition from laboratory based experiments to in vivo QUS to aid intraoperative glioma detection.