Computational Methods for Improved Interpretation of High-Density Diffuse Optical Tomography Data

Abstract

High-density diffuse optical tomography (HD-DOT) is a promising neuroimaging technique that can be used to produce three-dimensional reconstructions of brain activity. However, analysis methods applied to this data are not yet standardised, and often do not take advantage of the higher data resolution achievable with HD-DOT. Additionally, machine learning models, which are regularly applied to other neuroimaging modalities, have only been applied to HD-DOT in very limited ways. As such, this thesis primarily describes the development and adaptation of several data analysis methods for HD-DOT, along with the application of wearable HD-DOT in large scale data collection. First, this technology is applied to study the intra- and inter-subject variability of cortical sensitivity across healthy adults and develop a processing pipeline for dimensionality reduction of HD-DOT data. Then, the collection of a large HD-DOT dataset is described, comprising 160 adult participants performing an auditory task. Employing this data, a novel method of feature extraction is explained, using inherently interpretable and physiologically relevant features to perform machine learning classification. Finally, the adaptation of a common neuroimaging analysis method, the general linear model, is described in the context of HD-DOT, using an event-related paradigm to identify how prediction confidence may modulate brain responses in the prefrontal cortex. The research described in this thesis highlights how to improve interpretation of HD-DOT data while maintaining the benefits that this imaging modality offers. The processing methods developed demonstrate the potential for more complex modelling of neural responses using HD-DOT data, and should enable access to HD-DOT computational tools that are relevant across several application areas.