Bench, Ciaran; (2022) Data-driven quantitative photoacoustic tomography. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Data_driven_quantitative_photoacoustic_tomography.pdf - Other Download (15MB) | Preview

Abstract

Spatial information about the 3D distribution of blood oxygen saturation (sO2) in vivo is of clinical interest as it encodes important physiological information about tissue health/pathology. Photoacoustic tomography (PAT) is a biomedical imaging modality that, in principle, can be used to acquire this information. Images are formed by illuminating the sample with a laser pulse where, after multiple scattering events, the optical energy is absorbed. A subsequent rise in temperature induces an increase in pressure (the photoacoustic initial pressure p0) that propagates to the sample surface as an acoustic wave. These acoustic waves are detected as pressure time series by sensor arrays and used to reconstruct images of sample’s p0 distribution. This encodes information about the sample’s absorption distribution, and can be used to estimate sO2. However, an ill-posed nonlinear inverse problem stands in the way of acquiring estimates in vivo. Current approaches to solving this problem fall short of being widely and successfully applied to in vivo tissues due to their reliance on simplifying assumptions about the tissue, prior knowledge of its optical properties, or the formulation of a forward model accurately describing image acquisition with a specific imaging system. Here, we investigate the use of data-driven approaches (deep convolutional networks) to solve this problem. Networks only require a dataset of examples to learn a mapping from PAT data to images of the sO2 distribution. We show the results of training a 3D convolutional network to estimate the 3D sO2 distribution within model tissues from 3D multiwavelength simulated images. However, acquiring a realistic training set to enable successful in vivo application is non-trivial given the challenges associated with estimating ground truth sO2 distributions and the current limitations of simulating training data. We suggest/test several methods to 1) acquire more realistic training data or 2) improve network performance in the absence of adequate quantities of realistic training data. For 1) we describe how training data may be acquired from an organ perfusion system and outline a possible design. Separately, we describe how training data may be generated synthetically using a variant of generative adversarial networks called ambientGANs. For 2), we show how the accuracy of networks trained with limited training data can be improved with self-training. We also demonstrate how the domain gap between training and test sets can be minimised with unsupervised domain adaption to improve quantification accuracy. Overall, this thesis clarifies the advantages of data-driven approaches, and suggests concrete steps towards overcoming the challenges with in vivo application.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item