Motion and radiation dose reduction in quantitative CT perfusion imaging of acute stroke

Abstract

Computed tomography perfusion (CTP) imaging provides vital decision-support for physicians in the diagnosis and treatment planning for acute ischaemic stroke. Serial three-dimensional frames collected over 1-2 minutes during the transit of contrast agent enables visualisation of the integrity of the cerebral vasculature and underpins quantitative haemodynamic modelling to characterise stroke lesions. Notwithstanding the value of CTP imaging for stroke management, there are two areas of fundamental limitation: the increased likelihood of motion-induced corruption of the serial (4D) data compared to conventional 3D neuroimaging CT scans that complete within seconds, and the noise-limiting radiation exposure to patients to ensure that robust haemodynamic modelling can be performed. The overarching aim of this thesis was to develop methods to address these key limitations in CTP imaging, thereby improving the accuracy of image-based stroke analysis and long-term outcomes for patients. Our starting point was to characterise the prevalence, severity, temporal behaviour and dependencies of head movement during CTP imaging studies, and to quantify its clinical impact. Based on this understanding, a predictive model was established to identify patient-specific risk factors for motion. The model implicated stroke severity quantified by the National Institutes of Health Stroke Scale (NIHSS), patient age and time from stroke onset to imaging as the most important factors, all of which can be used pre-emptively to mitigate motion risk in CTP imaging. The results also showed that the accuracy of image interpretation and treatment decision making can potentially be improved for at least a fifth of CTP studies by developing retrospective intra-frame motion correction methods to augment conventional inter-frame motion correction. Although motion correction is well-recognised as an important pre-requisite to haemodynamic modelling in CTP image analysis, only inter-frame alignment is used and the impact of intra-frame corruption caused by continuous motion is ignored. We investigated the Intel RealSense D415 depth sensor, a compact, markerless and consumer-grade optical motion tracking device, for potential use in providing rapid and accurate pose estimates for continuous motion in CTP imaging. Suitability of the device was characterised with respect to thermal stability and jitter, static and dynamic six degree-of-freedom pose accuracy, and adaptability to the clinical setting. A conservative pose accuracy estimate for robotically controlled phantom motion was < 2 mm and < 1°, and for volunteer motion inside a clinical CT scanner was < 3 mm and < 1°. The device therefore shows promise for CTP motion correction but would likely need to be used in a multi-Intel D415 sensor configuration, or used to augment data-driven methods. To simultaneously reduce the radiation dose and the likelihood of motion during a CTP acquisition, we attempted to reduce the scan duration by reducing the number of frames acquired. This was achieved using a novel application of a stochastic adversarial video prediction approach trained to predict late CTP image frames from early frames, thereby avoiding the truncation of the wash-out phase of contrast agent transit. Using this approach to predict the last 18 CTP frames resulted in bolus shape characteristics deviating by < 4 ± 4% compared to the ground-truth. Average volumetric error of the hypo-perfused region was overestimated by 28.36 mL (22%) and the corresponding spatial agreement was 83% (mean dice coefficient). The results showed that predicting the last 18 frames can preserve the majority of clinical content of the images while simultaneously reducing the scan duration and radiation dose by 65% and 54.5%, respectively. The final strategy developed in this thesis was a radiation dose reduction method based on using a 3D generative adversarial network (GAN) to synthesise normal-dose CTP images from low-dose images. The method incorporated pre-processing aimed at leveraging the full spatio-temporal (4D) information of CTP data within a 3D GAN architecture. The quality of GAN-denoised images was assessed via image quality metrics, expert quality rating, and the preservation of the lesion characteristics. The results showed that prioritising temporal information in adapting 4D CTP data to the 3D GAN model resulted in better restoration of tissue haemodynamic information. The average lesion volumetric error reduced significantly by 18 - 29% and dice coefficient improved significantly by 15 - 22% at 50% and 25% of normal radiation dose using the GAN model. In summary, this thesis reports novel quantitative methods to improve our patient-specific understanding of the impact and dependencies of head motion during CTP imaging, the potential use of practical consumer-grade motion tracking devices for comprehensive motion-corrected CTP imaging, and two state-of-the-art deep learning-based approaches for radiation dose reduction in CTP imaging. The proposed methods lay the foundation for improved image-based stroke analysis and optimised CTP imaging workup and radiation dose, thereby providing more robust decision-support for physicians to improve patient outcomes.