Exploration of clinical applications of grating-based phase-contrast imaging
Talbot-Lau interferometry (TLI) is a recent innovation, that enables x-ray phase-contrast measurements using conventional x-ray tubes and detectors. This development allows clinical implementation of phase-contrast imaging. This had been anticipated for years, given the increased sensitivity of x-ray phase to variation in soft tissue density. Carestream Health engineered a prototype system, the CH-TLI setup, and this PhD work studied the clinical potential of the system. TLI systems generate three image types: a standard x-ray transmission image (Tr), and then two new types, the differential-phase (dP) image and the dark field (DF) image. The dP and DF images are very different in terms of physical origin and possible applications. Given these differences, their clinical potential might differ strongly and therefore the potential of each of the image types is discussed separately below. First the results obtained for dP imaging are presented, followed by those for DF imaging.
In the initial stages of this work, a wide ranging, experimental examination of potential applications using dP imaging revealed no advantage for dP over conventional Tr imaging. However, this conclusion was based on subjective evaluations since the different appearance of Tr and dP images prohibits direct comparisons made using quantitative measures such as contrast-to-noise ratios. Therefore, a method to benchmark dP imaging to Tr imaging was developed. Using detectability as a performance metric, virtual studies were conducted to find the threshold radiation dose at which a given object became detectable. Detectability was quantified using a four alternative forced choice paradigm. The relative dose needed for Tr and dP images to reach a given detectability level was used to quantify the relative performance. This type of observer study requires many accurately simulated images. Classical simulation frameworks for TLI in the literature based on numerical wave propagation methods are computationally expensive and are therefore rather impractical for these virtual simulation studies. This stimulated the development of a hybrid simulation platform, combining analytical equations with measured quality metrics to produce realistic Tr and dP images with levels of sharpness and noise characteristic of the CH-TLI system. General tasks, including the detection of a 4 mm spherical lesion in a uniform background were studied alongside mammography tasks, where the detectability of 5.3 mm non-spherical lesion embedded in a structured mammographic background was evaluated. These virtual studies confirmed our initial experimental findings that for the current CH-TLI system set up/dimensions, Tr imaging outperforms dP imaging for almost all of the evaluated tasks. The mammography detectability study demonstrated that while overlaying tissues complicate the detection of lesions in Tr images, their influence on the detectability in dP images is somewhat reduced. As well as giving a direct means of comparing Tr and dP imaging, the detectability method can also be used to estimate the system quality required for dP to outperform Tr imaging for a given task. By increasing the system visibility, G1-to-G2 distance or G2 frequency, the dP signal-to-noise values can be improved without affecting the Tr image. Their product, `the sensitivity-visibility product', scales with the relative performance of dP. Therefore, from the relative performance, the required senstivity-visiblity product to outperform the Tr image can be estimated in first approximation. It is only a rough estimation as the same magnification, detector and focal spot properties as the CH-TLI setup were assumed. While, in general, dP did not improve detectability in our setup, the most promising dP imaging application was found to be mammography or more specifically the detection of small lesions in structured backgrounds. The sensitivity-visibility product required to outperform Tr is estimated to be around 9.6 higher than the CH-TLI system quality. Such system quality has been published in the literature [Birnbacher2016] and is thus realistic to achieve.
In contrast to the dP imaging investigation, we found that dark field imaging yielded some successful results. In accordance with the literature [Hellbach2016, Schleede2012, Yaroshenko2014], improved visualization of lung tissue was seen using DF imaging. Therefore, dark field imaging was added to two preclinical studies running at the small animal department of the KULeuven, one of which investigated pulmonary aspergillus fungi in mice. It was hypothesised that the presence of the fungi lowers the DF signal produced by the lungs. DF measurements were combined with µCT and bioluminance scans to locate and identify the infected regions. Correlation between the µCT and DF scans was found, where at the suspicious regions the dark field signal indeed showed discontinuities. However, drawing conclusions solely based on the dark field images proved difficult. The second study investigated the effect of hyperoxia and prematurity on the alveolar development of pup rabbits. Preterm rabbits and humans can suffer from arrested development of alveoli, which can lead to bronchopulmonary dysplasia. Exposure to an excess of oxygen (hyperoxia) can increase this chance. In the earlier stages of alveolar development, the alveoli are larger and it was hypothesised that this would affect the DF signal measured. Preliminary results suggest that the DF signal in the lungs of full-term animals can be clearly distinguished from the preterm animals, however differentiation of the hyperoxia cases was more difficult. We can conclude that DF imaging visualizes the aerated lung with high contrast, but differentiating different stages in the alveolar development appears difficult, thus far.
It therefore appears that DF imaging shows much promise as tool for lung imaging. However, the applicability for human lung imaging is not a-priori guaranteed. Human alveoli are a factor five larger than murine alveoli and a corresponding drop in DF sensitivity is expected. On the other hand, the human thorax is thicker than those of mice and the aggregated DF contrast could well saturate, destroying the image. Before designing a setup for large animals, an estimate of the overall sensitivity of DF lung imaging in humans would prove extremely valuable. To this end, a second simulation framework was developed, this time using numerical full-wave propagation methods with the aim of generating realistic DF images rather than to predict system metrics, as is commonly done with these simulations in the literature. Software models of murine and human lung were built and the resulting DF signal simulated. The lungs were modelled as a volume of spheres with diameters matching those of the alveoli. For the murine lung, the settings of the CH-TLI system were applied, while for the human lung the design energy and pixel sizes were adapted to those of PA chest radiography. The estimated linear diffusion coefficient (and thus the corresponding DF signal) of human lung tissue was found to be 120 times smaller than that of murine lung tissue. However, due to the larger thorax thickness in humans, similar total DF contrast is achieved in projection imaging. Dark field imaging has thus great potential in human lung imaging applications and the simulation method developed here could be used to optimize TLI system dimensions for these applications.
To conclude, this thesis investigated the clinical potential of the CH-TLI system and TLI imaging in general. While using the CH-TLI system the dP contrast was not sufficient to outperform Tr imaging, the DF image showed promising results for lung imaging. However, with improved system dimensions, TLI could be a powerful tool for mammography as well. In this thesis we have proposed approaches to investigate the feasibility of applications for a given system setup. In the future they can be used to help estimate the required imaging performance to outperform current clinical modalities for specific applications and support the development of new system designs on a quantitative basis.