Chest radiography (CXR) remains the core imaging examination for the chest despite the availability of diagnostically superior imaging techniques like chest tomosynthesis and Computed Tomography (CT). CXR has many advantages, such as relatively low cost, fast acquisition times, widespread availability, and reduced dose to the patient. Chest radiography images are used in the management of numerous clinical tasks, like pulmonary diseases, bone fractures, visualization of catheters amongst many others. Acquisition protocols used nowadays have at their root those used in screen film systems, and in many cases have not been optimized for the latest generation of digital flat panel detectors. In digital radiography, the optimization process should be referred to the clinical task to be performed, which is the fundamental premise behind our project. As with all uses of radiation for medical purposes, optimization should find a balance between the image quality necessary to perform certain imaging task and the dose delivered to the patient. The main objective of the PhD thesis is the creation of a simulation framework that can be used in Virtual Clinical Trials (VCTs) in chest radiography. To allow a task-based optimization, the imaging chain should also comprise realistic anthropomorphic models including a wide range of tasks. Finally, the framework should allow the study of different elements of the imaging chain that influence clinical task performance and organ dose. The simulation platform was developed to match all these requirements by making use of ray tracing technique to generate noise free primary images and PENELOPE/penEasy Monte Carlo simulations for the generation of the scatter images. Additionally, the antiscatter grid is also included in the simulations. Real detector sharpness and noise characteristics are added to the simulated images via the Modulation Transfer Function (MTF) and the Normalized Noise Power Spectrum (NNPS) respectively. The established methodology was first validated by comparing signal difference to noise ratio (SDNR) from real and simulated images. The test object used consisted of a Poly (methyl methacrylate) (PMMA) block and a small aluminium detail. Large area contrast and noise were verified as a function of PMMA thickness, tube voltage and different dose levels, for antiscatter grid in and out cases. Additional validation of the noise and sharpness modification process was performed. Lastly, primary transmission (Tp) and total transmission (Tt) were measured experimentally in the physical antiscatter grid and compared to those for the simulated grid. Average differences between SDNR measured in real and simulated images were 6% and 5% for grid in and grid out respectively. Maximum differences of 12% were found between the simulated and measured MTF up to the Nyquist frequency. Average deviations below 5% were found between the NNPS measured in real and simulated images. Relative differences for Tp and Tt were below 6% and 13% respectively. Several image segmentation techniques and polygonal mesh modelling were used in the development of the computational phantoms. The models created were based on an existing polygonal mesh phantom, the Realistic Anthropomorphic Flexible (RAF) phantom. The polygonal mesh format of the RAF allowed the modelling of a more realistic lung background and different body types. To be used in task-based optimization, a set of lesions and/or devices commonly found in chest radiography were modelled within the phantoms. A library of 24 realistic anthropomorphic chest phantoms was created to model different type of patients, like male, female and different Body Mass Indexes. The clinical tasks modelled were pulmonary nodules, catheters, rib fractures, pneumothorax and pleural effusion. A set of models depicting Covid-19 disease was also included. The realism of the models anatomy and of the simulated clinical tasks was first validated by experienced radiologists. Then, the images were uploaded for analysis to an AI software which served as extra validation. The validated simulation framework and computational phantoms were used in a Virtual Clinical Trial to investigate the influence of different acquisition parameters in diagnostic performance. Posterior Anterior projections of 11 phantoms including clinical tasks were generated for a range of exposure conditions (i.e., dose level, tube voltage and antiscatter grid use). For the grid, two options were evaluated: antiscatter grid in place and removed. Five tube voltages values were used ranging from 60 to 140 in steps of 20 kVp and four dose levels: 0.62, 1.25, 2.5 and 5.0 μGy. Clinical image processing was applied to the simulated images, which were subsequently scored via a free-response observer study by four radiologists. The statistical analysis of the results was done using the jackknife-alternative free-response receiver operating characteristic (JAFROC) method. Additionally, organ dose calculations were performed for the different settings. The results from this study lead to several practical conclusions: • A reduction of dose by 50% can be achieved without decreasing the diagnostic performance for the clinical tasks studied, i.e., from a working level of 2.5 μGy to 1.25 μGy. • For grid out techniques, using 100 kVp and 80 kVp values did not show significant differences in diagnostic performance. Organ doses were on average 10% lower at 100 kVp compared to 80 kVp. More investigation on this is necessary, in which Anterior Posterior projections are simulated and organ doses are calculated to apply the results to bedside imaging. • For a constant target DAK and grid in technique, 120 kVp gave the lower organ doses with no significant decrease in diagnostic performance. In terms of tube voltage, this justifies the choice made in practice. • In general, no significant difference was found in images with and without the antiscatter device. Further analysis of the effect of the antiscatter grid is necessary. A comparison of different image processing techniques in the enhancement of contrast for grid out technique is relevant. The simulation platform developed was successfully applied to generate synthetic images of anthropomorphic phantoms including a range of clinical tasks on a VCT for chest radiography. Although developed for CXR, the simulation platform and the computational phantoms can be used in a wider range of applications. To our knowledge, this is the first VCT in the field of chest radiography to include this range of clinical tasks.
|Qualification||Master of Science|
|Date of Award||15 Jul 2021|
|State||Published - 15 Jul 2021|