SCATTERING-INFORMED TOMOGRAPHIC IMAGE RECONSTRUCTION FOR ION IMAGING

Radiotherapy stands as the leading treatment for cancer. Though the potential of X-rays is proven, the feasibility of using charged particles is gaining growing interest. While photons show an exponentially decreasing dose, ions release ever more energy along their path, culminating at the Bragg Peak. Thus, targeting accuracy and preservation of healthy tissues are improved. Nonetheless, ion imaging for treatment planning is still a long way from being a clinical reality: limited image quality and resolution due to Multiple Coulomb Scattering events across the medium were thoroughly assessed within this project. Ion imaging relies on two key quantities: the residual energy (stopping power), translating to Water Equivalent Thickness (WET), and particle deviation from a straight trajectory (scattering power). Upon combining these measurements with an estimation of the ion trajectory, iterative tomographic reconstruction algorithms were adopted to retrieve a map of the ion stopping power ratio relative to water, or Relative Stopping Power (RSP). Two detector setups were explored. List-mode detectors provide spatial and energy data. The Most Likely Path (MLP) algorithm is then applied to reconstruct the individual proton trajectory (scattering line). Since the actual scattering power can be measured only at the entry and exit location, the estimated path is affected by statistical uncertainty in the middle. By introducing a Gaussian scattering model into the MLP algorithm, a volumetric distribution of all possible trajectories was created (scattering spindle). Integration-mode detectors drop all spatial information, so the pencil beam trajectory is resolved as straight (nominal line). However, as its nominal dimension and direction are known, the scattering model can be adopted as a conical Gaussian distribution (scattering cone). Since WET quantities are carried by the ensemble of protons, a value statistically representative of the whole pencil beam was extracted from WET occurrence histograms. To exploit the whole histogram, a novel approach where a Gaussian cone is realized for each WET component (scattering cones) was explored. The images were reconstructed with and without embedding the Gaussian scattering models into the estimated trajectories. All models were tested in a water- and a tissue-based reconstruction. For list-mode images, the Gaussian spindle provides a better image quality and robustness to noise, with a relative improvement up to 10% over the scattering line. Integration-mode images see a relative improvement over 6% for the water cone over the nominal line, while the tissue cone is not able to provide a superior quality. The scattering cones, instead, show a slower convergence due to more abundant reconstruction data, but a lower error when compared to the respective single cone model (over 10% improvement). Though the tissue models are more accurate, the water ones generally offer better results. Since the water scenario usually implies higher WET integrals along the ion trajectory, the respective Gaussian envelop is more pronounced, thus adding an intra-reconstruction smoothing effect that favors image stability and convergence. Predominant in integration-mode configuration, the same phenomenon can occur in list-mode images. Future studies may explore the introduction of intra-reconstruction filtering methods in tissue models to investigate a possible compensation over the water-based scenario. Data-driven approaches may also benefit from these models.