Data-driven Solutions for Extremely Ill-posed Tomographic Imaging

Computational imaging, particularly involving radiography and tomography, has a critical role in a variety of industries including security systems, materials science, non-destructive testing, seismic imaging, and medical imaging. Current methods for generating images rely on conventional physical models or on AI-based models that learn patterns in data. These current methods are limited in applications with lower quality data, or where the patterns in the data are highly complex, or where the collection of additional data is expensive. This research seeks to develop improved methods that rely on novel combinations of conventional models and AI-based models. These combinations offer the ability to generate quantitatively accurate and reproducible images from current systems and from future lower cost systems. This project develops physics and learning-based models and methods for tomographic imaging reconstruction and acquisition in extremely limited and scatter-corrupted measurement setups. Developing effective machine learning approaches to enable high-quality object reconstructions in tomography with extremely sparse/limited or highly corrupted (by scatter, noise) measurements is a critical area of ongoing research. A key challenge is that many existing deep learning based methods often underperform in the extremely sparse or highly corrupted measurements regime, and often require large training sets, which can be time-consuming or impractical to generate in applications (e.g., medical imaging, or accurate scatter or hydrodynamic simulations). Moreover, machine learning methods for challenging applications of tomographic imaging (limited-views acquisition, scatter correction, and quantitative density or dynamic reconstruction) remain in early stages. This research aims to significantly advance the current algorithms for data acquisition, data correction, and image reconstruction and analysis in tomography with very limited/corrupted data by incorporating learning-based algorithms with limited training data throughout the imaging process. These techniques will both enable prediction of the optimal limited tomographic views to acquire and leverage a unified learning-based framework for scatter correction and image reconstruction that combines locally learned or robust feature-based scatter correction models and limited-view/low-dose reconstruction models that effectively combine learned neural network priors, sparsity-based models, and physics and statistical imaging models. In summary, this research develops and evaluates a machine learning-driven tomographic imaging framework that employs a unified approach combining machine learned and data-adaptive priors and physics and statistical models and simulations. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.