News & Events

Integrating Learning-based and Physical Model-based Methods for Elasticity Reconstruction

Narges Mohammadi

Supervised by Mujdat Cetin

601 Computer Studies Building

Elasticity images which provide quantitative visualization of tissue stiffness can be recon- structed by solving an inverse problem. Classical model-based methods are usually formulated in terms of constrained optimization problems consisting of a physical imaging model and an analytical regularizer where the reconstruction performance for complex elasticity patterns are significantly degraded. Thus, incorporating a suitable regularizer is essential for reducing the elasticity reconstruction artifacts for experimental measurements while finding the most suit- able one is challenging. On the other hand, the physical forward operator for the elasticity imaging model is directly influenced by the displacement measurements which results in an in- accurate forward operator in presence of noisy displacement acquisitions. Therefore, correcting the forward model is required when solving the optimization problem to reduce the elasticity reconstruction error with experimental displacement measurements.

The main concentration of this thesis is on improving computational medical imaging algorithms for ultrasound elasticity reconstruction and magnetic resonance elasticity (MRE) imaging using statistical physical modeling and machine learning. In particular, we present a new statistical representation for the physical imaging model which incorporates effective signal- dependent colored noise modeling. Moreover, this statistical representation is combined with supervised and unsupervised learning-based regularizers which enable us to capture complex and spatially-varying elasticity patterns as opposed to fixed regularizers. Furthermore, for noisy scenarios where no ground-truth displacement images and elasticity images are available, we propose a new self-supervised learning-based elasticity reconstruction framework that benefits from the data-driven gradient correction scheme and data-driven image denoiser as prior. We use fixed-point approaches and variants of gradient descent for solving this optimization task following learning-based plug-and-play (PnP) priors and regularization by denoising (RED) paradigms. Finally, we demonstrate the performance of the proposed approaches for ultrasound elastography and MRE.