Sparsity-Driven Ideal Observer
Image quality assessment is useful when optimizing the hardware for imaging systems. Quantities like signal-to-noise-ratio (SNR) and mean square error (MSE) have been extensively used in these optimization tasks. They are easy to compute, but they are sometimes vague and unreliable. Task-based image quality assessment avoids the problem by optimizing imaging system designs with respect to a specific classification or estimation task. Conventional task-based image quality assessment technology dictates that hardware should be optimized without consideration of the to-be-employed reconstruction method. However, in imaging systems that employ compressive sensing concepts, imaging hardware and (sparse) image reconstruction are innately coupled. Taking the relationship between hardware and software into account, we propose sparsity-driven ideal observer (SDIO). The SDIO and sparse reconstruction method are ‘matched’ in the sense that they both utilize the same statistical information regarding the class of objects to be imaged. We develop the algorithm employing tools recently developed for variational Bayesian inference. We are committed to studying its performance and features in different applications.
In medical imaging, from a Bayesian point of view, mode of the posterior distribution is commonly employed as an estimation of the to-be-imaged object. The so-called maximum a posteriori (MAP) estimation is widely used in image reconstruction tasks. However, such point estimates provide no information about the uncertainty in the computed estimate. Second-order statistics of the posterior distribution is highly desired to provide uncertainty information of the point estimate.
We designed the uncertainty estimation algorithm employing a computationally-efficient variational Bayesian inference approach. It’s much more computationally tractable than Bayesian inference methods and sampling methods such as Monte Carlo Markov Chain (MCMC) method. Our current work focus on cultivating valuable information from the uncertainty information.