22/10/15
You are invited to attend a lecture
By:
Y. Kaganovsky, Duke University.
Physics-Based Computational Imaging in the Era of Big Data: An X-Ray Perspective
Physics-Based Computational Imaging is the process of reconstructing a physical object represented by a digital image, which is often high-dimensional (e.g., 3D, hyperspectral, movie of 3D images), by computational inversion of measured data that is not directly related to the imaged object. This inverse problem is inherently ill-posed due to the non-direct and incomplete nature of the measurements, model uncertainties, and noise. In my talk I will discuss in detail two examples of computational imaging: (1) X-ray computed tomography (CT), which is used for 3D medical imaging of humans/animals, quality control in industrial manufacturing, and security inspections, to name a few; (2) Coded-aperture x-ray coherent scatter imaging, which is a novel imaging modality proposed by Prof. Brady’s group at Duke University that allows one to identify the molecular structure of materials in security and medical applications. Traditional methods used for data inversion in most actual imaging systems employ one-shot methods, e.g., Fourier-based inversion or filtered back-projection. In recent years, there has been an increased interest in statistical iterative inversion methods based on minimizing some cost function that incorporates additional physical information about the system, the measured signals, and prior knowledge about the imaged object, which makes the inverse problem less ill-posed. Despite their many advantages, iterative methods are computationally much more expensive (require more CPU time and memory) than one-shot methods. This is becoming a big challenge as sensors and detectors are becoming faster and more accurate, leading to more measurements and to the accumulation of big data, which increases storage requirements and processing time. There is also an increasing demand for high resolution images, resulting in very high dimensional solution spaces (another aspect of big data), which also increases reconstruction time, e.g., in 3D x-ray CT there are billions of image voxels to be reconstructed; this becomes an even bigger problem for hyperspectral or 4D images. In medical and security applications, time may be critical, so faster inversion methods need to be developed. Another challenge in iterative methods is the determination of model parameters, which are object dependent and therefore unknown a priori. I will present recent developments in iterative inversion algorithms and computational approaches that address the above challenges. I will highlight a non-conventional method called ``Automatic Relevance Determination (ARD)’’, which originated in machine learning, for principled automatic determination of model parameters. I will present an extension of ARD to physically-based models used in CT, called ``Variational ARD’’ [1], which can accurately account for photon shot noise and Beer's law. This method is over an order of magnitude faster than previous ARD methods for big data.
[1] Y. Kaganovsky et al., “Alternating Minimization Algorithm with Automatic Relevance Determination for Transmission Tomography under Poisson Noise”, accepted to the SIAM Journal on Imaging Sciences.
Thursday, October 22, 2015, at 15:00
Room 011, Kitot building