M.Sc Thesis

M.Sc StudentTsiper Shahar
SubjectPseudo-Polar Based CT Reconstruction
DepartmentDepartment of Electrical and Computer Engineering
Supervisor PROF. Yonina Eldar
Full Thesis textFull thesis text - English Version


X-ray computerized tomography (CT) is the most widely used imaging modality in medicine today.

Since its invention in the early 1970's, CT scanners have come a long way.

Today's scanners can perform full body scans in under 30 seconds, incorporating detector arrays with hundreds of elements that acquire more than a thousand different readings in every revolution of the scanner.

Despite recent progress in scanner manufacturing, there are many issues that impede CT scanners, such as high radiation dosage involved in each scan.

Although the X-Ray detectors incorporated into the scanners have tremendously evolved and improved throughout the years, the reconstruction algorithms used in scanners today have remained largely unchanged since CT's initial invention, more than 30 years ago.

The main goal in this research, is to explore new methods for processing tomographic measurements and improve image reconstruction, while keeping the radiation dosage involved to a minimum.

We aim to harness known concepts from generalized sampling and convex optimization, for implementation and adoption on actual CT devices.

The majority of reconstruction methods used in commercial systems are mostly one-step algorithms, based on the approximation of the discrete X-Ray readings to the continuous inverse Radon transform.

Unfortunately, these one-step algorithms are known to induce various artifacts on the reconstructed CT images, and generally require a large number of measurements to produce faithful reconstructions, which lead to increased radiation dosage.

In recent years, iterative reconstruction approaches, based on modern optimization theory, have been implemented and used in many imaging modalities with great success.

These iterative algorithms have been also adapted to CT imaging, and were shown to recover axial scans from fewer tomographic measurements, leading to a significant reduction in the X-Ray dosage involved.

The main reason one-step algorithms are still widely used, despite the higher radiation dosages associated with them, is due to the high computational and memory complexity that modern iterative methods require, rendering them infeasible on currently available hardware.

We propose a way to connect a recent theoretical advancement for processing discrete tomographic measurements, known as the pseudo-polar radon transform, to actual CT devices.

Our method offers a way to alleviate much of the computational burden that state-of-the-art iterative approaches require, by relying on modern sampling methods and optimization techniques.

The measurements are transformed to the pseudo-polar domain while incorporating prior information derived from the geometrical properties of the CT measurements.

This allows for higher reconstruction speeds even when operating under higher measurement noise settings.

Finally, we demonstrate a complete process that first transforms a reduced number of raw measurements and then reconstructs an axial scan with clinical quality from them, showing a potential reduction in radiation dosage while using our method.