דלג לתוכן (מקש קיצור 's')
אירועים

אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

event speaker icon
יוסי שטוק (הרצאה סמינריונית לדוקטורט)
event date icon
יום חמישי, 05.04.2012, 11:30
event location icon
Taub 337
event speaker icon
מנחה: Prof. Michael Elad, Dr. Michael Zibulevsky
The main problem with contemporary Computed Tomography (CT) imaging is the high radiation dose absorbed by patients during the screening. Reducing this dose may result in poor quality imaging when using the popular fast and direct reconstruction techniques. On the other hand, iterative methods powered by statistical models of the scan are better-performing in such cases, but are also very slow. To bridge the gap between these two solutions, various signal processing techniques that augment the direct reconstruction chain in different ways have been proposed. In this work we consider a brand of these techniques, which has a learning capability and involves an off-line example-based training process that improves the reconstructed images. Two state-of-the-art noise reduction techniques are adapted to the reconstcruction problem in this manner. In a different approach, an Artificial Neural Network (ANN) is invoked to build a mixture of experts for CT reconstruction, where the different experts are realized by basic reconstruction methods with varying values of a core parameter controlling their behavior. Our methods show capability of noise and artifacts reduction in low-dose CT images, effectively allowing for clinical dose savings by factor of ~4. In this supervised learning setup, we use a direct prior information in the form of high-quality reference images which serve as a training set. This is in contrast to almost all existing techniques in CT reconstruction, where at no stage the ground truth is available and the prior information is very implicit and unreliable. An important scenario considered is a Region Of Interest (ROI) reconstruction from limited scan data. Existing techniques offer substantial - up to 80% - dose savings, when compared to the standard full-scan reconstruction. We attack this problem using data-adaptive tools and propose an algorithm for local reconstruction that provides a reconstruction with excellent locality properties, rendering the more complex techniques unnecessary.