DPC CT reconstruction with limited projection views can be achieved using compressive sensing methods. The reconstructed tri-signature CT images can be further used for automatic material identification by machine learning. Our studies indicate that perfect identification can be achieved using DPC tri-signatures, which offer results superior to those provided by absorption signature alone.
Pattern recognition and machine learning:
In conventional X-ray CT applications of aviation security and biomedical imaging, human inspection can be problematic under the limits of human perception and cognition, especially when it is required to detect infrequent visual target signals among high levels of background clutter/tissues/organs. In fact, this problem becomes even more complicated when the conventional X-ray CT technique suffers from metal artifacts due do beam hardening of X-rays and other effects. Therefore, new inspection techniques such as DPC CT, as well as automatic discrimination systems, are highly desirable. In our lab, pattern recognition and machine learning techniques have been applied to automatic material discrimination and identification in DPC CT images. Based on our studies, we provide evidences indicating that X-ray-CT-based material discrimination can greatly benefit from machine learning, as well as from using DPC tri-signatures instead of absorption signature alone.
Compressive sensing (CS) techniques use limited information that is acquired below Nyquist sampling rate to recover the original signals. In DPC CT imaging, we are particularly interested in reducing the number of projection views needed to perform reconstruction, and CS methods have therefore been successfully developed and applied to DPC CT. Such CS techniques not only are particularly useful for aviation security screening in the situations where only limited-view X-ray projections are available, but also greatly benefit biomedical imaging applications of DPC CT, in which dose reduction is highly desirable.
There is currently no content classified with this term.