| Popis: |
Photon counting detectors used in spectroscopic CT are often ba sed on small pixels and therefore o er onlylimited space to include energy discriminators and their associated co unters in each pixel cell. For this reason,it is important to make e cient use of the available energy discriminato rs in order to achieve an optimizedmaterialcontrast at a radiationdose aslow as possible. Unfortuna tely, the complexity of evaluatingeverypossiblecombination of energy thresholds, given a xed number of counter s, rapidly increases with the resolution at whichthis search is performed, and makes brute-force approaches to this problem infeasible. In this work, we introducemethods from machine learning, in particular sparse regression, to perform a feature selection to determineoptimal combinations of energy thresholds. We will demonstrate ho w methods enforcing row-sparsity on a linearregression's coe cient matrix can be applied to the multiple response problem in spectroscopic CT, i.e. the casein which a single set of energy thresholds is sought to simultaneously r etrieve concentrations pertaining to amultitude of materials in an optimal way. These methods are applied to CT images experimentally obtainedwith a Medipix3RX detector operated in charge summing mode and with a CdTe sensor at a pixel pitch of110 m. We show that the least absolute shrinkage and selection operato r (lasso), generalized to the multipleresponse case, chooses four out of 20 possible threshold position s that allow discriminating PMMA, iodine andgadolinium in a contrast agent phantom at a higher accuracy than wit h equally spaced thresholds. Finally, weillustrate why it might be unwise to use a higher number of energy thre sholds than absolutely necessary.Keywords: spectroscopic CT, photon counting, threshold optimization, cont rast agents, material discrimina-tion, Medipix, sparse regression |
