Reweighted ℓ1 referenceless PRF shift thermometry

Abstract

Temperature estimation in proton resonance frequency (PRF) shift MR thermometry requires a reference, or pretreatment, phase image that is subtracted from image phase during thermal treatment to yield a phase difference image proportional to temperature change. Referenceless thermometry methods derive a reference phase image from the treatment image itself by assuming that in the absence of a hot spot, the image phase can be accurately represented in a smooth (usually low order polynomial) basis. By masking the hot spot out of a least squares (ℓ(2)) regression, the reference phase image's coefficients on the polynomial basis are estimated and a reference image is derived by evaluating the polynomial inside the hot spot area. Referenceless methods are therefore insensitive to motion and bulk main field shifts, however, currently these methods require user interaction or sophisticated tracking to ensure that the hot spot is masked out of the polynomial regression. This article introduces an approach to reference PRF shift thermometry that uses reweighted ℓ(1) regression, a form of robust regression, to obtain background phase coefficients without hot spot tracking and masking. The method is compared to conventional referenceless thermometry, and demonstrated experimentally in monitoring HIFU heating in a phantom and canine prostate, as well as in a healthy human liver.