A regional convolution kernel algorithm for scatter correction in dual-energy images: comparison to single-kernel algorithms

Single kernel scatter correction algorithms are based on the model that the scatter field can be predicted by convolution of the primary intensity (Iprim) with a spatially invariant scatter point-spread function (PSF). Practical limitations (Iprim unknown) suggest the substitution of the total detected intensity (Idet) for Iprim as the source image in the convolution. In regions of high scatter fraction (SF), Idet is a poor approximation of Iprim, thereby causing an overestimation of scatter originating in the region. This contributes to errors in estimating detected scatter in the mediastinum and neighboring regions. A technique using a regionally variable point-spread function that significantly reduces RMS error in estimation of the primary image as compared to the single PSF method is investigated. The regionally variable convolution method employs a larger PSF in the mediastinum and a smaller PSF in the lungs to reduce the error in estimating the scatter throughout the image. The method to allow for patient differences has also been expanded and various implementations of these methods have been compared. Results show that the dual-kernel algorithm is always more effective than an equivalent single-kernel algorithm. The dual-kernel algorithm using a predicted scatter fraction curve gives an overall RMS error in the primary of as low as 20.8% which is equivalent to 8.7% RMS error in the scatter. The dual-kernel method using a predicted scatter fraction curve approaches the accuracy of the single-kernel method using patient specific scatter measurements. Because using individual scatter measurements is a less desirable method for clinical use, we feel that the dual-kernel algorithm which uses two regions specific convolution kernels and a variable scatter fraction curve is the preferable method.