Iterative guided image filtering for multimodal medical imaging

Abstract

Image denoising is an important task in many medical imaging applications. There have been many developed algorithms such as Gaussian kernel based filtering, anisotropic diffusion, bilaterial filtering, non local means filtering, and total variation based denoising. In the computer vision community, the guided image filtering (GF) has been introduced, which demonstrated powerful denoising performance and fast computation time. GF was also extended to Iterative GF to improve the filter output quality (e.g. rolling guidance filtering or RGF). Recently, GF was applied to PET denoising problem in simultaneous PET-MR with matched spatial resolution of MR. In this work, we propose a new iterative GF method called iterative guided image filtering (IGF). Unlike RGF, our new method uses an intermediate guide image. The proposed method was evaluated using a toy example as well as SPECT-CT Monte Carlo (MC) simulation with XCAT phantom. When there were matched edges between noisy image and guide image, then both GF and IGF showed significant improvement in image quality due to edge-preserving property of GF for toy example images. However, when there are mismatched structures between noisy and guide images, severe blurring was observed in GF result image, but relatively good edge information recovery was observed and IGF using side information image yielded the best the root mean squared error (RMSE). Similar tendency was observed in the SPECT-CT MC simulation. Both GF with CT side information and IGF without CT information yielded about 40% reduction in RMSE. When both IGF and CT information were used, 55% RMSE reduction was achieved. Our proposed method can potentially be applied to many multimodal medical imaging applications including PET-CT, SPECT-CT, and simultaneous PET-MR.