Bilateral filter for image derived input function in MR-BrainPET

Abstract

In positron emission tomography (PET) a post-filtering step may be used to reduce image noise. For that purpose a moving average filter with a Gaussian shape is frequently used. However, such a filter decreases the spatial resolution and increases the spillover between adjacent structures. These effects become important when dealing with small structures such as the carotid arteries with the aim to derive an image derived input function (IDIF). In this work, a bilateral filter which involves the anatomical information from a segmented magnetic resonance image (MRI) is proposed. To test the filter, dynamic FDG images were simulated with GATE (Geant4 Application for Tomographic Emission) for the BrainPET scanner. To evaluate the filter, the signal to noise ratio (SNR) of the IDIF was calculated. Moreover, three approaches to estimate the IDIF were examined, which were based on: i) the carotid volume of interest (VOl) average, ii) the hottest voxels per plane in carotid VOl and iii) the hottest voxels in the carotid VOl. These were evaluated with the area under the curve (AVe) as well as with partial volume coefficients. The results show that the bilateral filter increases the SNR and reduces the differences between the simulated and estimated IDIF. In conclusion, compared to moving average Gaussian filtering the proposed filter reduces the PVE and increases the SNR.