Abstract
1828
Objectives 18F-FDG PET images are usually filtered after reconstruction to improve the overall signal-to-noise ratio (SNR) and facilitate visual interpretation. In clinical practice this is often performed using a Gaussian filter (GF) which is known for degrading contrast and spatial resolution. In this work we compared the potential improvement achieved by replacing GF with edge-preserving filters (EPF) recently proposed for PET denoising.
Methods Two EPF with parameters previously optimized for PET were considered, namely the bilateral (BF) (1) and the wavelet-curvelet (WCF) (2) filters. Simulated PET acquisitions of the NEMA-94 phantom, as well as a dataset of 15 tumor clinical images reconstructed using OSEM without post-reconstruction filtering, were used. The performance of the applied filters was compared using different figures of merit, including SNR improvements used to quantify denoising, while shifts in i) local contrast, ii) absolute quantitative intensities and iii) heterogeneity quantification through textural features, were used to evaluate the preservation of qualitative and quantitative characteristics of reconstructed images.
Results GF led to higher SNR improvements (25±7%) over both EPFs (19±5% vs. 17±6% for BF and WCF respectively), although it was associated with substantial loss of local contrast (-38±12%), quantitative mean intensity bias (-33±8%) and loss of heterogeneity information (-15±9%). Both EPF led to images with higher quantitative accuracy and preservation of structural information such as edges and intratumor heterogeneity. WCF nonetheless significantly outperformed BF with -4.4±3.8%, 0.1±1.2% and -5.1±3.2% vs. -11.2±7.2%, -1.8±3.4% and -9.3±7.1%, for loss of local contrast, quantitative bias and loss of heterogeneity respectively.
Conclusions Although GF improved SNR in reconstructed PET images, it was associated with a significant loss of quantitative information. Both BF and WCF provided significant SNR improvement and were more efficient at ensuring the quantitative image accuracy compared to Gaussian filtering, although WCF outperformed BF.
Research Support None.