Dual domain spatial-transform smoothing of whole-body PET images

Objectives: Positron emission tomography (PET) images commonly suffer from the high noise level, thus resulting in poor signal-to-noise ratio (SNR). This factor adversely affects PET imaging performance in terms of lesion identification and quantification. Denoising of PET images is a challenging task owing to the inherent low SNR. Therefore, different denoising techniques have been introduced in either spatial or transform (wavelet) domains, achieving different levels of success. In this work, we propose a hybrid dual-domain PET denoising method taking advantage of both spatial and transform domain filtering to preserve image patterns and minimize quantification uncertainty.

Methods: Spatial domain techniques excel at preserving high-contrast patterns while transform domain filtering outperforms other methods in recovering low contrast details of the image. We exploited the non-local mean (NLM) method for the spatial domain filtering due to its promising performance in denoising high contrast features and multi-scale curvelet denoising for transform domain filtering due its capability to recover small details smoothed out by NLM denoising. The proposed hybrid method was implemented as a post-reconstruction technique and compared to commonly used Gaussian and edge preserving bilateral filters. Computer simulations of phantom containing small lesions, physical Jaszczak phantom and clinical studies of patients presenting with small lesions were used to evaluate the different PET denoising techniques. The evaluation was performed using a number of metrics including the SNR, quantification bias and contrast to noise ratio (CNR) improvement.

Results: The proposed hybrid filter resulted in SNR increase from 8.1 (non-filtered PET image) to 35.6 for small lesions in computer phantom studies, while Gaussian and bilateral filtering led to SNRs of 24.3 and 28.6, respectively. Moreover, the bias decreased from 19.6% for Gaussian filtering to 9.3% using the proposed hybrid method, while the bilateral filter achieved a bias of 14.5%. The CNR in the physical Jaszczak phantom improved from 5.08 using the Gaussian smoothing to 7.42 and 13.04 using the bilateral and proposed hybrid filters, respectively. Comparison between the proposed and conventional filters using clinical PET studies demonstrated that hybrid filtering outperforms the other approaches (Fig. 1).

Conclusion: The proposed hybrid smoothing method proved to improve lesion contrast, SNR and reduces quantification bias compared to Gaussian and bilateral filtering approaches using simulated, experimental phantom and clinical studies. The proposed approach could be utilised as an alternative for post-reconstruction filtering of clinical whole-body PET images. Research Support:  This work was supported by the Swiss National Science Foundation under grant SNFN 31003A-149957 and the Swiss Cancer Research Foundation under Grant KFS-3855-02-2016.

• Download figure
• Open in new tab
• Download powerpoint