Deep learning enhanced bilateral post-filtering of noisy PET data

 

Ziel/Aim PET images can exhibit high noise levels which adversely affects qualitative and quantitative image evaluation. Especially challenging are respiratory gated studies and dynamic studies. In such cases, Gaussian filtering is routinely used to improve the signal to noise ratio. However, this degrades the spatial resolution and leads to reduced contrast recovery (CR) in small lesions. Edge preserving bilateral filtering is able to overcome this shortcoming but requires careful tuning of its 2 parameters on a per case basis in order to produce optimal results. In this work we evaluate the potential of using a deep neural network for automatic edge preserving image filtering utilizing a training set of manually filtered PET images.

Methodik/Methods We collected unfiltered gated PET data from clinical PET/MR (Philips PET/MR) and PET/CT (Siemens PET/CT) systems and interactively optimized bilateral filtering to achieve the best combination of noise reduction and preservation of spatial resolution. The set of pairs of corresponding unfiltered and filtered images was randomly split into training, validation, and testing sets. The convolutional neural network (CNN) was trained to generate the filtered images from the unfiltered ones. The resulting network model was then evaluated using the ROVER software package regarding its denoising and CR performance and also for presence of artifacts.

Ergebnisse/Results With the preliminary data available so far, evaluation of the images filtered with CNN yielded results closely resembling these obtained with manually tuned bilateral filtering in terms of noise level and CR. No apparent image artifacts were found.

Schlussfolgerungen/Conclusions Our initial results indicate that the CNN-based post-filtering produces images comparable to interactively optimized filtering. However, more thorough analyses with more image data for testing and training is required to draw definite conclusions about reliably of the proposed solution and will be performed in the coming months. Furthermore, integration of the derived network into a new respiratory motion compensation framework is planned.

Publication History

Article published online:
14 April 2022

© 2022. Thieme. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany