Unsupervised deep learning strategies for noise reduction in 18F-FDG brain PET imaging

Introduction: High noise levels would limit the clinical value and diminish the certainty of findings in PET imaging. In this work, unsupervised noise reduction approaches were compared with conventional post-reconstruction filtering in PET imaging. The aim of this study is to propose and evaluate two unsupervised deep learning-based noise reduction strategies in ¹⁸F-FDG brain PET imaging and compare them with conventional denoising filters.

Methods: Unsupervised noise reduction techniques rely on the concept of noise-to-noise training and knowledge transformation. In the noise-to-noise technique, the raw PET data were divided into two independent 50% subset dataset (two independent 50% low-dose PET images). Then, a deep learning-based technique was trained in a noise-to-noise manner to learn noise reduction of PET images. The knowledge transformation technique relies on a model trained to estimate full-dose PET images from the corresponding 50% noisy versions. Subsequently, the trained models, noise-to-noise, and knowledge transformation were applied to the full-dose PET images to reduce noise and maintain the underlying signals in the original PET images. The noise-to-noise approach (trained to estimate two independent 50% low-dose images), as well as the knowledge transformation technique, rely on the fact that the underlying PET signals are the same (or at least there is a high correlation) in 50% low-dose versions as well as the full-dose PET image, while the noise component would be different (with very low correlation) in these images. Hence, the model would learn to discriminate the underlying signals from the added noise compartments. The performance of these two techniques was compared with conventional post-reconstruction Gaussian, Bilateral, and non-local mean (NLM) filters. Forty ¹⁸F-FDG clinical brain PET/CT studies were used to evaluate/compare the performance of the abovementioned noise reduction techniques.

Results: Superior performance of the noise-to-noise approach (relying on two 50% low-dose versions) was observed in the clinical studies wherein a reasonable compromise was established between levels of smoothness and activity bias in the different brain regions. The noise-to-noise approach led to a quantification bias of 1.4±2.1% while the knowledge transformation technique, Gaussian, Bilateral, and NLM filters led to quantification baises of 2.2±3.0%, -11.5±4.7%, -5.8±3.9%, -3.0±3.8%, respectively, in the different regions of the brain (p<0.05).

Conclusions: This study demonstrated the effectiveness of the unsupervised noise-to-noise network training for noise reduction in PET imaging (compared to conventional post-reconstruction filters and conventional network training). The merit of this technique is that it does not require a ground-truth dataset for training.

• Download figure
• Open in new tab
• Download powerpoint