Abstract
571
Objectives: Imaging with radionuclides remains challenging due to limitations of administered dose and scan time. Recent advancements in machine learning (ML) based denoising techniques have shown promise toward reducing image noise. We have implemented and compared results for several approaches in terms of quantitation, reduction of noise and presence of artifacts. Our goal was to recover the quantitation and image quality found in long-duration images (‘long duration’) by denoising images with ¼ statistics (‘standard clinical’ duration).
Methods: Data from 38 clinical FDG studies from a GE Discovery 710 PET/CT system were used for this evaluation. Three of the bed positions over the lungs and upper-abdomen, as part of a whole-body study, were acquired at a longer duration of 6 minutes per position. From PET list-mode data, shorter 90-second sub-scan data was generated. Default clinical settings were used for image reconstruction. ML was used to denoise standard clinical duration input images using training and testing datasets with pairs of (standard duration, long duration) images. Several machine learning algorithms were implemented, including 2D U-Net, 3D DeepMedic and a GE-developed 2D patch-based method. The number of trainable model parameters ranged from 156k to 43M, with the GE model being the simplest. Several combinations of input data were used, including single and multi-slice axial image data and PET-only or PET/CT. Model training/validation used 30 of the subjects (22 train, 8 validation) and the remaining 8 subjects were tested using the trained model. Results across all these approaches were compared against the quantitative and visual image results from the long duration images. Evaluations emphasized visual image quality (noise texture, artifacts) as well as quantitative comparison of liver noise (pixel standard deviation) and SUVmax. In the 8 testing subjects, 29 tracer-avid regions of interest were analyzed and used for comparison as well as a ~15mL volume-of-interest in the liver.
Results: The ML denoising models performed better using multi-slice 2D versus using single-slice 2D data. Also, inclusion of CT data improved performance. Interestingly, the use of a 3D model did not result in significant benefits - potentially because the multi-slice 2D was of as much use in denoising as the 3D information. In general, all denoising models reduced image noise but typically liver noise was 15-25% lower than that of the long duration images. Averaged across models, feature contrast of the denoised images was approximately 16 ± 12% lower than the feature contrast of the long duration images, whereas the standard clinical images mostly preserved feature contrast as compared to the long duration images (0 ± 6%). Changes in image artifact included the appearance of high-contrast feature borders (e.g., myocardium), sometimes with the appearance of ringing-like behavior. The use of one model with multiple different input data randomizations demonstrated an inherent quantitation and noise variation of about 5%.
Conclusions: Both 2D and 3D CNN-based denoising methods can successfully reduce image noise and improve image quality. However, the denoising models struggle to preserve small high-contrast features in the PET images, leading to reduced feature quantitation. Further approaches continue to be investigated to improve the tradeoff between maintaining feature contrast while reducing noise.