TY - JOUR T1 - Noise reduction with cross-tracer transfer deep learning for low-dose oncological PET JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 108 LP - 108 VL - 60 IS - supplement 1 AU - Hui Liu AU - Jing Wu AU - Wenzhuo Lu AU - John Onofrey AU - Yi-Hwa Liu AU - Chi Liu Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/108.abstract N2 - 108Objectives: Deep convolutional neural networks can be robust and effective in noise reduction for low-dose FDG PET thanks to large amount of training datasets. However, clinically, PET datasets from new or uncommonly used tracers may not be adequately available and acquisitions of high-count images from tracers with long half-lives such as Zr-89 are difficult. Therefore, image de-noising using a deep learning method for these tracers could be challenging. In this work, we investigated the feasibility of transferring a pre-trained neural network from widely-used FDG tracer to a less frequently used tracer, fluoromisonidazole (FMISO), for image noise reduction. Methods: A 3D deep convolutional neural network using a U-Net architecture was adopted for noise reduction. The network was trained to minimize an L2 loss function using the Adam optimizer. A patch size of 64x64x16 was used for training and testing. Our datasets consisted of two groups of PET lung studies acquired with a Siemens Biograph mCT (9 scans for the FDG group and 11 scans for the FMISO group). The scans were performed for 90 min after 10 mCi FDG injection for the FDG group and were for 120 min with 5 mCi FMISO injection for the FMISO group. The list-mode data ranging from 60 to 80 min for the FDG group and from 70 to 120 min for the FMISO group were extracted to generate full-count sinograms. Subsequently, 10 samples of low-count sinograms were obtained by independent 10% down-sampling of the full-count list-mode data. The extracted low-dose sinograms had similar noise levels between both groups. The resulting sinograms were reconstructed using the OSEM algorithm with 3 iterations and 21 subsets. The image size was 400×400×109 with 2.04×2.04×2.03 mm3 voxel size, and was cropped to 300×180×70 for training. In 3D U-Net training, low-count images were used as the input with full-count images as the target. Three U-Net networks were trained and obtained: FDG U-Net trained with the FDG group data from scratch, fine-tuned U-Net based on pre-trained FDG U-Net with fine-tuning using 1 scan of the FMISO group as transfer learning, and FMISO U-Net trained with 9 scans of the FMISO group from scratch. The remaining 2 scans of the FMISO group were used to generate the predicted de-noised images for testing. The quantification results inside nodule region of interest (ROI) were calculated, and the predicted ROI mean values obtained from FDG U-Net and fine-tuned U-Net were compared with those from FMISO U-Net across 10 samples for each scan. The ROI relative bias (using full-count image as ground truth) and background noise for the predicted images of the three U-Nets were also calculated and compared with those obtained from post-reconstruction anatomical-guided non-local mean (ANLM) filtering. Results: All the three U-Nets were capable of reducing the noise of low-count images to the level similar to the full-count images. For both FMISO test scans, there was no significant difference in the profiles of the predicted images obtained from FDG U-Net, fine-tuned U-Net and FMISO U-Net. Furthermore, when quantitatively comparing FDG U-Net, fine-tuned U-Net with FMISO U-Net, the average correlation coefficient of the predicted ROI means across 10 samples was higher than 0.96 for the two scans and the regression line was close to the line of identity in the linear regression analysis. There was underestimation for the predicted ROI mean value in the de-noised low-count images when compared with full-count images, whereas the underestimations were similar for all the three U-Nets. Relative bias versus background curves demonstrated that all three U-Nets performed better than the ANLM method in terms of lower background noise at the same bias level, indicating that the feasibility of both FDG U-Net and fine-tuned U-Net for FMISO image noise reduction. Conclusions: Transfer deep learning across different tracers for low-dose PET noise reduction is feasible, which may be beneficial for tracers that are not widely used and with long half-lives. ER -