TY - JOUR T1 - Low-dose PET reconstruction using deep learning: application to cardiac imaged with FDG JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 573 LP - 573 VL - 60 IS - supplement 1 AU - Claes Ladefoged AU - Philip Hasbak AU - Joachim Hansen AU - Andreas Kjær AU - Liselotte Højgaard AU - Flemming Andersen Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/573.abstract N2 - 573Objectives: 18F-FDG-PET is widely used for viability testing in patients with chronic ischemic heart disease. Guidelines recommend injection of 300 MBq 18F-FDG, but reduction of radiation exposure has become increasingly important, with the goal to achieve exposures As Low As Reasonably Achievable (ALARA). Injecting a lower tracer dose might reduce the diagnostic accuracy due to the increased noise in the images. Recently, methods utilizing deep learning (DL) for reduction of this noise have been proposed for e.g. neurological applications, but the impact on the clinical accuracy has not been examined. The aim of this study was to explore the use of a common DL network for noise reduction in low-dose PET images, and to validate its accuracy using the cardiac volumes as clinical quantitative metrics in patients imaged with 18F-FDG-PET. Methods: We included 166 fully anonymized patients injected with 300 MBq 18F-FDG and imaged on a simultaneous PET/CT. We simulated, using E7-tools (Siemens Healthcare, USA), a reduced dose by keeping counts at thresholds 1% and 10%, corresponding to 3 MBq and 30 MBq, respectively. We reconstructed both static and gated images for the 1%, 10% and 100% doses. We used 3D U-net with four blocks in the encoding part, each consisting of convolution, batch normalization and reLU activation. We used strided convolution to down-sample between the blocks and a filter size increasing from 64 to 512 in the encoding part and vice-versa in the decoding part. We used residual learning to learn the difference between the standard-dose images and the low-dose images rather than the images directly. Spatial and structural information is obtained by having 16 adjacent slices from PET and CT as input to the model. We used a batch size of 2, MAE as loss function, learning rate of 10-4 and trained the models for 2000 epochs. We trained in total four models, static and gated for each dose reduction threshold. We used 20 of the patients for validation and the remaining 146 for training. During training, we randomly sampled 16 neighboring slices within a randomly selected patient. For the gated models, we also randomly selected the gate (1-8). During validation, we predicted the residual noise, which was added to the low-dose images to obtain a denoised PET image. The uncorrected and denoised images were compared in Corridor4DM (INVIA Medical Imaging Solutions, USA) to the full-dose PET images. We used the default segmentation of the ventricles to extract quantitative the metrics end-diastolic volume (EDV), end-systolic volume (ESV) and ejection fraction (EF) from the gated images, and extend from the static images. Results: Qualitatively, the denoised images resemble the full-dose images to a high degree, as is exemplified in Fig. 1 A. Quantitatively, the difference between the uncorrected and denoised low-dose relative to the full-dose images are shown in Fig. 1 B. For the 1% low-dose images, the error is most pronounced for EDV and ESV, where the average underestimation is 25%. The bias is nearly removed using the proposed DL denoising method. The EDV, ESV and EF metrics are more robust to the 10% dose reduction, with average underestimation of 2%, and only minor improvements when using the denoised images. The same results are obtained when inspecting extend. Conclusions: A significant dose reduction can be achieved for 18F-FDG images of the heart without significant loss of diagnostic accuracy when using our deep learning model for noise reduction. Both 1% and 10% dose reductions are possible with clinically quantitative metrics comparable to that obtained with a full dose. Figure 1: A: Example images over the heart region for a single representative patient, shown for static (upper row) and gated images (bottom row). B: Difference in clinical metrics when reducing dose with and without denoising using deep learning. Note, ESV, EDV and EF is given as relative percent difference between reduced and full dose image, where extend is the absolute difference in percentage-points. ER -