Realizing 32-time Scan Duration Reduction of 18F-FDG PET Using Deep Learning Model with Image Augmentation

Purpose: 32-time scan duration reduction of 18F-Fluorodeoxyglucose (18F-FDG) Positron Emission Tomography (PET) images through the generation of standard scan duration images using a multi-slice cycle-consistent Generative Adversarial Network (cycle-GAN) was studied. Also, the effect of the image augmentation methods on the performance of the cycle-GAN model was evaluated. Materials and Methods: Four subsets of standard and 32-time short scan duration PET image pairs, each contacting image data of 10 patients were used to train and test (80 percent for training and 20 percent for testing) a multi-slice cycle-GAN separately. Another patient’s image data was used as the validation dataset for different training subsets. When training the cycle-GAN model for each subset, two approaches were followed: with and without image augmentation. Common image quality metrics of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Normalized Root Mean Squared Error (NRMSE) were used to assess the generation performance of the cycle-GAN model. Paired sample t-test statistical testing with a confidence interval of 0.95 was used to determine whether the differences between approaches were statistically significant or not. Results: For subsets 1-3, both training approaches improved the image quality of the short scan duration inputs (p < 0.001) while for subset 4 only the training approach with image augmentation was capable of improving the image quality. However, the training approach with image augmentation offered better results than the approach without image augmentation (p < 0.001). Conclusion: Employing the training approach with image augmentation, the cycle-GAN model was capable of improving the image quality of 1/32nd short scan duration images through the generation of synthetic standard scan duration images. In the case of the training approach without image augmentation, except for subset 4, the model trained on all subsets 1-3 was capable of improving the image quality. Image augmentation does indeed improve the performance of the cycle-GAN model, especially in the case of insufficient available training datasets.