Synthetic CT Generation via Variant Invertible Network for Brain PET Attenuation Correction

Attenuation correction (AC) is essential for the generation of artifact-free and quantitatively accurate positron emission tomography (PET) images. Nowadays, deep-learning-based methods have been extensively applied to PET AC tasks, yielding promising results. Therefore, this paper develops an innovative approach to generate continuously valued CT images from non-attenuation corrected PET images for AC on brain PET imaging. Specifically, an invertible neural network combined with the variable augmentation strategy that can achieve the bidirectional inference processes is proposed for synthetic CT generation. On the one hand, invertible architecture ensures a bijective mapping between the PET and synthetic CT image spaces, which can potentially improve the robustness of the prediction and provide a way to validate the synthetic CT by checking the consistency of the inverse mapping. On the other hand, the variable augmentation strategy enriches the training process and leverages the intrinsic data properties more effectively. Therefore, the combination provides for superior performance in PET AC by preserving information throughout the network and by better handling of the data variability inherent PET AC. To evaluate the performance of the proposed algorithm, we conducted a comprehensive study on a total of 1480 2D slices from 37 whole-body 18F-FDG clinical patients using comparative algorithms (such as Cycle-GAN and Pix2pix). Perceptual analysis and quantitative evaluations illustrate that the invertible network for PET AC outperforms other existing AC models, which demonstrates the feasibility of achieving brain PET AC without additional anatomical information.