High resolution brain PET image reconstruction with deep progressive learning

Positron emission tomography (PET) can be used for the diagnosis and therapeutic monitoring of brain tumors and neurodegenerative diseases. However, brain PET images typically require high resolution and high image quality to obtain sufficient diagnostic value due to the complex and delicate anatomical structure of the brain. It further limits the use of low-dose and fast acquisition schemes on conventional PET systems. In this work, we proposed a deep progressive learning based super-resolution reconstruction (DPR) method for brain PET imaging. Specifically, a denoising network and an enhanced super-resolution network were designed and used in the reconstruction. The algorithm was evaluated using an XCAT digital brain phantom based on GATE simulation, and compared to OSEM with different pixel sizes. The evaluation results indicate that DPR achieves the best performance in metrics as SSMI, PSNR, Contrast, and background COV, especially under low count conditions. In conclusion, the proposed DPR algorithm makes it possible to obtain high-resolution and high-quality brain PET images under low-dose and/or fast scanning conditions.