Enhancing low-dose PET image recovery via edge perception and Mamba-guided network architecture

Positron emission tomography (PET) is a sensitive molecular imaging technique used extensively in cancer diagnosis, neurology, and cardiovascular disease. However, low-dose PET (LPET) imaging often results in decreased signal-to-noise ratio and loss of detail. To address this challenge, we propose ED-Mamba, a novel brain LPET image recovery network that leverages edge perception and Mamba guidance. ED-Mamba employs an edge perception module (EdPM) and an auxiliary guidance Mamba module (AGMM) to capture multi-scale information, enhance edge details, and model global dependencies. Experimental results on public brain datasets demonstrate that, compared to the current mainstream diffusion probabilistic model (DDPM), ED-Mamba increases PSNR from 25.624dB to 26.237dB (+2.39%) and SSIM from 0.963 to 0.967 (+0.42%), while maintaining a lightweight architecture with only 16.07M parameters. Furthermore, additional evaluations conducted on the patient dataset further confirm that ED-Mamba demonstrates excellent robustness and generalizability. This work highlights the potential of integrating edge perception with Mamba guidance for enhancing LPET image recovery quality. The source code is available athttps://github.com/Ethevliu/ED-Mamba.