Exploiting network optimization stability for enhanced PET image denoising using deep image prior

PET is affected by statistical noise due to constraints on tracer dose and scan duration, impacting both diagnostic performance and quantitative accuracy. While deep learning (DL)-based PET denoising methods have been used to improve image quality, they may introduce over-smoothing, compromising quantitative accuracy. We propose a method for making a DL solution more reliable and apply it to the conditional deep image prior (DIP). We introduce the idea of stability information in the optimization process of conditional DIP, enabling the identification of unstable regions within the network's optimization trajectory. Our method incorporates a stability map, which is derived from multiple intermediate outputs of moderate network at different optimization steps. The final denoised image is then obtained by computing linear combination of the DIP output and the original reconstructed image, weighted by the stability map. Our method effectively reduces noise while preserving small structure details in brain FDG images. Results demonstrated that our approach outperformed existing methods in peak-to-valley ratio and noise suppression across various low-dose levels. Region-of-interest analysis confirmed that the proposed method maintains quantitative accuracy without introducing under- or over-estimation. We applied our method to full-dose PET data to assess its impact on image quality. The results revealed that the proposed method significantly reduced background noise while preserving the peak-to-valley ratio at a level comparable to that of unfiltered full-dose PET images. The proposed method introduces a robust approach to DL-based PET denoising, enhancing its reliability and preserving quantitative accuracy. This strategy has the potential to advance performance in high-sensitivity PET scanners, demonstrating that DL can extend PET imaging capabilities beyond low-dose applications.