Comparison of network structures for deep learning-based cardiac PET scatter correction

Le résumé fourni par la source

Recently, deep learning-based methods have emerged as potential alternatives to conventional scatter correction techniques, particularly in static PET applications. In this study, we evaluate the efficacy of two prominent deep neural network architectures, U-Net and ResNet, for dynamic cardiac PET scatter correction. These architectures are selected for their capacity to handle intricate data relationships and reduce noise, aiming to optimize PET reconstruction and improve diagnostic efficacy in cardiac imaging.We retrospectively selected three clinical Rb-82 myocardial perfusion PET scans: two datasets containing stress and rest scans were allocated for training, while the remaining dataset with only rest scans was designated for testing. Each dataset was reconstructed into 15 frames with and without model-based scatter correction. The frames without scatter correction served as inputs for the U-Net and ResNet networks during training, with the corresponding scatter-corrected frames serving as labels.During testing on the remaining dataset, we calculated the Mean Absolute Relative Difference (MARD) for the body region and assessed local relative differences (RD) in specific volumes of interest (VOI) for each frame. Across the 15 frames, the MARDs for U-Net and ResNet were 6.16% $\pm 3.14 \%$ and $5.45 \% \pm 1.61 \%$, respectively. For RDs in the left ventricle, U-Net achieved $5.91 \% \pm$ $38.8 \%$, while ResNet achieved $1.94 \% \pm 23.09 \%$. In the myocardium VOI, U-Net exhibited RDs of $-4.92 \% \pm 12.95 \%$, whereas ResNet showed RDs of $1.92 \% \pm 8.7 \%$.In conclusion, the ResNet architecture demonstrated lower quantitative error and superior edge preservationin the dynamic cardiac PET scatter correction task compared to the U-Net architecture in our testing dataset. Further studies are necessary to validate these findings with additional data and to explore the impact of these results on time-activity curves and parameter estimates derived from the dynamic images.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.