Validation of low-dose lung cancer PET-CT protocol and PET image improvement using machine learning

Purpose To conduct a simplified lesion-detection task of low-dose (LD) PET-CT protocol for frequent lung screening using 30% the effective PETCT dose and to investigate feasibility increasing clinical value low-statistics scans machine learning. Methods We acquired 33 SD PET images, which 13 had actual LD (ALD) PET, simulated (SLD) images at seven different count levels from scans. employed image quality transfer (IQT), learning algorithm that performs patch-regression map parameters low-quality high-quality images. At each level, patches extracted 23 pairs SD/SLD were used train three IQT models – global linear, single tree, random forest regressions with cubic patch sizes 3 5 voxels. The then estimate level 10 unseen subjects. Lesion-detection was carried out on matched lesion-present lesion-absent Results yielded lesion detectability sensitivity 0.98 specificity 1. Random size allowed further 11.7% reduction in without compromising detectability, but underestimated SUV by 30%. Conclusion validated detection ALD Substantial improvement or additional while preserving values can be achieved methods though quantification may biased adjustment our research is required use.