A Pilot Study on Quantitative Accuracy and Radiomic Feature Stability of Deep Progressive Learning Reconstruction in NSCLC PET

Deep progressive learning reconstruction (DPR) is a novel deep learning-based algorithm for PET imaging, yet its impact on quantitative metrics and radiomic feature stability is not fully characterized. This preliminary study systematically evaluated DPR against conventional ordered-subset expectation maximization (OSEM) in non-small cell lung cancer (NSCLC) PET imaging. In this retrospective study of 24 NSCLC patients, PET data were reconstructed using OSEM and three DPR strength levels. We compared standardized uptake values (SUV), contrast-to-noise ratio (CNR), and background noise. As a secondary objective, the stability of 93 radiomic features was quantified using an intra-patient coefficient of variation (COV_RF) across all four reconstruction methods. DPR significantly increased SUV, particularly in smaller tumors, but this came at the expense of image quality, with only the lowest DPR strength improving CNR. The stability analysis revealed a stark stratification of radiomic features. While 31 features (33.3%) were robust against algorithmic changes (median COV_RF ≤ 10%), a larger group of 38 features (40.9%), including the commonly used glcm_Contrast, proved highly unstable. In conclusion, DPR presents a critical trade-off between enhanced SUV quantification and image quality, requiring careful parameter optimization. Furthermore, our findings demonstrate that the stability of radiomic features is highly algorithm-dependent. The reliable application of advanced reconstruction techniques like DPR in quantitative and radiomic pipelines is therefore contingent upon a rigorous, evidence-based selection of features verified to be robust.