Assessing Respiratory Motion Stability of Novel ¹⁸F-Fluorodeoxyglucose Positron Emission Tomography-Derived...

Background/Objectives: Novel hotspot displacement radiomic features (normalized hotspot-to-centroid distance [NHOC]/normalized hotspot-to-perimeter distance [NHOP]) are robust against image resampling and spatial resolution variations. However, their reproducibility under respiratory motion remains unvalidated. This study aimed to evaluate the reproducibility, reliability, and survival prognostic value of NHOC/NHOP features in thoracic ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) images with and without respiratory motion correction and to determine whether these features maintain stability and predictive performance for overall survival (OS) compared with respiratory-stable reference features. Methods: We analyzed 138 patients (203 lesions) who underwent ¹⁸F-FDG PET/CT with and without data-driven respiratory gating. Reproducibility and reliability were assessed using the coefficient of variation (CoV) and intraclass correlation coefficient (ICC), respectively. OS prediction was evaluated using Cox regression and concordance index (c-index) analyses. Results: Except for NHOCmax and NHOPpeak, which showed ICC values of 0.782 and 0.93, respectively, the novel morphological features generally exhibited poor reproducibility and moderate reliability (CoV > 20% and ICC < 0.75). In contrast, reference features (entropy-based and sphericity) demonstrated excellent robustness. Motion-corrected NHOCmax showed significant OS prediction for both spatially resampled and non-resampled images. No significant differences in c-indices were observed between motion-corrected and non-corrected features. Conclusions: The marked sensitivity of novel hotspot-displacement features to respiratory motion substantially limits their clinical applicability in thoracic disease. To ensure reproducibility and generalizability in future research, prioritizing inherently robust radiomic parameters, such as entropy-based features, is strongly recommended.