Longitudinal Evolution of Radiomic Features in Radiation-Induced Necrosis During Follow-Up of Brain Metastases: A Pilot Study

Background: Radiomic analysis enables quantitative characterization of post-radiotherapy brain lesions beyond conventional imaging. However, the longitudinal behavior of radiomic features, particularly across different MRI scanners, remains insufficiently explored. Purpose: This pilot study aimed to explore the longitudinal evolution of radiomic features in radiation-induced necrosis following radiotherapy for brain metastases by comparing trajectories obtained from MRI follow-up performed on the same scanner with follow-up acquired across different scanners. Methods: This retrospective pilot study included 40 radiation necrosis lesions divided into a same-scanner cohort (20 lesions) and a mixed-scanner cohort (20 lesions). Each lesion underwent three sequential post-contrast T1-weighted MRI examinations, resulting in 120 longitudinal lesion-level observations. Images underwent standardized preprocessing, three-dimensional lesion segmentation, and radiomic feature extraction. Feature robustness to segmentation variability was assessed using the coefficient of variation and intraclass correlation coefficient, and only robust features were retained. Longitudinal coherence was evaluated using the Δ-feature metric quantifying longitudinal variability. Results: Seventy-six of 107 radiomic features (71%) were robust to segmentation perturbations. Substantial longitudinal variability was observed even in the same-scanner cohort. Shape features demonstrated lower longitudinal variability, whereas texture-based features showed descriptively higher variability in the mixed-scanner cohort. However, these differences were not statistically significant after false discovery rate correction. Conclusions: This pilot study highlights the importance of validating longitudinal radiomic feature stability before interpreting temporal radiomic changes as biological phenomena or incorporating radiomic biomarkers into clinical decision-support models.