Author: Michael Dohopolski, Jiaqi Liu, Hao Peng, Robert Timmerman, Zabi Wardak, Haozhao Zhang 👨🔬
Affiliation: Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose:
This study introduces a gradient-based radiomics framework to enhance outcome prediction in Personalized Ultra-Fractionated Stereotactic Adaptive Radiotherapy (PULSAR) for brain metastases (BMs). By leveraging spatial segmentation and temporal changes, this approach provides robust, interpretable insights for personalized treatment planning.
Methods:
MRI data from 39 patients with 69 BMs treated with PULSAR were retrospectively analyzed, including pretreatment scans (prior to fraction one) and intra-treatment scans (after three fractions and before the final two). Gradient-based features—gradient magnitude (GM), radial gradient (RG), and radial deviation (RD)—were extracted from intratumoral and peritumoral regions, segmented into eight octant margins. Tumor masks were manipulated using 3 mm expansions/contractions for large tumors and 1 mm for small tumors, aligning with 90% and 50% isodose lines. Features were quantified across core, margin, and octant regions using mean, standard deviation, and coefficient of variation, yielding 90 features. Analyses evaluated these regions independently and in combination. Fifteen Support Vector Machine (SVM) models classified lesions by ≥20% volume reduction, with an ensemble feature selection (EFS) model integrating top-performing features. Validation was performed on a non-PULSAR fractionated stereotactic radiotherapy (fSRT) cohort to assess the generalizability of proposed gradient-based method.
Results:
The EFS model achieved the highest performance (AUC: 0.995; F1: 0.940), outperforming individual models. Gradient-based features from octant margins demonstrated superior predictive power compared to core and margin features alone. Pre-treatment models outperformed intra-treatment and delta-mode models. In non-PULSAR cohort, gradient-based features (AUC: 0.933) exceeded conventional radiomics (AUC: 0.839), highlighting their robustness and applicability across datasets.
Conclusion:
Gradient-based radiomics leverages dose-driven volume and tumor morphology to extract biologically meaningful features from both intratumoral and peritumoral regions. By integrating gradient metrics, this framework provides enhanced interpretability and predictive power, advancing precision oncology. This framework enhances interpretability and predictive power, supporting adaptive radiotherapy and potentially improving treatment outcomes in photon and particle therapy.