2695 - Gradient-Based Radiomics Analysis for Outcome Prediction in Personalized Ultra-Fractionated Stereotactic Adaptive Radiotherapy (PULSAR)

Purpose/Objective(s):

The study introduces a novel gradient-based radiomics framework by differentiating peritumoral and intratumoral regions for outcome prediction. By leveraging spatial segmentation and temporal changes, the framework is anticipated to improve its predictive power and help enhance interpretability.

Materials/Methods:

Magnetic resonance imaging (MRI) data from 39 patients with 69 brain metastases (BMs) treated by Gamma Knife were retrospectively analyzed, including both pretreatment and intra-treatment scans. Three gradient-based features—gradient magnitude, radial gradient, and radial deviation—from intratumoral and peritumoral regions. Peritumoral regions were defined based on tumor volume and segmented into eight octant margins, for large tumors (>4000 mm³) with a 3-mm margin expansion/contraction, and for small tumors (<4000 mm³) with a 1-mm expansions/contractions. Features were quantified across core, margin, and octant regions using mean, standard deviation, and coefficient of variation. Fifteen support vector machine (SVM) models were tested to classify whether lesions exhibited =20% volume reduction at 3-month follow up, plus an ensemble feature selection (EFS) model integrating top-performing features. Statistical significance was evaluated using Welch’s t-test with Bonferroni correction (p<0.002). The framework was also performed on a non-PULSAR fractionated stereotactic radiotherapy (fSRT) cohort to assess its generalizability.

Results:

The EFS model integrating both peritumoral and intratumoral gradient features achieve the best performance (AUC: 0.995; F1 score: 0.940). Models based solely on peritumoral octant features significantly outperform those utilizing only features from intratumoral core or peritumoral margin. Features extracted from the pretreatment MRI demonstrate greater discriminative power compared to models using intra-treatment MRI or delta features. In non-PULSAR cohort, the gradient-based model (AUC: 0.933) is found to outperform the conventional radiomics model (AUC: 0.839).

Conclusion:

Our gradient-based radiomics approach, combining spatial segmentation and temporal features, help simplify feature selection, improve outcome prediction and dynamic decision making. Its superior performance compared to conventional radiomics, coupled with its effectiveness in both PULSAR and non-PULSAR cohorts, highlights its potential as a robust tool for personalized treatment planning in neuro-oncology.