2441 - Automated Radiation Response Assessment with Deep Learning and Radiomics in Pediatric Diffuse Midline Glioma

Purpose/Objective(s): Diffuse Midline Glioma (DMG) is an aggressive, treatment-resistant brain tumor that predominantly affects pediatric and young adult populations. Tumors have heterogeneous cellular responses to radiation therapy (RT) and quantitative imaging features (i.e. radiomics) may help assess tumor heterogeneity and predict response. We hypothesized that an automated segmentation and radiomics extraction pipeline could guide RT response assessment and prognosis.

Materials/Methods: We curated a single institution pediatric DMG cohort with pre-/post-RT, and longitudinal T2 MRI. We applied a previously validated deep learning-based segmentation pipeline to define tumor regions, followed by a PyRadiomics pipeline with bias-field correction, intensity normalization, and standardized bin width, to extract radiomic features. To ensure spatial stability, tumor segmentations were randomly perturbed (tolerance: 1-3 mm), retaining features with coefficients of variation (CV) < 0. Highly correlated features (r > 0.95) were removed, and the rest were standardized. 3D volumetrics pre-/post-RT were grouped by RAPNO criteria: partial response (³25% decrease), progressive (³25% increase), or stable. A Random Forest model was trained for pre- vs. post-RT classification to determine radiomics most indicative of RT receipt using a 10-fold stratified cross-validation and evaluated with the area under the ROC curve (AUC). Feature importance was assessed by mean decrease in Gini. Predictors of overall survival (OS) were evaluated using Cox regression.

Results: Following curation, 66 patients (441 MRIs) were identified (median follow-up: 315 days [range: 13 – 1739]); 34 had died at 1-year (52%). Mean 3D volume change pre- to post-RT was -6.9% (std: 24.5%). Of patients, 47 (71%) had partial response, 10 (15%) were stable, and 9 (13.6%) had progressive disease. Partial response was associated with improved survival vs. stable disease (median OS: 326 days vs 172 days, p<0.001). A Random Forest classifier using 377 radiomic features achieved an AUC of 0.84 for RT prediction. Analysis of feature importances revealed five top predictive features (all textural features). The inclusion of the five RT-predicting delta radiomic features did not improve survival prediction compared to volumetric response alone (C-index: 0.55 vs 0.61).

Conclusion: Automated tumor volumetrics can inform RT response in DMG and volumetric change with RT predicts prognosis. Textural radiomic features are associated with RT delivery, but larger studies with multi-institutional data will be necessary to determine if radiomics add value beyond volumetrics.