ID #1008 Trajectory-Based Modeling For Pre-Treatment Prediction of Radiotherapy Response in Diffuse Midline Glioma
Daria Laslo, Thien Phuc Dac Nguyen, Benjamin Lerman, Schuyler Stoller, Truman Knowles, Dror Suhami, Nabaan Mir, Atlas Haddadi Avval, Andrea Webster Carrion, Cassie Kline, Andrea Franson, Timothy Müller, Andreas Rauschecker, Catherine Jutzeler, Sabine Mueller, Sarah BrüningkAbstract
Background
Radiotherapy (RT) is a cornerstone of H3K27-altered diffuse midline glioma (DMG) treatment, yet therapeutic responses vary and remain difficult to quantify. Conventional fixed timepoint response assessments are frequently confounded by early pseudo-progression, limiting their ability to capture comprehensive treatment benefits. We hypothesize that RT response in DMG is more accurately reflected by longitudinal tumor dynamics. This study assesses whether RT-response metrics can be predicted from pre-treatment tumor genomic profiling and MRI imaging.
Methods
We segmented whole-tumor in longitudinal multi-contrast MRIs from 139 DMG patients across four international cohorts (n = 654 images), yielding response trajectories. Two complementary RT-response labeling strategies were assessed: (i) trajectory-based classification of sustained control (≥100 days); (ii) volumetric-cutoff-based(≥25%) assessment of the MRI closest to 120 days post-RT. Pre-treatment MRIs were encoded using embeddings from a 3D-brain-MRI foundation model or handcrafted tumor radiomic features. For a patient subset (n = 99), pre-treatment tumor biopsies underwent next-generation sequencing to characterize somatic mutations and gene fusions. Machine-learning models were trained by five-fold cross-validation to predict response labels from all feature sets.
Results
For trajectory-based labelling, the best-performing model achieved an F1-score of 0.79±0.02 from radiomic tumor features. Brain-MRI foundation model embeddings did not improve performance further. We further demonstrate that the inclusion of genomic information minimally improved prediction performance (F1-score from 0.69±0.09 to 0.72±0.07) in the respective subpopulation, with TP53 and NF1 mutation status emerging as the most informative genomic predictors. Models trained to predict static volumetric response at a fixed post-RT time performed no better than random.
Conclusion
We assessed the importance of comprehensive, longitudinal RT-response metrics, which, based on predictions from pre-treatment radiographic and genomic biomarkers, could support individualized RT planning, early identification of patients unlikely to derive durable benefit, and more appropriate selection of patients for treatment intensification.