NIMG-74. A NOVEL APPROACH FOR PERSONALIZED RADIOTHERAPY TARGET VOLUME DEFINITION USING METABOLIC MRI AND DEEP LEARNING-DRIVEN PREDICTIONS OF GLIOBLASTOMA RECURRENCE

Abstract BACKGROUND Machine-learning models have demonstrated great promise in predicting tumor infiltration beyond anatomical margins in glioblastoma. Yet standard-of-care (SOC) radiation therapy (RT) planning only utilizes a uniform, isotropic 2cm-expansion of the T1-post-contrast-lesion or T2-lesion to generate a clinical target volume (CTV), without considering heterogeneity in tumor infiltration. We hypothesize that using a novel deep-learning approach to predict regions of tumor progression with metabolic and diffusion-weighted MRI will result in personalized CTVs that are more sensitive to detecting normal-appearing infiltrating tumor beyond the T2-lesion, sparing more healthy brain compared to SOC-CTVs. METHODS Anatomical, diffusion-weighted, and metabolic MRI from 101 patients with glioblastoma, acquired post-surgery but prior to chemoradiation, were retrospectively used to predict regions of new contrast-enhancement or T2-hyperintensity at confirmed progression. A segmentation-based predictive deep-learning approach with personalized loss-functions and evaluation metrics (“Progression-Coverage-Coefficient”, PCC) that weight Dice coefficient terms by tumor size was employed with a 66/17/18 train/validation/test split of patients and the resulting DL-CTV compared to two SOC-CTVs: 1) the RTOG[T2-FLAIR+2cm]-CTV, 2) the EORTC[T1c+2cm]-CTV. RESULTS Our best deep-learning-CTV trained using anatomical, diffusion, and metabolic MRI achieved significantly improved median test specificity (0.87 vs. 0.77; p<0.005) and comparable sensitivity (0.92 vs. 0.95; p=0.1) and PCC (0.81 vs. 0.79; p=0.1) to the SOC RTOG[T2-FLAIR+2cm]-CTV, sparing more normal brain tissue. Compared to the more conservative EORTC[T1c+2cm]-CTV, our approach had significantly higher test sensitivity (0.92 vs. 0.85; p<0.00001) and PCC (0.81 vs. 0.77; p=0.003) in covering the progressed lesion, while maintaining similar specificity (0.87 vs. 0.89; p=0.3). Test performance dropped by 9.6% and 12.4% when removing metabolic and both metabolic and diffusion metrics, respectively (p<0.001). CONCLUSION This study demonstrates the benefit of using metabolic and diffusion MRI with deep learning for personalized RT planning by targeting regions of anatomically normal infiltrative disease that are highly likely to progress, while reducing treatment to normal brain.