Abstract
PURPOSE: Standard-of-care (SOC) radiation therapy (RT) planning utilizes a uniform, isotropic 2cm-expansion of the T1-contrast-enhancing or T2-hyperintensity lesion (CEL, T2L) to generate a clinical target volume (CTV), without considering anisotropic tumor infiltration. We hypothesize that incorporating vision-transformers into our segmentation-based deep-learning approach for generating CTVs from predictions of tumor progression using either pre-surgical or pre-RT metabolic and diffusion-weighted MRI will result in personalized CTVs that are more sensitive to detecting infiltrating tumor and spare more healthy brain compared to SOC-CTVs. METHODS: Anatomical, diffusion-weighted, and metabolic MRI from 193 patients with glioblastoma (92 acquired pre-surgery, 101 post-surgery, pre-chemoradiation) were retrospectively used to predict regions of new contrast-enhancement or T2-hyperintensity at recurrence (60/30/10% train/validation/test split). Two segmentation-based predictive deep-learning approaches with personalized loss-functions and evaluation metrics that incorporate tumor size were employed and performance was compared to the SOC T2L+2cm expansion CTV. 8 tissue samples with known coordinates on the progression MRI scan taken from 4 patients in the test-set were assessed to further validate model predictions. RESULTS: Our deep-learning, vision-transformer-predicted CTV achieved significantly improved coverage of the progressed lesion compared to our prior U-Net model for the presurgery test-set [0.963(95%CI:0.889,0.986) vs 0.85(95%CI:0.770,0.871); p<0.001], and similar sensitivity to both the U-Net applied to pre-RT test-set (0.95 vs 0.92) and SOC-T2L+2cm CTV [0.985(95%CI:0.84,0.996)], but with a smaller mean treatment volume [361.82cm(3) vs 373.91cm(3)]. For 3 test-set patients with research scans pre-surgery, pre-RT, and at progression allowing for direct comparison between pre-surgery and pre-RT models, 9-80% and 14-75% increases in sensitivity and personalized Progression-Coverage-Coefficients were consistently observed for pre-surgery models for all patients. All 8 tissue samples from regions of suspected recurrence were confirmed as recurrent tumor and located within the predicted CTV; 75% of which were outside the original pre-treatment CEL. CONCLUSION: This study demonstrates the potential benefit of using pre-surgical metabolic and diffusion MRI with a segmentation-based, deep-learning approach that utilizes vision-transformers for personalized RT planning.