Deep learning derived tumor infiltration maps for personalized target definition in Glioblastoma radiotherapy

Glioblastoma is routinely treated by concomitant radiochemotherapy. Current target definition guidelines use anatomic MRI (magnetic resonance imaging) scans, taking into account contrast enhancement and the rather unspecific hyperintensity on the fluid-attenuated inversion recovery (FLAIR) sequence.

We applied deep learning based free water correction of diffusion tensor imaging (DTI) scans to estimate the infiltrative gross tumor volume (iGTV) inside of the FLAIR hyperintense region. We analyzed the resulting iGTVs and their impact on target volume definition in a retrospective cohort of 33 GBM patients.

iGTVs were significantly smaller compared to standard pre- and post-operative gross tumor volume (GTV) definitions. Two novel infiltrative tumor GTVs (nGTV_PRE-OP and nGTV_POST-OP) defined as the conjunction volume of the standard GTV and the iGTV showed only a moderate increase in size compared to standard GTV definitions. On postoperative scans, the iGTV was predominantly covered by the two clinical target volume (CTV) concepts CTV_EORTC and CTV_ROTG1. A novel infiltrative tumor CTV (nCTV) [nGTV_POST-OP + 2 cm margin] was significantly smaller compared to CTV_ROTG1 but larger than CTV_EORTC. The overlap volume and conformity index demonstrated a distinct spatial configuration of the nCTV. Tumor recurrences overlapped with the iGTV in all but one patients and were completely covered by the nCTV in all patients. After reducing the margin to 1 cm recurrences coverage was at least in-field in all patients.

To conclude, free water corrected DTI scans may help to define infiltrative tumor areas of GBM that could ultimately be used to individualize RT treatment planning in terms of dose sparing or dose escalation.