Testing diffusion-derived orientation priors for streamline modelling of the MRI-visible glioblastoma core: the BRIAN framework

Abstract

Glioblastoma spreads diffusely beyond the abnormality visible on conventional magnetic resonance imaging, and because tumour cells migrate preferentially along white-matter pathways, growth models that assume isotropic spread may misrepresent the geometry of invasion relevant to radiotherapy planning. This work introduces and evaluates BRIAN, a diffusion-informed simulator that biases tumour propagation along directions derived from diffusion MRI. Patient tumour masks from the UPENN-GBM cohort are transferred onto healthy host brains from the Human Connectome Project through the MNI152 template as a proxy, diffusion orientation distribution functions (dODFs) are reconstructed on each host, and stochastic streamline propagation with the MRtrix3 iFOD2 algorithm yields volumetric occupancy maps. Propagation parameters are fitted per tumour under a three-stage curriculum that tightens a constrained five-metric objective. Across 30 unifocal tumours, each propagated onto 65 validation hosts (1 950 tumour-host pairings), the simulator reached a per-tumour median Dice coefficient of 0.748 (0.745 pooled across all pairings) and a median bounded Hausdorff agreement of 0.776. Because parameters are calibrated against each tumours own reference mask and the train/validation split is over hosts, these figures quantify reproduction and host-transfer of a known lesion; prediction of unseen tumours is outside their scope. On a purposively selected ten-tumour subset, controlled comparisons tested both the contribution of directional information and whether a richer angular reconstruction improved performance. Relative to a direction-blind isotropic null, orientation-informed tracking improved all five evaluated metrics (dz = 0.4-1.2), although only surface Dice remained significant after Holm correction. A second comparison replaced the multi-shell SHORE dODF with a single-tensor dODF while keeping the tracking framework unchanged. No statistically detectable differences were observed between the two directional models on this subset, providing no evidence that resolving crossing fibres improved agreement with the visible tumour envelope. Directional sampling improved agreement with the MRI-visible tumour core, but the selected subset provided no evidence that the SHORE dODF outperformed the tensor dODF.