High precision segmentation of glioma and surroundings – a feasibility study using multiparametric MRI and deep learning

Glioblastoma (GBM) is an aggressive primary brain cancer in which precise spatial characterisation of tumour sub-compartments and surrounding anatomy may support research into targeted treatment. We designed and assessed a feasibility pipeline for subject-specific localisation of glioma and brain regions from standard multiparametric MRI (T1, T1-Gadolinium, T2, FLAIR), combining deep-learning segmentation with robust anatomical parcellation rather than atlas coregistration. Coregistered images yield three tumour compartments: central non-enhancing/necrotic tumour, enhancing tumour, and surrounding edoema. From this multichannel representation, we derive tumour volumes and regional tumour burden “hit-plots” that show, at the voxel level, which brain regions each compartment intersects. In a cohort of n=50 UCSF-PDGM glioblastoma subjects, deep-learning segmentations agreed closely with the reference masks: median Dice of 0.90 (whole tumour), 0.94 (tumour core), and 0.86 (enhancing tumour), with corresponding HD95 values of 4.1, 2.2, and 2.0 mm. In one longitudinal LUMIERE subject (Patient-048, six timepoints), the pipeline recovered concordant volumetric trajectories and evolving regional-burden patterns with respect to specific brain structures, in agreement with independent comparator segmentations. This subject-specific regional tumour profile may in the future serve as a descriptive anatomical context for radiotherapy target delineation.