Posters

Background: Pituitary adenomas are common intracranial tumours that may compress adjacent neurovascular structures, including the optic apparatus and pituitary stalk. Accurate delineation of the tumour and surrounding anatomy is critical for neuro-oncological surgical planning; however, manual segmentation is time-consuming and subject to inter-observer variability. Although contrast-enhanced MRI is frequently used, contrast administration is not always feasible, and imaging protocols vary substantially across centres.

Methods: Automated segmentation of pituitary adenomas and clinically relevant surrounding structures was evaluated using routinely acquired T1-weighted MRI. Models were trained and tested under pre-contrast, post-contrast, and mixed contrast conditions, using training datasets of 15, 26, and 52 cases. Segmentation accuracy was quantified using Dice similarity coefficients. Cross-contrast experiments were performed to assess robustness to variation in contrast protocols.

Results: For tumour segmentation, Dice scores ranged from approximately 0.67 to 0.73 across contrast conditions at the largest training size, with no statistically significant differences between pre-contrast, post-contrast, and mixed-contrast training when testing conditions were matched. Segmentation of the optic apparatus achieved Dice scores of approximately 0.47–0.57 at higher training volumes. Increasing training dataset size reduced performance variability and improved cross-contrast robustness, with absolute Dice performance gaps decreasing by up to ~75% between the smallest and largest training cohorts. Mixed-contrast training consistently demonstrated improved generalisation across contrast protocols.

Conclusion: Clinically meaningful automated segmentation of pituitary adenomas and adjacent critical structures can be achieved without dependence on contrast-enhanced MRI. Contrast-robust approaches may support wider clinical adoption of automated tools for surgical planning, particularly in heterogeneous imaging environments and resource-limited settings.

Background: Hypofractionated radiotherapy (HFRT) reduces treatment burden in glioblastoma (GBM), but whether MGMT promoter methylation and IDH status should guide hypofractionated strategy selection is unclear. We mapped molecular reporting and patient-centred endpoints in prospective HFRT trials.

Methods: Scoping review of PubMed and Cochrane Library (1997-March 2025) for prospective trials of hypofractionated stereotactic/dose-painted RT in newly diagnosed GBM. Data were charted for fractionation, biomarkers, survival, toxicity, and QoL/cognition.

Results: Four trial-level sources met criteria (three completed early-phase studies; one phase II protocol), all post-surgery with temozolomide-based therapy. Omuro et al. tested dose-painted HFSRT (6×6 Gy to enhancement; 6×4 Gy to FLAIR) plus temozolomide/bevacizumab (n=40): 1-year OS 93% and median OS 19.0 months. Azoulay et al. delivered CyberKnife 40 Gy/5 with temozolomide (n=30): median OS 14.8 months; MGMT-methylated OS 19.9 vs 11.3 months (p=0.031) and no grade ≥3 late adverse radiation effects. Pollom et al. reported dose-escalated 25–40 Gy/5 with temozolomide (n=30): median OS 14.2 months, 12-month progression incidence 64%, and HRQoL stable across most domains with communication decline mainly after progression. HSCK-010 (planned n=50) evaluates IMRT 20 Gy/10 + HSRT 30 Gy/5 with prospective QoL (EORTC QLQ-C30), cognition (MMSE), CTCAE v5.0 toxicity, and advanced imaging-supported response assessment.

Conclusion: Evidence for biomarker-informed HFRT in GBM is sparse and heterogeneous; molecular subgroup reporting beyond MGMT and linkage of toxicity/QoL/cognition to biomarker status remain inconsistent. Standardised MGMT/IDH stratification and harmonised endpoint/BED reporting, with prespecified biomarker-stratified toxicity and QoL analyses, are priorities.