Related Experiment Video
Updated: Aug 13, 2026
Reproducible 3D Glioblastoma Migration Assay with Magnetic Nanoparticle Mediated Spheroid Localization Under Hypoxic Conditions
Published on: May 12, 2026
A single model for glioblastoma segmentation with and without T2-FLAIR: independent validation of a targeted dropout
Marco Öchsner1, Lena Kaiser2, Robert Stahl1
1Institute of Neuroradiology, LMU University Hospital, LMU, Munich, Germany.
Objectives:
To evaluate targeted T2 fluid-attenuated inversion recovery (T2-FLAIR) dropout for robust automated glioblastoma segmentation and volumetry when T2-FLAIR is unavailable, while preserving performance when the full MRI protocol is available.
Materials And Methods:
In this retrospective multi-dataset study, 3D nnU-Net models were developed on BraTS 2021 after excluding UPenn-GBM cases (remaining n = 848) and evaluated in the independent withheld UPenn-GBM cohort (n = 403). Models were trained with or without targeted T2-FLAIR dropout by zeroing the T2-FLAIR channel during training. Testing used prespecified T2-FLAIR-present and T2-FLAIR-unavailable scenarios, where the unavailable scenario was simulated by zeroing the T2-FLAIR channel at inference. The primary endpoint was per-patient overall region-wise Dice similarity coefficient (DSC). Secondary endpoints were region-specific DSC, 95th percentile Hausdorff distance and Bland-Altman whole-tumor volume bias.
Results:
In the UPenn-GBM validation cohort, performance was preserved with the full MRI protocol: overall median DSC was 94.8% [interquartile range (IQR) 90.0-97.1%] with 35% dropout and 95.0% (IQR 90.3-97.1%) without dropout. In the T2-FLAIR-unavailable scenario, targeted dropout improved overall median DSC from 81.0% (IQR 75.1-86.4%) to 93.4% (IQR 89.1-96.2%). Whole-tumor DSC improved from 60.4 to 92.6%, whole-tumor 95th percentile Hausdorff distance from 17.24 mm to 2.45 mm, and whole-tumor volume bias from -45.6 mL to 0.83 mL. A dedicated three-sequence nnU-Net achieved similar performance without T2-FLAIR (overall DSC 93.8%, WT DSC 93.5%), suggesting that much of this recovery reflects adaptation to the reduced-input setting rather than dropout training specifically.
Conclusion:
In the independent withheld UPenn-GBM cohort, targeted T2-FLAIR dropout preserved complete-protocol performance and remained robust when T2-FLAIR was unavailable. Because a dedicated three-sequence model matched this performance without T2-FLAIR, the value of targeted dropout lies not in superior missing-sequence accuracy but in providing a single model that operates across both complete and T2-FLAIR-unavailable inference.