2026 Poster Presentations
P139: DEEP LEARNING SEGMENTATION OF SMALL AND IRREGULAR MENINGIOMAS USING NNU-NET ENSEMBLING AND TVERSKY FOCAL LOSS: AN EXPLORATORY APPROACH
Mohamad Khalili, BEng, MMA; Pierre-Olivier Champagne, MD, PhD, FRCSC; Martin Côté, MD, FRCSC; Laval University Faculty of Medicine
BACKGROUND: Accurate segmentation of meningiomas on MRI is essential for precise volumetric assessment and surgical planning. While convolutional neural networks such as nnU-Net provide strong baseline performance in biomedical image segmentation, segmenting small and irregularly-shaped meningiomas, commonly encountered in neurosurgical practice, remains a significant challenge for these baseline models. These more complex cases are clinically critical, yet cause standard models to underperform due to difficulties in contour recognition and volumetric delineation. This study explores ensemble modeling of 2D and 3D nnU-Net models with the incorporation of a Tversky Focal Loss function to improve segmentation performance on such challenging cases.
METHODS: We retrospectively gathered 120 meningioma cases from the 2024 multicentric BraTS dataset, comprising MRI scans with neuroradiologist annotations. Segmentation was performed on T1-weighted post-contrast (T1c) MRI images with gadolinium. A 5-fold cross-validation strategy (80% training, 20% validation) was employed to ensure robustness. Baseline nnU-Net 2D and 3D models were trained and benchmarked against modified models implementing a Tversky Focal Loss function (α = 0.3, β = 0.7, γ > 1). The loss function was designed to penalize false positives more heavily, encouraging improved recognition of meningioma voxels, prioritizing specificity. Ensemble models were constructed by averaging voxel-wise probabilities from 2D and 3D models. Model performance was evaluated using the Dice similarity coefficient, with 95% confidence intervals and Mann-Whitney U testing applied to address skewed distributions and small sample sizes. Final model testing was done on unseen T1c MRI scans from CHU de Québec-Université Laval.
RESULTS: Baseline models achieved mean Dice scores of 0.803 (2D) and 0.752 (3D). Incorporating Tversky Focal Loss modestly improved the 2D model (mean Dice 0.814). The more important effect was a consistent narrowing of interquartile ranges with improved performance in lower quartiles, indicating enhanced performance on irregular or small lesions. Gains in the 3D models were inconsistent (mean Dice 0.749), reflecting its tendency to prioritize volumetric continuity at the expense of fine contour recognition. Ensembling a Tversky-modified 2D model with a baseline 3D model yielded the most robust performance, balancing volumetric awareness with contour sensitivity: In folds where baseline models were weakest, improvements exceeded 4%. Testing the ensemble model on the unseen cases demonstrated clinically significant performance in recognizing small meningiomas and delineating irregular tumour margins. Twelve unseen cases had meningioma volumes between 2.6 and 8.4 mL, with a mean Dice of 0.743. On meningiomas larger than 8.4 mL, mean Dice was 0.832.
CONCLUSIONS: This exploratory study demonstrates that augmenting nnU-Net with Tversky Focal Loss yields incremental yet clinically meaningful improvements in the segmentation of small and irregular meningiomas, particularly within 2D models and ensemble configurations. While the overall magnitude of Dice score gains was limited, the improvements in lower-performing cases highlight the potential for loss function engineering and ensembling strategies to address clinically challenging segmentation tasks. Future directions include expanding training cohorts, integrating additional MRI sequences (T2, FLAIR), and a sensitivity analysis of the Tversky Focal Loss function’s parameters.
KEYWORDS: meningioma, MRI, nnU-Net, segmentation, deep learning, Tversky
