2026 Poster Presentations
P148: AUTOMATIC RESIDUAL MENINGIOMA SEGMENTATION USING DEEP LEARNING: AN EXPLORATORY APPROACH IN ARTIFICIAL INTELLIGENCE
Marc-Olivier Comeau; Mohamad Khalili; Guilherme Gago, MD; François Gascon; Martin Cote, MD; Pierre-Olivier Champagne, MD, PhD; Université Laval
Context: The monitoring of postoperative residual meningioma requires long-term imaging surveillance for detection of progression, which is both time and resource intensive. While convolutional neural networks such as nnU-Net provide a robust segmentation of large well-defined lesions, complex cases involving small lesions adjacent to iso-intense neighboring structures remain a significant challenge. This study represents the first attempt in automatic artificial intelligence segmentation of residual meningioma disease.
Methods: This retrospective cohort study included 32 adult patients who underwent intracranial meningioma resection between 2010 and 2021 with a postoperative finding of residual disease on imaging. Preoperative and all postoperative T1-weighted gadolinium enhanced MRI sequences were exported under DICOM format for segmentation. Segmentations were performed on each MRI by the model and by two independent observers, whose volume measurements were averaged into a single value. A modified nnU-Net model comprising of an ensemble of a Tversky-modified 2D model, and a baseline 3D model were utilized. The Tversky-modified 2D component aimed to address any class imbalance and improve recognition of small lesions, while the 3D component aimed to provide volumetric awareness. Model performance was evaluated against the observers through volumetric error, relative and absolute errors, and the Dice Similarity Coefficient. For all tests, 95% confidence intervals were utilized and p-values < 0.05 were considered significant.
Results: A total of a 176 MRI segmentations were performed, 32 preoperative and 144 postoperative. Mean preoperative meningioma volume were 36.6 cm3 (± 33.2) with the model and 34.8 cm3 (± 35.0) with the observers. Mean residual meningioma volumes were 4.6 cm3 (± 8.7) with the model and 4.1 cm3 (± 7.8) with the observers. Mean relative volumetric errors were 5.1 cm3 (± 10.1) (max: 51.2) preoperatively and 1.7 cm3 (± 1.5) (max: 8.5) postoperatively. Absolute errors were greater on residual disease, with means of 32.1% (± 65.9) (max: 341) preoperatively and 108.4 % (± 136.8) (max: 867) postoperatively (p = 0.006). Errors on residual disease trended down as volumes increased (fig. 1).
Fig. 1. Relative error against observer remnant volume.
Dice Coefficient decreased significantly on residual disease, with means of 0.73 (± 0.30) preoperatively and 0.31 (± 0.35) postoperatively (p < 0.001). In remnants, the model segmentation had no overlap with the observers’ segmentations in 73 cases (50.7%) (Dice Coefficient = 0). Dice Coefficient was significantly higher in remnants with volumes above the sample mean of 4.1 cm3 (p < 0.001) (fig. 2).
Fig. 2. Dice Coefficient against A) observer remnant volume and B) volumes below and above the mean.
Absolute errors and Dice Coefficient according to remnant location are displayed below (fig. 3).
Fig. 3. A) Absolute error and B) Dice Coefficient according to remnant location.
Conclusion: This exploratory study displayed the significant underperformance of our novel deep learning model in the segmentation of residual meningioma disease. While we were able to replicate the clinically relevant performance previously demonstrated in large lesions, small complex lesions remain a significant challenge. Future directions include the addition of different MRI sequences.