Fully automatic meningioma segmentation with nnUNet using T1-weighted contrast-enhanced MR images by leveraging publicly available data and different types of annotations

Meningiomas (a kind of primary brain tumor) can be automatically segmented using T1-contrast (T1c) brain scans. Moreover, public datasets are available for model training. Unfortunately, deep learning models suffer from domain shift. Therefore, this study tests if including data of the center where the model will be applied improves segmentations. Moreover, commonly only the whole tumor (comprising the enhancing tumor and tumor core) is used during training. This study tests if using multi-label segmentations comprising the enhancing tumor, the tumor core, and edema during training improves whole tumor segmentations. NnUNet was trained on different combinations of our large (n = 374) dataset of clinically collected T1c scans with whole tumor segmentations and the BraTS2023 Meningioma dataset with either whole tumor segmentations or multi-label segmentations. We employed a novel loss function using both sets of labels during training. Model performances were compared against inter-rater reliability (n = 20 patients) and example segmentations were provided. The model using both datasets and whole tumor segmentations (Dice 0.929) performed best, matching the inter-rater variability (Dice 0.925). Differences between models were small (≤0.002 in Dice score). Relatively poor test-set predictions with a Dice score at the 10th percentile required minimal corrections. Automatic segmentation was possible, matched inter-rater reliability, and required little manual corrections. Including data from the center where the model will be applied resulted in a negligible improvement, likely due to the large variance in the BraTS dataset. Using multi-label segmentations did not improve whole tumor segmentations. Future studies should test if these results translate to different centers.