Automated Segmentation and Volumetry for Meningioma Using Deep Learning

U-Net-based architectures will be applied to 500 contrast-enhanced axial MR images of different patients from a single institution after manual segmentation of meningioma, of which 50 were used for testing. After preprocessing with Z-isotropification and intensity normalization of images, 3 U-Net-based networks (2D U-Net, Attention U-Net, 3D U-Net) and 3 nnU-Net-based networks (2D nnU-Net, Attention nnU-Net, 3D nnU-Net) will be trained with meningioma-segmented images. For applying to 3D networks, sagittal and coronal images will be reconstructed using axial images. After prediction, the cut-off of the probability function, which is a trade-off, will be obtained with the Gaussian Mixture Modeling algorithm using the probability density function. The voxels having a probability function higher than that will be finally predicted as meningioma. Tumor volume is calculated as the sum of the product of segmented area and thickness of axial images. For performance evaluation, dice similarity coefficient (DSC), precision, and recall will be evaluated compared with manually segmented voxels for validation datasets. The results of volumetry of each model will be compared with manual segmentation-based volume through Pearson's correlation analysis.

Inclusion criteria

• Radiologically diagnosed meningioma by MRI

Exclusion criteria

• under 18 years old
• Multiple meningiomas
• Orbital meningioma
• Any prior treatment for intracranial meningioma before registration