MenUnet: An End-to-End 3D Neural Network for Meningioma Segmentation from Multiparametric MRI

Meningioma is one of the most common and serious intracranial tumors. Multiparametric MRI (mpMRI) has been shown to provide significant information for precise tumor diagnosis. Manual segmentation of the meningioma is both lengthy and challenging. Automatic segmentation with high accuracy is of high interest. However, the state-of-art studies investigating automatic meningioma segmentation are unsatisfying in accuracy. An end-to-end network, MenUnet, is proposed to directly segment the whole 3D volume of MRIs. It captures the correlation among the surrounding slices in the 3D volume for accurate meningioma segmentation. It is validated in the public Brain Tumor Segmentation (BraTS) 2023 Meningioma Challenge, achieving the 5th in the validation phase. The dice scores of Enhancing Tumor (ET), Tumor core (TC), and whole tumor (WT), respectively, are 82.62%, 82.21%, and 83.00%. In the testing phase, the dice scores of Enhancing Tumor (ET), Tumor core (TC), and whole tumor (WT), respectively, are 83.01%, 82.01%, and 76.05%. The results show that MenUnet has the potential to assist the clinical practice of meningioma diagnosis.