Title: Deep Learning-based Segmentation of Cerebral Aneurysms in 3D TOF-MRA using Coarse-to-Fine Framework

Abstract: BACKGROUND AND PURPOSE: Cerebral aneurysm is one of the most common cerebrovascular diseases, and SAH caused by its rupture has a very high mortality and disability rate. Existing automatic segmentation methods based on DLMs with TOF-MRA modality could not segment edge voxels very well, so that our goal is to realize more accurate segmentation of cerebral aneurysms in 3D TOF-MRA with the help of DLMs. MATERIALS AND METHODS: In this research, we proposed an automatic segmentation framework of cerebral aneurysm in 3D TOF-MRA. The framework was composed of two segmentation networks ranging from coarse to fine. The coarse segmentation network, namely DeepMedic, completed the coarse segmentation of cerebral aneurysms, and the processed results were fed into the fine segmentation network, namely dual-channel SE_3D U-Net trained with weighted loss function, for fine segmentation. Images from ADAM2020 (n=113) were used for training and validation and images from another center (n=45) were used for testing. The segmentation metrics we used include DSC, HD, and VS. RESULTS: The trained cerebral aneurysm segmentation model achieved DSC of 0.75, HD of 1.52, and VS of 0.91 on validation cohort. On the totally independent test cohort, our method achieved the highest DSC of 0.12, the lowest HD of 11.61, and the highest VS of 0.16 in comparison with state-of-the-art segmentation networks. CONCLUSIONS: The coarse-to-fine framework, which composed of DeepMedic and dual-channel SE_3D U-Net can segment cerebral aneurysms in 3D TOF-MRA with a superior accuracy.