Deep Learning-based Segmentation of Cerebral Aneurysms in 3D TOF-MRA using Coarse-to-Fine Framework
This work addresses the need for more accurate automatic segmentation of cerebral aneurysms in medical imaging, which is crucial for diagnosis and treatment planning, though it appears incremental as it builds on existing networks like DeepMedic and U-Net.
The paper tackled the problem of inaccurate edge voxel segmentation in cerebral aneurysms using 3D TOF-MRA by proposing a coarse-to-fine deep learning framework, achieving a Dice Similarity Coefficient (DSC) of 0.75 on validation and outperforming state-of-the-art methods on an independent test set with a DSC of 0.12.
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.