Segmenting Brain Tumors in Multi-modal MRI Scans Using a 3D SegNet Architecture.

Gliomas are the most common type of primary brain tumor, and high-grade gliomas are typically treated using a combination of chemotherapy, radiation therapy, and surgical excision. For the latter two therapy options, precise knowledge about the location of the tumor and its components is required, which can be obtained using MRI scans. Manually labeling the tumor area in those 3-dimensional images is a tedious and time-consuming task, hence major efforts have been made to provide automated segmentation. We present our solution to the BraTS 2021 challenge Task1, where we segment gliomas in MRI scans using a SegNet-based approach, achieving competitive and stable performance across tumor types and components. Compared to previous solutions using UNet architectures, our model achieves improved segmentation of the peritumoral edema and comparable performance for the other classes while reducing the number of parameters.