Automated brain tumor segmentation using roto-translation equivariant network

Abstract

Reliable delineation of tumor boundaries in magnetic resonance imaging (MRI) is a prerequisite for surgical planning and treatment monitoring, yet conventional convolutional networks remain sensitive to the orientation and positional variability inherent to clinical acquisitions. We propose a segmentation network built on a modified Special Euclidean group, SE(2), that embeds rotation and translation symmetry directly within its layers. It is utilizing a lifting operation, group convolutions, and group transposed convolutions in the decoder. Evaluated on the public BraTS2020 benchmark, the proposed network reached a Dice coefficient of 0.9503 and an Intersection-over-Union of 0.9616, clearly exceeding U-Net and self-configuring nnU-Net baselines while preserving sharp boundary localization. These findings proves that encoding geometric priors offers a practical route to stable, high-fidelity tumor segmentation suitable for integration into clinical imaging workflows.