Token Mixer Placement Matters: A Systematic Encoder–Decoder Ablation Study of Mamba for Brain Tumour Segmentation on BraTS-Africa

Accurate 3D brain tumour segmentation requires both fine-grained local feature extraction and effective modelling of long-rangespatial dependencies. While convolutional neural networks excel at lo-cal feature extraction, they often struggle to capture global context.Conversely, Transformer-based models improve contextual modellingbut incur prohibitive memory costs in 3D settings. State space modelssuch as Mamba have emerged as a promising alternative by offeringlinear-complexity sequence modelling, yet their optimal placement withinvolumetric medical image segmentation networks remains insufficientlyunderstood. Compounding this issue, many commonly used benchmarksdo not adequately reflect the imaging and clinical heterogeneity encoun-tered in real-world practice. We present a systematic encoder-decoderablation of Mamba-based architectures for brain tumour segmentation onBraTS-Africa 2025, a 146-case multi-centre dataset representing diversesub-Saharan African imaging environments. Using a unified MetaUNETRframework, we evaluate controlled CNN and Mamba configurations andperform encoder–decoder ablations to determine where Mamba con-tributes most effectively. We further benchmark the best-performingconfigurations against nnU-Net, TransUNet, and SwinUNETR. Resultsfrom this study indicate that incorporating Mamba within the encoder con-sistently yields the strongest performance among MetaUNETR variants,whereas Mamba-based decoding was associated with poorer boundaryreconstruction, particularly for WT, within the evaluated configuration.These findings demonstrate that token mixer placement is critical, suggest-ing that Mamba-based encoders may provide a promising architecturalalternative for developing brain tumour segmentation models suited forresource-constrained settings.