Knowledge Distillation for Computationally Tractable Brain Tumour Segmentation in Sub-saharan Africa

Brain tumour segmentation plays a vital role in diagnosis and treatment planning, but its benefits are often inaccessible in low-resource settings, particularly in the Global South, due to the need for high-quality imaging and computationally intensive models. This paper presents a proof-of-concept segmentation system designed to perform on low-quality MRI scans and run on extremely limited hardware. The lightweight model leverages knowledge distillation from a high-performing 3D U-Net variant developed for the BraTS-Africa challenge for brain tumour segmentation. While the model achieves a low dice score of 0.09 and a moderate Hausdorff score of 103.46, this inference process is possible on a Raspberry Pi 3b - an outdated and resource-constrained device with only a single gigabyte of RAM available. This work does not propose a clinically viable system but instead demonstrates the potential of extreme model compression and architectural adaptations such as depthwise convolutional layers to enable research into accessible medical AI tools for rural and under-resourced regions.