Ceefour18/brats-africa-3dunet-segmentation

Lightweight 3D U-Net for glioma segmentation on the BraTS-Africa 2024 dataset — CPU-only training and Streamlit deployment. # BraTS-Africa 3D U-Net Segmentation **🔗 Live demo:** brats-africa-3dunet.streamlit.app Lightweight 3D U-Net for automated glioma segmentation on Sub-Saharan African MRI data (BraTS-Africa 2024), trained end-to-end on **CPU only** and deployed as an interactive Streamlit app. > **Attribution:** This project adapts the open-access SPARK / MAI Lab tutorial *"Lightweight Brain Tumor Segmentation on Low-Resource Systems: A Step-by-Step Guide with 3D U-Net"* (Oladele et al., 2025). The pipeline, 3D U-Net implementation, and Streamlit app are built on that tutorial and on Bhattiprolu's U-Net reference. See Acknowledgements for full credit and My contributions for what I added on top. This repository is a learning and portfolio project — it is **not** a validated clinical tool. --- ## Overview Automated brain tumor segmentation delineates tumor sub-regions on multi-modal MRI to support diagnosis, treatment planning, and monitoring. Most production models depend on GPUs, which limits training and skills development in resource-constrained settings. This project implements a **lightweight 3D U-Net** that trains and runs on a standard CPU, using the **BraTS-Africa 2024** dataset — a collection built specifically to improve representation of African populations in brain tumor imaging benchmarks. The goal is a reproducible, CPU-only segmentation pipeline from raw NIfTI volumes through preprocessing, training, evaluation, and deployment. The workflow is organised in four phases: 1. **Data collection, preparation & preprocessing** — load NIfTI, scale intensities, stack modalities, crop, filter low-tumor volumes, split. 2. **Model building** — a reduced-capacity 3D U-Net with patch extraction and augmentation for memory efficiency. 3. **Training & evaluation** — combined Dice + focal loss, per-region Dice / IoU / HD95 metrics, resource tracking. 4. **Deployment** — local inference script and a Streamlit app for uploading scans and downloading predictions. --- ## Dat …