Brain Tumor Segmentation: Glioma Segmentation in Sub-Saharan Africa Patients Using nnU-Net

Accurate and automatic segmentation of glioblastoma multiforme (GBM) is crucial for effective treatment planning, disease diagnosis, surgical planning, and brain tumor tracking. MRI and other imaging modalities are used to capture the complex nature of GBM, however the intrinsic heterogeneity in tumor features adds to the difficulty of segmentation. Consequently, advanced computational algorithms and machine learning approaches are being developed to improve the accuracy and efficiency of GBM segmentation. Recently, deep learning-based U-Net architecture has been the state-of-the-art method for segmenting medical images. This work builds on our prior work on U-Net for GBM segmentation and proposes state-of-the-art nnU-Net for adult glioma segmentation in the brain tumor segmentation challenge-2023 (BraTS-Africa Adult Glioma). The nnU-Net allows training of different multiple networks: 2D U-Net, 3D U-Net to perform semantic segmentation of 3D images with high accuracy and performance. We find the best configurations from different variants of nn-UNet and combine an ensemble of these variants to improve segmentation performance. We utilize 3D U-Net to perform segmentation of brain tumor for BraTS Africa dataset. In addition, we compare the result with our 3D UNetcontext encoding model which is pretrained on BraTS Adult glioma data and then fine-tuned on BraTS Africa dataset. The accuracy of our proposed nn-UNet for brain segmentation from multi-modal MRI is evaluated using a 5-fold cross-validation over 15 manually segmented images from the BraTS 2023 challenge. The mean of the lesion-wise Dice Similarity Coefficient (DSC) of the BraTS -Africa validation dataset is 0.7445, 0.7244, and 0.8526 for Enhance Tumor (ET), Tumor Core (TC), and Whole Tumor (WT), respectively. The performance of our model on the BraTS-Africa data set indicates that higher segmentation accuracy may be attained utilizing the latest nn-UNet method.