Brain tumors, particularly gliomas, pose substantial challenges in clinical diagnosis, primarily due to the inherent difficulties in achieving accurate segmentation of magnetic resonance imaging (MRI) scans. Traditional manual segmentation techniques are not only labor-intensive but also susceptible to inter-observer variability, which further complicates the diagnostic process. This issue is particularly pronounced in resource-limited settings, such as Sub-Saharan Africa (SSA), where the quality of neuroimaging data may vary significantly. While machine learning-based segmentation approaches are increasingly applied for automated brain tumor detection, the absence of representative and diverse datasets constrains their generalizability to local clinical conditions and patient populations.
To address these challenges, we employ the Swin UNETR architecture and its attention mechanisms. The attention mechanisms are more effective at leveraging smaller datasets with less overfitting, making SwIn UNETR a potential efficient approach for the limited datasets in SSA. In this study, the Swin UNETR model was pre-trained on anatomical brain MRI scans from Brain Tumor Segmentation (BraTS) Adult Glioma dataset (N=1251), then fine-tuned using the BraTS Africa dataset (N=60), and subsequently evaluated on a separate BraTS Africa test set (N=35). Our model achieved an overall Dice score of 0.73, highlighting the model’s promising performance in segmenting tumor subregions. This model can be further fine tuned and evaluated on clinical cases across SSA to confirm the potential of novel vision transformer models across clinical settings, particularly in resource-limited settings.