Brain tumor segmentation from multimodal 3D magnetic resonance imaging (MRI) is critical for diagnosis and treatment planning of brain tumors, yet it remains challenging due to limited data and high computational de-mands. We propose LiMSAUNet, a lightweight mode-selective 3D U-Net architecture designed for efficient and accurate brain tumor segmentation. Our model employs separate encoder pathways for each of the four MRI modalities with an attention-based modality selection mechanism to fuse information, all within a compact 3D U-Net framework. To enhance performance, we pre-trained LiMSA-UNet on the large Brain Tumor Segmentation (BraTS) 2021 Challenge dataset and fine-tuned on a relatively smaller BraTS-Africa Challenge dataset, leveraging transfer learning to improve generalization. LiMSA-UNet achieved competitive legacy Dice scores of 77%, 69%, 70% for the tumor subregions (whole tumor, enhancing tumor, tumor core, respectively), using only 6.4 million parameters which are significantly fewer than many state-of-the-art models. These results demonstrate that our lightweight model, with modality-selective fusion and pre-training, can potentially segment tumors with high accuracy, while requiring much less computational resources. This makes it attractive for deployment in resource-constrained clinical settings.