Brain tumor segmentation is vital for diagnosis and treatment planning but remains challenging due to variations in tumor characteristics across MRI scans. While deep learning models like 3D U-Net have shown promise, training on all four standard MRI modalities (T1, T2, T1ce, FLAIR) is computationally expensive. Transfer learning offers a practical solution, yet most pre-trained 3D U-Net architectures accept only three-channel inputs, complicating multimodal integration. This study investigates optimal modality combinations for brain tumor segmentation using a 3D U-Net with VGG19, DenseNet201, and ResNet50 backbones pre-trained on ImageNet. We propose strategies to incorporate multimodal inputs while preserving the benefits of transfer learning. Results show that combining T1ce, T2, and FLAIR consistently yields the best performance, with DenseNet201 achieving the highest IoU of 81.16% and Mean DSC of 84.14%. Compared to recent methods, our model improves Mean DSC and closely matches top-performing approaches, confirming its robustness and accuracy. These findings highlight the potential of our model as a reliable and efficient tool for brain tumor segmentation with clinical relevance.