Transformer-based architectures have recently emerged as powerful models for 3D medical image segmentation, with UNETR demonstrating notable success by leveraging self-attention to capture global contextual information. Despite its effectiveness, UNETR still faces challenges in interpretability and computational efficiency. To address these limitations, we propose UKANETR, a novel UNET variant that employs kansformers, i.e., transformers equipped with Kolmogorov–Arnold Network multilayer perceptrons (KAN-MLPs). By integrating KAN-MLPs into the transformer blocks, UKANETR enhances model interpretability while improving its ability to represent complex nonlinear functions with fewer parameters. In addition, UKANETR incorporates a squeeze-and-excitation mechanism within the skip connection pathways to explicitly capture channel interdependencies. This design improves feature recalibration, allowing the network to emphasize informative channels and suppress less relevant ones, thereby strengthening multi-level feature fusion between encoder and decoder stages. The proposed UKANETR was rigorously evaluated on the BRaTS 2020 brain tumor MRI dataset. Comparative analysis against the traditional UNETR and other state-of-the-art models shows that UKANETR consistently achieves superior performance in multi-level brain tumor segmentation, demonstrating improvements in Dice score, sensitivity, and boundary delineation accuracy. These results highlight the potential of integrating kansformers and channel attention mechanisms into transformer-based UNET models, offering a more interpretable and efficient solution for 3D medical image segmentation tasks.