Segmentation models achieved expert-level performance in a large variety of medical applications. However, their robustness to rotations is rarely discussed and can be crucial for clinical use with the risk of discarding subtle but diagnostically relevant anatomical structures. In medical images, complex structures appear in a wide range of positions and rotations, requiring rotation robustness. In this work, we investigate the robustness to rotations of a standard 3D nnU-Net in the context of two segmentation tasks: the hippocampus in MRI and the pulmonary airway system in CT. In addition, we introduce a 3D Locally Rotation Invariant (LRI) operator based on the bispectrum to achieve high robustness to input rotations. It is compared to a standard nnU-Net, a nnU-Net with extended rotational data augmentation and XEdgeConv, a state-of-the-art approach for RI. While all models performed similarly in terms of Dice score for right-angle rotations, the Bispectral U-Net outperformed other designs in the context of finer and more realistic rotations. Furthermore, the Bispectral U-Net and the XEdgeConv were more stable w.r.t. input rotation, i.e. the predictions are significantly more consistent across input rotations. Important inconsistencies of the nnU-Net were observed for lung airway segmentation, suggesting potential risks of using the model in clinical routine.