A novel and universally applicable architectural modification is proposed, can substantially enhance performance of 3D convolution-based segmentation networks by supplanting conventional skip connections with 2D U-Net-inspired connections. 3D CNNs often exhibit deficiencies in capturing inter-slice feature commonalities due to anisotropic voxel spacing inherent in 3D medical imaging. To address this limitation, we propose 2D U-Net connection that harnesses 2D feature extraction capabilities on 3D UNets. A dual feature fusion module then incorporates input features undergoes upsampling with concomitant channel depth augmentation at each upsampling stage. This architecture can be instantiated on any 3D U-Net backbones. Empirical validation on benchmark datasets, including FLARE, OIMHS, and FeTA, reveals that our proposed method achieves state-of-the-art results, markedly outperforming extant methods. The simple yet efficacious connection can be seamlessly incorporated into most existing 3D segmentation architectures.