The paper introduces UNet++, a novel neural architecture for medical image segmentation that addresses two key limitations of existing models: unknown optimal depth and restrictive skip connections. UNet++ improves performance by incorporating an efficient ensemble of U-Nets with varying depths, redesigned skip connections for flexible feature fusion, and a pruning scheme to accelerate inference. The redesigned skip connections allow aggregation of features at different semantic scales, enhancing segmentation quality for varying-sized objects. The model is evaluated on six medical image segmentation datasets, including CT, MRI, and EM, demonstrating superior performance over baseline models in both semantic and instance segmentation tasks. Pruned versions of UNet++ achieve significant speedup with minimal performance degradation. The architecture is also shown to be extensible to various backbone encoders and medical imaging modalities. UNet++ outperforms the original U-Net and Mask R-CNN, and its performance is further enhanced by deep supervision, which enables model pruning during inference. The paper also highlights the benefits of collaborative learning among multi-depth U-Nets embedded within UNet++, leading to improved segmentation performance. The implementation and pre-trained models are available on GitHub.The paper introduces UNet++, a novel neural architecture for medical image segmentation that addresses two key limitations of existing models: unknown optimal depth and restrictive skip connections. UNet++ improves performance by incorporating an efficient ensemble of U-Nets with varying depths, redesigned skip connections for flexible feature fusion, and a pruning scheme to accelerate inference. The redesigned skip connections allow aggregation of features at different semantic scales, enhancing segmentation quality for varying-sized objects. The model is evaluated on six medical image segmentation datasets, including CT, MRI, and EM, demonstrating superior performance over baseline models in both semantic and instance segmentation tasks. Pruned versions of UNet++ achieve significant speedup with minimal performance degradation. The architecture is also shown to be extensible to various backbone encoders and medical imaging modalities. UNet++ outperforms the original U-Net and Mask R-CNN, and its performance is further enhanced by deep supervision, which enables model pruning during inference. The paper also highlights the benefits of collaborative learning among multi-depth U-Nets embedded within UNet++, leading to improved segmentation performance. The implementation and pre-trained models are available on GitHub.