The paper introduces UNet++, a novel neural architecture designed to address the limitations of existing U-Net and fully convolutional network (FCN) models in medical image segmentation. The key contributions of UNet++ include: 1. **Ensemble of U-Nets**: UNet++ incorporates an ensemble of U-Nets of varying depths, allowing for flexible network depth selection and sharing of encoders among the U-Nets. This ensemble approach leverages deep supervision to train the U-Nets simultaneously, improving overall segmentation performance. 2. **Redesigned Skip Connections**: The skip connections in UNet++ are redesigned to aggregate features of varying semantic scales at the decoder sub-networks, providing a more flexible and effective fusion scheme. This redesign enables the model to handle objects of different sizes more efficiently. 3. **Model Pruning**: UNet++ introduces a pruning scheme that removes redundant decoder layers during inference, significantly reducing inference time while maintaining only modest performance degradation. This pruning is achieved through deep supervision, which allows the model to be deployed in two modes: ensemble mode and pruned mode. 4. **Collaborative Learning**: The architecture facilitates collaborative learning among the constituent U-Nets, where each U-Net is trained with gradients from both strong (loss from ground truth) and soft (losses from adjacent deeper nodes) supervision. This enhances the performance of both shallow and deeper networks. The paper evaluates UNet++ on six medical image segmentation datasets, covering various imaging modalities such as CT, MRI, and electron microscopy. The results demonstrate that UNet++ consistently outperforms baseline models in semantic segmentation tasks and enhances the segmentation quality of varying-size objects. Additionally, Mask RCNN++ (a version of Mask R-CNN with UNet++ design) outperforms the original Mask R-CNN in instance segmentation tasks. The implementation and pre-trained models are available online.The paper introduces UNet++, a novel neural architecture designed to address the limitations of existing U-Net and fully convolutional network (FCN) models in medical image segmentation. The key contributions of UNet++ include: 1. **Ensemble of U-Nets**: UNet++ incorporates an ensemble of U-Nets of varying depths, allowing for flexible network depth selection and sharing of encoders among the U-Nets. This ensemble approach leverages deep supervision to train the U-Nets simultaneously, improving overall segmentation performance. 2. **Redesigned Skip Connections**: The skip connections in UNet++ are redesigned to aggregate features of varying semantic scales at the decoder sub-networks, providing a more flexible and effective fusion scheme. This redesign enables the model to handle objects of different sizes more efficiently. 3. **Model Pruning**: UNet++ introduces a pruning scheme that removes redundant decoder layers during inference, significantly reducing inference time while maintaining only modest performance degradation. This pruning is achieved through deep supervision, which allows the model to be deployed in two modes: ensemble mode and pruned mode. 4. **Collaborative Learning**: The architecture facilitates collaborative learning among the constituent U-Nets, where each U-Net is trained with gradients from both strong (loss from ground truth) and soft (losses from adjacent deeper nodes) supervision. This enhances the performance of both shallow and deeper networks. The paper evaluates UNet++ on six medical image segmentation datasets, covering various imaging modalities such as CT, MRI, and electron microscopy. The results demonstrate that UNet++ consistently outperforms baseline models in semantic segmentation tasks and enhances the segmentation quality of varying-size objects. Additionally, Mask RCNN++ (a version of Mask R-CNN with UNet++ design) outperforms the original Mask R-CNN in instance segmentation tasks. The implementation and pre-trained models are available online.