R2U

R2U, short for Recurrent Residual U-Net, is a deep learning architecture designed for medical image segmentation. It extends the U-Net framework by incorporating recurrent residual convolutional neural network blocks within the encoder and decoder paths. The recurrence enables the network to iteratively refine features, effectively increasing the receptive field without proportionally increasing the number of parameters.

In each stage, the standard convolutional block is replaced with an RRCNN block that applies a small number of shared-weight convolutions over several time steps, with a residual connection around the block. This allows the network to model complex textures and boundaries while maintaining gradient flow during training. The architecture preserves the U-Net’s symmetric encoder–decoder structure and its skip connections, enabling precise localization.

R2U has been applied to various medical segmentation tasks, such as cell nuclei, gland segmentation, and other organ/tissue delineation in 2D images. Training typically uses cross-entropy or Dice loss, possibly with data augmentation, and evaluation uses metrics such as the Dice coefficient, IoU, precision, and recall.

Impact and variants: R2U-inspired modules have influenced subsequent segmentation networks that seek to balance depth and parameter efficiency. Variants may use deeper recurrence, 3D inputs, or additional attention mechanisms, while retaining the core recurrent residual philosophy.

See also: U-Net; Residual networks; Recurrent convolutional networks; Medical image segmentation.