RP-Unet: a Unet-based network with RNNPool enables computation-efficient polyp segmentation

The incidence of colon cancer has shown an upward trend in recent years, and the appearance of colon polyps is one of the signs of colon cancer. The detection and segmentation of colon polyps are one of the doctors' auxiliary diagnostic methods. However, the increasing number of model parameters and inference memory requirements make the engineering of polyp segmentation models a challenging task. In this paper, an efficient polyp segmentation model based on Unet and RNNPool named RP-Unet is proposed. The first two blocks consisted of two convolutional and max pooling layers in Unet are replaced with the proposed RNNPool Down and Fuse (RDF) modules to rapidly downsample and fuse the input feature maps, and they also provide feature maps for skip connection. The last two blocks in the encoder are replaced with the proposed Double Convolution with Residual connection and RNNPool (DCRR) modules, in which the convolution layers are residually connected, and the max pooling layer is replaced directly with RNNPool. In the two proposed modules, up mapping and channel mapping are used to strengthen feature propagation by mapping activation maps logically instead of allocating unnecessary memory. The proposed RP-Unet is evaluated on two polyp segmentation datasets, and experiments show that the peak inference memory is reduced by almost 22%, while the segmentation accuracy is not significantly reduced.