The Importance Of Skip Connections In Encoder-Decoder Architectures For Colorectal Polyp Detection

Abstract

Accurate polyp detection during the colonoscopy procedure impacts colorectal cancer prevention and early detection. In this paper, we investigate the influence of skip connections as the main component of encoder-decoder based convolutional neural network (CNN) architectures for colorectal polyp detection. We conduct experiments on long and short skip connections and further extend the existing architecture by introducing dense lateral skip connections. The proposed segmentation architecture utilizes short skip connections in the contracting path, moreover it utilizes dense long and lateral skip connections in between the contracting and expanding path. Results obtained from the MICCAI 2015 Challenge dataset show progressive improvement of the segmentation result with expanded utilization of skip connections. The proposed colorectal polyp segmentation architecture achieves performance comparable to the state-of-the-art under significantly reduced number of model parameters.