PCRNet: Parent–Child Relation Network for automatic polyp segmentation

Abstract

Colorectal cancer (CRC) is the third most common cancer worldwide in terms of both incidence and mortality rates. On the other hand, its slow development process is very beneficial for early diagnosis and effective treatment strategies in reducing mortality rates. Colonoscopy is considered the standard approach for early diagnosis and treatment of the disease. However, detecting early-stage polyps remains challenging with the current standard colonoscopy approach due to the diverse shapes, sizes, and camouflage properties of polyps.

To address the issues posed by the different shapes, sizes, colors, and hazy boundaries of polyps, we propose the Parent–Child Relation Encoder Network (PCRNet), a lightweight model for automatic polyp segmentation. PCRNet comprises a parent–child encoder branch and a decoder branch equipped with a set of Boundary-aware Foreground Extraction Blocks (BFEB). The child encoder is designed to enhance feature representation while considering model size and computational complexity. The BFEB is introduced to accurately segment polyps of varying shapes and sizes by effectively handling the issue of hazy boundaries.

PCRNet is evaluated both quantitatively and qualitatively on five public datasets, demonstrating its effectiveness compared to more than a dozen state-of-the-art techniques. Our model is the most lightweight among current approaches, with only (5.0087) million parameters, and achieves the best Dice Score of (0.729%) on the most challenging dataset, ETIS. PCRNet also has an average inference rate of (36.5) fps on an $Intel$® $Cor{e}^{TM}$ i7-10700K CPU with 62 GB of memory, using a GeForce RTX 3080 (10 GB).