GUNet: Duplex Global Graph Relationship Mining for Polyp Segmentation

Colorectal Cancer is one of the most common cancers worldwide, and as the probability of cure decreases as the stages progress, the early-stage detection of possible indications of colorectal cancer (i.e., polyps) during colonoscopy exams can contribute greatly to its prevention and curation. Since some of the indications of colorectal cancer are hard to identify by the human eye, it becomes crucial to develop an algorithm that can assist in the early identification of abnormal tissue walls that may be signs of early-stage cancer. This paper proposes the GUNet, where duplex global graph relationships are mined through graph convolution operations. The core of this framework is the extraction and utilization of both spatial and semantic information from diverse encoded feature maps to acquire contextual semantic information, done so to construct a fuller and more diverse global representation. We also train the GUNet on a combination of three public datasets to enhance its generalization abilities. Experiments show that the GUNet can serve as a competitive alternative to the U-Net, achieving a 64.47% IOU on the combined dataset.