Combining Deep CNNs and Superpixels for Polyp Segmentation in Colonoscopy Images

Polyps can be missed as a result of fatigue or inattention during prolonged or consecutive colonoscopies, with an estimated average rate of 16.8-28%. Consequently, patients with undetected polyps are less likely to survive late-stage colorectal cancer than those without polyps. The detection of polyps in colonoscopy images has been automated using a variety of approaches in order to address this issue. As a result, polyps will be prevented from being overlooked by humans, and the time and expense associated with screening numerous colonoscopies or wireless capsule endoscopies will be reduced. Our primary goal is to develop a system that precisely defines the location of polyps on an image, therefore increasing overall detection accuracy and enabling timely identification of colorectal cancer. In this work, we aim to combine UNet architecture with the pretrained model Inception-Resnet-V2 in the Encoder, and then a phase of segmentation Refinement based superpixels with Simple Linear Iterative Clustering (SLIC) for superpixels generation. The proposed methods’ average Dice on CVC-Clinic612 dataset is 0.8903, with MIoU and MPA of 0.8085 and 0.9796 respectively, and the average Dice on Kvasir-SEG dataset is 0.8758 with MIoU and MPA of 0.7936 and 0.9567 respectively. Comparing the results to state-of-the-art methods, the results show high performance.