Colonoscopy is an effective technique for detecting colorectal polyps, whichare highly related to colorectal cancer. In clinical practice, segmentingpolyps from colonoscopy images is of great importance since it providesvaluable information for diagnosis and surgery. However, accurate polypsegmentation is a challenging task, for two major reasons: (i) the same type ofpolyps has a diversity of size, color and texture; and (ii) the boundarybetween a polyp and its surrounding mucosa is not sharp. To address thesechallenges, we propose a parallel reverse attention network (PraNet) foraccurate polyp segmentation in colonoscopy images. Specifically, we firstaggregate the features in high-level layers using a parallel partial decoder(PPD). Based on the combined feature, we then generate a global map as theinitial guidance area for the following components. In addition, we mine theboundary cues using a reverse attention (RA) module, which is able to establishthe relationship between areas and boundary cues. Thanks to the recurrentcooperation mechanism between areas and boundaries, our PraNet is capable ofcalibrating any misaligned predictions, improving the segmentation accuracy.Quantitative and qualitative evaluations on five challenging datasets acrosssix metrics show that our PraNet improves the segmentation accuracysignificantly, and presents a number of advantages in terms ofgeneralizability, and real-time segmentation efficiency.