Source-linked AI summary
PraNet: Parallel Reverse Attention Network for Polyp Segmentation
Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, Ling Shao
TL;DR
Polyp segmentation is important for colonoscopy-based colorectal cancer prevention but is difficult because polyps vary in appearance and their boundaries are often blurred. PraNet aggregates high-level features for global guidance and uses recurrent reverse attention to refine boundaries; experiments report strong accuracy, generalization, and real-time efficiency across five challenging datasets.
Problem
Accurate polyp segmentation matters for colonoscopy-based colorectal cancer prevention, but variable polyp appearance and blurred polyp–mucosa boundaries make it challenging.
Method
PraNet aggregates high-level features with a parallel partial decoder, generates a global map, and uses recurrent reverse attention to model boundaries and calibrate misaligned predictions.
Results
PraNet outperforms state-of-the-art approaches by more than 5% across five challenging datasets and achieves mean Dice = 0.898 on Kvasir without pre-/postprocessing.
Takeaways & Limitations
PraNet combines strong learning and generalization ability with real-time segmentation efficiency for colonoscopy images.
Abstract
from arXiv · showhide
Colonoscopy is an effective technique for detecting colorectal polyps, which are highly related to colorectal cancer. In clinical practice, segmenting polyps from colonoscopy images is of great importance since it provides valuable information for diagnosis and surgery. However, accurate polyp segmentation is a challenging task, for two major reasons: (i) the same type of polyps has a diversity of size, color and texture; and (ii) the boundary between a polyp and its surrounding mucosa is not sharp. To address these challenges, we propose a parallel reverse attention network (PraNet) for accurate polyp segmentation in colonoscopy images. Specifically, we first aggregate the features in high-level layers using a parallel partial decoder (PPD). Based on the combined feature, we then generate a global map as the initial guidance area for the following components. In addition, we mine the boundary cues using a reverse attention (RA) module, which is able to establish the relationship between areas and boundary cues. Thanks to the recurrent cooperation mechanism between areas and boundaries, our PraNet is capable of calibrating any misaligned predictions, improving the segmentation accuracy. Quantitative and qualitative evaluations on five challenging datasets across six metrics show that our PraNet improves the segmentation accuracy significantly, and presents a number of advantages in terms of generalizability, and real-time segmentation efficiency.
1 Introduction
PraNet addresses polyp-segmentation difficulty caused by variable polyp appearance and blurred polyp–mucosa boundaries. It combines coarse area prediction with reverse attention to model boundaries and calibrate misaligned predictions.
- Polyp segmentation is clinically important because colonoscopy provides location and appearance information that enables removal of polyps before colorectal cancer develops.
- Variable size, color, and texture within polyp types, together with blurred polyp–mucosa boundaries, make accurate segmentation challenging.
- Hand-crafted-feature methods often miss polyps because their representations poorly handle high intra-class variation and low interclass variation with hard mimics.
- PraNet first predicts coarse polyp areas and then implicitly models boundaries through reverse attention, targeting better learning, generalization, and training efficiency.
- The architecture uses a parallel partial decoder and three reverse attention modules to support accurate polyp segmentation.
2 Method
PraNet aggregates high-level features into a global guidance map, then progressively refines predictions with reverse attention and deep supervision. Its weighted losses emphasize difficult pixels while supervising global and side-output maps.
- Feature Aggregating via Parallel Partial Decoder: PraNet uses a parallel partial decoder to aggregate high-level features and generate a semantic global map for subsequent segmentation steps.
- Reverse Attention Module: Reverse attention progressively erases estimated foreground regions to mine discriminative boundaries and refine coarse predictions into accurate, complete maps.
- Reverse Attention Module: The reverse attention module multiplies high-level side-output features by reverse attention weights to produce reverse attention features.
- Learning Process and Implementation Details: The weighted IoU and binary cross-entropy losses impose global and pixel-level restrictions, giving greater attention to hard pixels.
- Learning Process and Implementation Details: Deep supervision up-samples the three side-outputs and global map to the ground-truth size, while end-to-end training uses 352×352 inputs and Adam optimization.
3 Experiments
Experiments evaluate PraNet against established baselines across five datasets using multiple segmentation metrics, including seen-dataset learning, unseen-dataset generalization, qualitative accuracy, efficiency, and component ablations.
- Datasets and Baselines: Experiments compare PraNet with U-Net, U-Net++, ResUNet-mod, ResUNet++, and SFA across five polyp segmentation datasets.The datasets include ETIS, CVC-ClinicDB/CVC-612, CVC-ColonDB, EndoScene, and Kvasir.
- Learning Ability: PraNet outperforms all compared methods on Kvasir and CVC-612, with mean Dice improvements of about > 7% across both datasets and all metrics.These experiments assess learning ability on two seen datasets.
- Generalization Capability: PraNet significantly improves over classical baselines and SFA on all three unseen datasets, while SFA drops dramatically on those datasets.The unseen datasets are CVC-ColonDB, ETIS, and the CVC300 test set of EndoScene.
- Qualitative Results: Qualitative results show that PraNet precisely locates and segments polyps in challenging cases involving varied size, homogeneous regions, and different textures.These examples are reported on the Kvasir test set.
- Training and Inference Analysis: PraNet converges in 20 epochs, approximately 0.5 hours, and runs at approximately 50fps for 352×352 inputs.The reported timings were measured on an Intel i9-9820X CPU and TITAN RTX GPU; the authors state that this speed supports colonoscopy-video implementation.
- Ablation Study: Ablations show that PPD improves the backbone, RA raises CVC-612 mean Dice from 0.747 to 0.888, and the combined PraNet setting is generally best.RA also increases Sα from 0.735 to 0.912; PPD is deployed only on high-level features to reduce training time.
4 Conclusion
PraNet consistently outperforms state-of-the-art approaches across five challenging datasets while achieving high accuracy without preprocessing or postprocessing. It also combines strong generalization with real-time segmentation efficiency.
- >5% improvement over all state-of-the-art approaches across five challenging datasets demonstrates PraNet’s consistent segmentation advantage.
- PraNet achieves a mean Dice of 0.898 on the Kvasir dataset without any pre-/postprocessing.
- PraNet is reported to provide strong learning ability, generalization ability, and real-time segmentation efficiency compared with current top-ranked SFA models.
- The architecture is universal and flexible, allowing additional effective modules to be added for further accuracy improvements.