[Paper Review] PraNet: Parallel Reverse Attention Network for Polyp Segmentation

Deng-Ping Fan, Ge-Peng Ji+5 more|arXiv (Cornell University)|Jun 13, 2020

Advanced Image and Video Retrieval Techniques38 references109 citations

TL;DR

PraNet introduces a Parallel Partial Decoder and Reverse Attention modules to achieve real-time, high-accuracy polyp segmentation in colonoscopy images, improving generalization across datasets.

ABSTRACT

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.

Motivation & Objective

• Motivate automatic, accurate polyp segmentation to aid colorectal cancer screening.
• Address large intra-polyp appearance variance and blurred polyp-boundary boundaries.
• Develop a fast, generalizable network suitable for real-time colonoscopy video.
• Leverage area and boundary cues through novel architectural components to improve precision.

Proposed method

• Use a parallel partial decoder (PPD) to aggregate high-level features and produce a global polyp map.
• Employ multiple reverse attention (RA) modules to progressively mine boundary cues by erasing predicted polyp regions and refining predictions.
• Train with a joint loss consisting of weighted IoU and weighted BCE applied to global and side-output maps.
• Adopt deep supervision on multiple outputs to stabilize training and enable end-to-end learning.
• Maintain real-time performance (~50 fps) on 352x352 inputs.

Experimental results

Research questions

• RQ1 Can a parallel decoding strategy improve high-level feature aggregation for polyp segmentation?
• RQ2 Do reverse attention mechanisms effectively refine boundaries by erasing previous predictions?
• RQ3 Is PraNet capable of real-time inference with superior accuracy across diverse polyp datasets?
• RQ4 How do PPD and RA interact to improve both learning speed and segmentation quality?

Key findings

• PraNet outperforms state-of-the-art methods across five datasets and multiple metrics.
• On Kvasir, PraNet achieves mean Dice of 0.898 and mean IoU of 0.840.
• On CVC-612, PraNet achieves mean Dice of 0.899 and mean IoU of 0.849.
• PraNet demonstrates strong generalization to unseen datasets, with significant gains over baselines.
• Inference speed is real-time at ~50 fps for 352x352 inputs, with training convergence in ~20 epochs (~0.5 hours).
• Ablation shows PPD and RA contribute additively to performance, with the full combination (PPD+RA+Backbone) yielding the best results.