Abstract:Automatic and precise polyp segmentation in colonoscopy images is highly valuable for diagnosis at an early stage and surgery of colorectal cancer. Nevertheless, it still posed a major challenge due to variations in the size and intricate morphological characteristics of polyps coupled with the indistinct demarcation between polyps and mucosas. To alleviate these challenges, we proposed an improved dual-aggregation polyp segmentation network, dubbed Dua-PSNet, for automatic and accurate full-size polyp prediction by combining both the transformer branch and a fully convolutional network (FCN) branch in a parallel style. Concretely, in the transformer branch, we adopted the B3 variant of pyramid vision transformer v2 (PVTv2-B3) as an image encoder for capturing multi-scale global features and modeling long-distant interdependencies between them whilst designing an innovative multi-stage feature aggregation decoder (MFAD) to highlight critical local feature details and effectively integrate them into global features. In the decoder, the adaptive feature aggregation (AFA) block was constructed for fusing high-level feature representations of different scales generated by the PVTv2-B3 encoder in a stepwise adaptive manner for refining global semantic information, while the ResidualBlock module was devised to mine detailed boundary cues disguised in low-level features. With the assistance of the selective global-to-local fusion head (SGLFH) module, the resulting boundary details were aggregated selectively with these global semantic features, strengthening these hierarchical features to cope with scale variations of polyps. The FCN branch embedded in the designed ResidualBlock module was used to encourage extraction of highly merged fine features to match the outputs of the Transformer branch into full-size segmentation maps. In this way, both branches were reciprocally influenced and complemented to enhance the discrimination capability of polyp features and enable a more accurate prediction of a full-size segmentation map. Extensive experiments on five challenging polyp segmentation benchmarks demonstrated that the proposed Dua-PSNet owned powerful learning and generalization ability and advanced the state-of-the-art segmentation performance among existing cutting-edge methods. These excellent results showed our Dua-PSNet had great potential to be a promising solution for practical polyp segmentation tasks in which wide variations of data typically occurred.

A novel non-pretrained deep supervision network for polyp segmentation

PDCA-Net: Parallel dual-channel attention network for polyp segmentation

PraNet: Parallel Reverse Attention Network for Polyp Segmentation

Shallow Attention Network for Polyp Segmentation

PPNet: Pyramid pooling based network for polyp segmentation

DAN-PD: Domain adaptive network with parallel decoder for polyp segmentation

DLGRAFE-Net: A double loss guided residual attention and feature enhancement network for polyp segmentation

Dual encoder–decoder-based deep polyp segmentation network for colonoscopy images

Attention based multi-scale parallel network for polyp segmentation

Probabilistic Modeling Ensemble Vision Transformer Improves Complex Polyp Segmentation

Improved dual-aggregation polyp segmentation network combining a pyramid vision transformer with a fully convolutional network

DBG-Net: A Double-Branch Boundary Guidance Network for Polyp Segmentation

Dual-branch multi-information aggregation network with transformer and convolution for polyp segmentation

TCCNet: Temporally Consistent Context-Free Network for Semi-supervised Video Polyp Segmentation

Polyp segmentation with distraction separation

PDLFBR-Net: Partial Decoder Localization and Foreground-Background Refinement Network for Polyp Segmentation

Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation.

Learnable Oriented-Derivative Network for Polyp Segmentation

SSN: A Stair-Shape Network for Real-Time Polyp Segmentation in Colonoscopy Images.

DSFNet: Dual-GCN and Location-fused Self-attention with Weighted Fast Normalized Fusion for Polyps Segmentation

DBE-Net: Dual Boundary-Guided Attention Exploration Network for Polyp Segmentation