Source-linked AI summary
Lesion-aware Dynamic Kernel for Polyp Segmentation
Ruifei Zhang, Peiwen Lai, Xiang Wan, De-Jun Fan, Feng Gao, Xiao-Jian Wu, Guanbin Li
TL;DR
Polyp segmentation is challenging because polyps vary in appearance and concealed polyps can have low contrast against surrounding regions. LDNet uses a dynamic segmentation head and tailored attention modules, and the authors report superior performance across public benchmarks and a collected dataset.
Problem
Polyp segmentation is challenging because diverse polyp appearances and low contrast with surrounding regions make concealed polyps difficult to segment.
Source: §1 · p. 1Method
Results
LDNet outperforms other methods across four public benchmarks on most metrics and achieves 78.43% and 74.37% Dice scores on the two unseen datasets, improving by 2.19% and 1.99%.
Source: §3.3 · pp. 6, 8Takeaways & Limitations
Experiments and ablation studies support the effectiveness of LDNet for polyp segmentation.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Automatic and accurate polyp segmentation plays an essential role in early colorectal cancer diagnosis. However, it has always been a challenging task due to 1) the diverse shape, size, brightness and other appearance characteristics of polyps, 2) the tiny contrast between concealed polyps and their surrounding regions. To address these problems, we propose a lesion-aware dynamic network (LDNet) for polyp segmentation, which is a traditional u-shape encoder-decoder structure incorporated with a dynamic kernel generation and updating scheme. Specifically, the designed segmentation head is conditioned on the global context features of the input image and iteratively updated by the extracted lesion features according to polyp segmentation predictions. This simple but effective scheme endows our model with powerful segmentation performance and generalization capability. Besides, we utilize the extracted lesion representation to enhance the feature contrast between the polyp and background regions by a tailored lesion-aware cross-attention module (LCA), and design an efficient self-attention module (ESA) to capture long-range context relations, further improving the segmentation accuracy. Extensive experiments on four public polyp benchmarks and our collected large-scale polyp dataset demonstrate the superior performance of our method compared with other state-of-the-art approaches. The source code is available at https://github.com/ReaFly/LDNet.
1 Introduction
Polyp segmentation is difficult because polyps vary in appearance and can have low contrast against surrounding regions. LDNet addresses this with an input-conditioned dynamic segmentation head, tailored attention modules, and experiments across public and collected benchmarks.
1 Introduction
- Appearance diversity and low contrast make accurate polyp segmentation challenging, while prior methods had not addressed flexibility in feature learning through network structure.The introduction links accurate automatic segmentation to early polyp detection and removal, which can reduce colorectal cancer incidence.Source: §1 · pp. 1–2
- LDNet generates its segmentation head from global image features and iteratively updates it using lesion features refined from segmentation predictions.The authors propose this dynamic-kernel scheme to adapt feature learning to polyps with diverse shapes, sizes, and appearances.Source: §1 · pp. 2–3
- ESA captures global feature relations, while LCA enhances contrast between lesions and background regions to help detect concealed polyps.The authors present both attention modules as additions to the dynamic-kernel network.Source: §1 · pp. 2–3
- Experiments on four public benchmarks and a collected large-scale dataset are reported as demonstrating LDNet's effectiveness.The contribution summary identifies all five evaluation datasets as evidence for the proposed method.Source: §1 · p. 3
2 Methodology
LDNet combines an encoder–decoder with a segmentation head generated from global image features and iteratively updated using lesion features. ESA captures long-range context, while LCA uses lesion representations to enhance polyp–background feature contrast.
2 Methodology
- LDNet pairs a five-block Res2Net encoder with five decoder stages and an input-conditioned head updated from decoder features and prior predictions; final-stage P1 is the output.Each stage's prediction is supervised by its corresponding down-sampled ground truth.Source: §2 · pp. 3–4
Kernel Generation
- Pooled global context features generate the initial 64-channel kernel, which is applied to the deepest decoder feature to produce the first polyp prediction.Adaptive average pooling produces a K × K feature map before a 1 × 1 convolution generates the kernel.Source: Kernel Generation · p. 4
Kernel Update
- The prior prediction gates decoder features into lesion representations, which are gated and integrated with the previous kernel to make the next prediction iteratively.The gate mechanism is used to filter noise when a prior polyp prediction may be inaccurate.Source: Kernel Update · p. 4
Efficient Self-Attention
- ESA uses multi-head attention with pooled, multiscale global features as keys and values, adding long-range relations while reducing computation.The pooling scales are 1 × 1, 3 × 3, and 5 × 5; the per-head dimension is C_i.Source: Efficient Self-Attention · p. 5
Lesion-aware Cross-Attention
- LCA uses lesion representations extracted from decoder features and predictions as keys and values, enhancing lesion characteristics to increase polyp–background feature contrast.Regions more similar to the lesion receive greater enhancement, supporting detection of concealed polyps.Source: Lesion-aware Cross-Attention · p. 6
3 Experiments
LDNet is evaluated on four public benchmarks and a 5,175-image collected dataset. Comparisons and ablations report strong performance across these settings.
3.1 Datasets
- The public evaluation trains on 80% of Kvasir-SEG and CVC-ClinicDB, validates on 10%, and tests on their remaining images plus CVC-ColonDB and ETIS; the collected dataset has 5,175 images split 60/20/20.Source: §3.1 · p. 6
3.2 Implementation Details and Evaluation Metrics
- Training uses 256 × 256 images, SGD, and combined Binary Cross-Entropy and Dice losses; performance is assessed with eight metrics, including Dice, IoUp, IoUb, mIoU, and Accuracy.Source: §3.2 · p. 6
3.3 Experiments on the Public Polyp Benchmarks
- On four public benchmarks, LDNet leads on most metrics; unseen datasets yield Dice 78.43% and 74.37%, gains of 2.19% and 1.99% over second-best approaches.On Kvasir and CVC-ClinicDB, it records the best Dice and mIoU scores.Source: §3.3 · pp. 6, 8, Table 1 · p. 7
3.4 Experiments on the Collected Large-Scale Polyp Dataset
- On the collected dataset, LDNet achieves the best performance: Dice 91.66% and mIoU 91.34%.Source: Table 2 · p. 7, §3.4 · p. 8
3.5 Ablation Study
- Adding the dynamic kernel improves Dice by 1.58%; ESA and LCA additions further boost Dice and mIoU by 0.86%/0.68% and 0.92%/0.96%, respectively.The study progresses from a static-head baseline to Baseline+DK, Baseline+DK+ESAs, and Ours.Source: Table 2 · p. 7, §3.5 · p. 8
4 Conclusion
LDNet combines a globally conditioned, iteratively updated lesion-aware dynamic kernel with ESA and LCA. The paper reports that experiments and ablations demonstrate the method's effectiveness.
- LDNet's segmentation kernel is generated from global information and updated using multi-level lesion features, giving the model flexibility to attend to diverse polyp regions.Source: §4 · pp. 8–9
- ESA and LCA improve feature representation and context contrast, which the paper says is beneficial for detecting concealed polyps.Source: §4 · p. 9
- The paper reports that extensive experiments and ablation studies demonstrate the proposed method's effectiveness.Source: §4 · p. 9