Source-linked AI summary
Adaptive Context Selection for Polyp Segmentation
Ruifei Zhang, Guanbin Li, Zhen Li, Shuguang Cui, Dahong Qian, Yizhou Yu
TL;DR
Polyp segmentation is important for colorectal cancer diagnosis and treatment but remains difficult because methods rarely account explicitly for varied polyp size, shape, and spatial context. ACSNet addresses this with local hard-region attention, global context extraction, and adaptive channel-wise context selection in an encoder-decoder framework. The method achieves the best reported performance on the evaluated EndoScene and Kvasir-SEG comparisons and shows progressive gains in ablation experiments.
Problem
Existing polyp-segmentation algorithms rarely explicitly account for polyp size, shape, and complex spatial context, although different polyp sizes rely on different local and global information.
Method
ACSNet combines LCA for hard-region local context, GCM for global context, and ASM for channel-wise adaptive context selection within an encoder-decoder framework.
Results
ACSNet achieves the best performance over all reported metrics on EndoScene and outperforms other compared methods by large margins on Kvasir-SEG.
Takeaways & Limitations
Adaptive selection of local and global context supports more consistent and accurate polyp segmentation across varied polyp sizes, shapes, and complex samples.
Abstract
from arXiv · showhide
Accurate polyp segmentation is of great significance for the diagnosis and treatment of colorectal cancer. However, it has always been very challenging due to the diverse shape and size of polyp. In recent years, state-of-the-art methods have achieved significant breakthroughs in this task with the help of deep convolutional neural networks. However, few algorithms explicitly consider the impact of the size and shape of the polyp and the complex spatial context on the segmentation performance, which results in the algorithms still being powerless for complex samples. In fact, segmentation of polyps of different sizes relies on different local and global contextual information for regional contrast reasoning. To tackle these issues, we propose an adaptive context selection based encoder-decoder framework which is composed of Local Context Attention (LCA) module, Global Context Module (GCM) and Adaptive Selection Module (ASM). Specifically, LCA modules deliver local context features from encoder layers to decoder layers, enhancing the attention to the hard region which is determined by the prediction map of previous layer. GCM aims to further explore the global context features and send to the decoder layers. ASM is used for adaptive selection and aggregation of context features through channel-wise attention. Our proposed approach is evaluated on the EndoScene and Kvasir-SEG Datasets, and shows outstanding performance compared with other state-of-the-art methods. The code is available at https://github.com/ReaFly/ACSNet.
1 Introduction
Polyp segmentation is clinically important but difficult because polyps vary in shape and size, while existing approaches rarely adapt explicitly to these differences and complex spatial context. ACSNet addresses this gap by selecting local and global context adaptively within an encoder-decoder framework.
- Automatic and accurate polyp segmentation supports colorectal cancer diagnosis and treatment, but manual colonoscopy is expensive and has a higher misdiagnosis rate.Colorectal cancer is described as a serious health threat, and polyps are critical precursors.
- Polyp segmentation remains challenging because polyps exhibit diverse shapes and sizes.
- Earlier fully convolutional networks produced rough, low-resolution segmentations, while UNet-based structures progressively restore resolution.The passage describes this as a progression in receptive-field enlargement and boundary recovery.
- ACSNet adaptively selects context information according to polyp-region size, combining LCA, GCM, and ASM within an encoder-decoder architecture.LCA and GCM mine local and global context, while ASM performs channel-wise feature fusion and selection.
2 Method
ACSNet enhances an encoder-decoder network with local hard-region attention, global context extraction, and channel-wise adaptive context selection. These modules route complementary context through decoder stages to refine polyp predictions.
- Architecture: ACSNet uses a five-block ResNet34 encoder and five decoder blocks, each producing a supervised prediction map at a different resolution.Each decoder block contains two Conv-BN-ReLU combinations.
- Local Context Attention Module (LCA): LCA replaces encoder-decoder skip connections and uses the previous decoder prediction to emphasize uncertain, harder regions with local context enhancement.The attention map is spatially defined and supports layer-wise feature complementation and prediction refinement.
- Local Context Attention Module (LCA): LCA attention increases as prediction values approach threshold T, with T set to 0.5 in the experiments.Features are multiplied by attention values and summed with the original features.
- Global Context Module (GCM): GCM extracts multi-scale global context through 1 × 1, 3 × 3, and 5 × 5 pooling outputs plus a non-local identity branch, then feeds concatenated features to decoder ASMs.The four feature maps are upsampled and concatenated before dense delivery to decoder stages.
- Adaptive Selection Module (ASM): ASM aggregates LCA features, GCM features, and preceding decoder features, using channel-wise attention to select informative context and suppress unhelpful responses.It incorporates a Squeeze-and-Excitation block and sigmoid-normalized channel weights.
3 Experiments
Experiments evaluate ACSNet on EndoScene and Kvasir-SEG using segmentation metrics, comparisons with state-of-the-art methods, visualizations, and module ablations. ACSNet achieves strong benchmark performance, with improvements attributed to context selection and hard-region mining.
- Datasets: ACSNet is evaluated on EndoScene and Kvasir-SEG, with fixed image sizes of 384×288 and 320 × 320, respectively.EndoScene contains 912 images; Kvasir-SEG contains 1000 images and uses a 60%/20%/20% train-validation-test split.
- Implementation Details and Evaluation Metrics: Evaluation uses Recall, Specificity, Precision, Dice Score, IoUp, IoUb, mIoU, and Accuracy.Training uses augmentation, random 224 × 224 crops, SGD, and a binary cross-entropy plus Dice loss.
- Results on the EndoScene Dataset: 86.59% Dice on EndoScene gives ACSNet a 3.66% improvement over the second-best algorithm and the best performance across all reported metrics.The comparison includes FCN8s, UNet, UNet++, SegNet, and SFANet.
- Results on the EndoScene Dataset: Visual comparisons show ACSNet predictions closest to ground truth, with hard regions commonly located near polyp borders.The reported robustness includes variation in polyp region size and image brightness, while hard-region mining supports boundary positioning during prediction refinement.
- Results on the Kvasir-SEG Dataset: On Kvasir-SEG, ACSNet achieves the best performance and outperforms UNet, UNet++, SegNet, ResUNet, and SFANet by large margins.The authors describe these results as demonstrating robustness and effectiveness.
- Ablation study: Adding LCAs, GCM, and ASMs progressively improves Dice by 0.79%, 1.28%, and 1.02%, respectively, over successive ablation variants.The baseline is a U-shape encoder-decoder with skip-connections; modules are added as Baseline+LCAs, Baseline+LCAs+GCM, and Ours.
4 Conclusion
The paper concludes that efficient perception of local and global context is essential for polyp localization and segmentation. ACSNet combines hard-region local context extraction, global feature enhancement, and adaptive context aggregation, with experiments and ablations supporting its effectiveness.
- 4 Conclusion: ACSNet uses LCA for hard-region-based local context extraction, GCM for global feature representation and enhancement, and ASM for contextual information aggregation and selection.The framework is presented as an adaptive context selection encoder-decoder approach.
- 4 Conclusion: Extensive experiments and ablation studies demonstrate the effectiveness and superiority of the proposed method.