ICGNet: Integration Context-based Reverse-Contour Guidance Network for Polyp Segmentation

Xiuquan Du, Xuebin Xu, Kunpeng MaIJCAI 2022 · DOI

TL;DR

This paper introduces ICGNet, a novel deep learning architecture designed to improve polyp segmentation in colonoscopic images. By integrating reverse-contour guidance, adaptive local-global context, and multi-level feature fusion, ICGNet tackles the challenges of blurry boundaries and varying polyp shapes, leading to more accurate and reliable polyp detection for early colorectal cancer diagnosis.

What Problem Does This Paper Address?

Colorectal cancer (CRC) is a significant global health concern, and polyps are its primary precursor. Early detection and removal of polyps are crucial for preventing CRC. Colonoscopy is the gold standard for polyp identification, but it's prone to missing polyps and requires skilled human labor. Automatic polyp segmentation using deep learning offers a promising solution, but it's hindered by several challenges:

  • Blurry Boundaries: The transition between polyps and the background can be indistinct, making it hard for algorithms to precisely delineate the polyp's edge.
  • Varied Shapes and Sizes: Polyps exhibit a wide range of shapes and sizes, which complicates feature representation and makes it difficult for models to generalize across different polyp appearances.

Existing methods often struggle with these issues, either by relying too heavily on contour or reverse information (leading to noise or missed detections) or by failing to effectively capture both local details and global context across different scales.

Methods

ICGNet is built upon a U-Net-like encoder-decoder architecture and introduces three key novel modules:

  • Reverse-Contour Guidance Module (RCG): This module aims to address the blurry boundary problem. It combines low-level edge features (which contain rich boundary details but are often obscured by noise) with reverse information (derived from previous predictions) to highlight target edges and suppress false-negative regions. This integrated approach helps compensate for lost boundary details and focuses attention on unpredicted areas within the polyp's potential boundary.
  • Adaptive Local-Global Context Module (ALGM): To handle the diverse shapes and sizes of polyps, ALGM adaptively extracts both local and global contextual information. It uses a non-local operation to capture long-range spatial dependencies (global context) and a series of cascaded dilated convolutions with increasing dilation rates to create larger and denser receptive fields (local context). This module also integrates complementary information from previous layers to enhance feature representation.
  • Hybrid Pyramid Pooling Fusion Module (HPPF): This module is designed to refine the segmentation results by effectively fusing multi-level features from the decoder. It employs a combination of global average pooling (GAP) to capture overall target information and global max pooling (GMP) with different kernel sizes to focus on prominent polyp areas and reduce background influence. This hybrid approach aims to produce a more accurate and refined segmentation mask.

Datasets

The proposed ICGNet was evaluated on three publicly available polyp segmentation datasets:

  • EndoScene: Contains 44 colonoscopy sequences from 36 patients.
  • Kvasir-SEG: A dataset of 1000 images with polyp regions, manually annotated by an experienced doctor.
  • CVC-ColonDB: Consists of 380 annotated frames from 15 colonoscopy sequences.

The paper does not mention the introduction of any new benchmarks.

Key Insights and Ideas

  • Synergy of Contour and Reverse Information: The RCG module demonstrates that combining edge information with reverse attention, rather than using them in isolation, is crucial for robust boundary segmentation and reducing false positives.
  • Adaptive Context is Key: The ALGM highlights the importance of adaptively capturing both local details and global context to handle the significant variations in polyp size and shape, a common challenge in medical image segmentation.
  • Beyond Simple Pooling: The HPPF module shows that a more sophisticated fusion strategy, incorporating both global average and local maximum pooling, can lead to better feature representation and more refined segmentation masks by focusing on salient polyp regions.
  • Lightweight Yet Effective: The ablation studies indicate that the proposed modules (RCG, ALGM, HPPF) are relatively lightweight and can significantly improve performance with minimal increases in model parameters.

Key Results

ICGNet achieved state-of-the-art performance on the Kvasir-SEG and EndoScene datasets, outperforming several advanced methods.

  • On the Kvasir-SEG dataset, ICGNet achieved a Dice score of 92.35% and a mean IoU of 91.99%.
  • On the EndoScene dataset, it achieved a Dice score of 87.93% and a mean IoU of 89.56%.
  • The model also demonstrated strong generalization capabilities on the CVC-ColonDB dataset, achieving a Dice score of 74.97% and a mean IoU of 82.64%.
  • Qualitative results showed that ICGNet was better at handling blurry boundaries and complex polyp shapes compared to other methods, with fewer instances of missed detections or inaccurate segmentations.

Limitations and Future Work

  • Limitations: While the paper highlights the effectiveness of its modules, it acknowledges that the algorithm is "still flawed." Specific details on these flaws are not elaborated upon, but the continuous improvement in metrics suggests ongoing refinement is possible. The paper also notes that using only edge information or reverse attention alone can lead to incorrect detections or noise, implying that the balance and integration of these features are critical and might still have room for optimization.
  • Future Work: The paper implicitly suggests future work could involve further refining the integration of the proposed modules and exploring their application to other medical imaging segmentation tasks. The success in generalization also points towards potential for broader application in clinical settings. The authors' focus on improving accuracy and reducing missed detections suggests that future work might aim for even higher precision and robustness in challenging scenarios.

Summary updated

BibTeX
@inproceedings{du2022icgnet,
  author = {Xiuquan Du and Xuebin Xu and Kunpeng Ma},
  title = {{ICGNet: Integration Context-based Reverse-Contour Guidance Network for Polyp Segmentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/123},
}
Figure