Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Contextual Attention Network: Transforming Spatial Representations into Adaptive Insights for Endoscopic Polyp Diagnosis

Teja Krishna Cherukuri, Nagur Shareef Shaik, Sribhuvan Reddy Yellu, Jun-Won Chung, Dong Hye Ye

TL;DR

The paper tackles the challenge of explainable endoscopic polyp diagnosis by addressing localization and contextual awareness in imaging. It introduces the Dynamic Contextual Attention Network (DCAN), a multi-stage attention framework that converts spatial feature maps into adaptive contextual insights using Spatial Attention, Gating, and Adaptive Attention Refinement, built on EfficientNetV2B0 features and a Regularized Classification Head. Empirical results on GMC Polyp and HyperKvasir demonstrate that DCAN delivers higher accuracy and robust explainability, with compelling GradCAM++ visualizations and ablation evidence showing the added value of each component. By enabling reliable, interpretable polyp detection without explicit localization modules, DCAN holds promise for improving colorectal cancer screening and clinical decision-making, with future work exploring multi-modal integration with text data for automated report generation.

Abstract

Colorectal polyps are key indicators for early detection of colorectal cancer. However, traditional endoscopic imaging often struggles with accurate polyp localization and lacks comprehensive contextual awareness, which can limit the explainability of diagnoses. To address these issues, we propose the Dynamic Contextual Attention Network (DCAN). This novel approach transforms spatial representations into adaptive contextual insights, using an attention mechanism that enhances focus on critical polyp regions without explicit localization modules. By integrating contextual awareness into the classification process, DCAN improves decision interpretability and overall diagnostic performance. This advancement in imaging could lead to more reliable colorectal cancer detection, enabling better patient outcomes.

Dynamic Contextual Attention Network: Transforming Spatial Representations into Adaptive Insights for Endoscopic Polyp Diagnosis

TL;DR

The paper tackles the challenge of explainable endoscopic polyp diagnosis by addressing localization and contextual awareness in imaging. It introduces the Dynamic Contextual Attention Network (DCAN), a multi-stage attention framework that converts spatial feature maps into adaptive contextual insights using Spatial Attention, Gating, and Adaptive Attention Refinement, built on EfficientNetV2B0 features and a Regularized Classification Head. Empirical results on GMC Polyp and HyperKvasir demonstrate that DCAN delivers higher accuracy and robust explainability, with compelling GradCAM++ visualizations and ablation evidence showing the added value of each component. By enabling reliable, interpretable polyp detection without explicit localization modules, DCAN holds promise for improving colorectal cancer screening and clinical decision-making, with future work exploring multi-modal integration with text data for automated report generation.

Abstract

Colorectal polyps are key indicators for early detection of colorectal cancer. However, traditional endoscopic imaging often struggles with accurate polyp localization and lacks comprehensive contextual awareness, which can limit the explainability of diagnoses. To address these issues, we propose the Dynamic Contextual Attention Network (DCAN). This novel approach transforms spatial representations into adaptive contextual insights, using an attention mechanism that enhances focus on critical polyp regions without explicit localization modules. By integrating contextual awareness into the classification process, DCAN improves decision interpretability and overall diagnostic performance. This advancement in imaging could lead to more reliable colorectal cancer detection, enabling better patient outcomes.
Paper Structure (15 sections, 9 equations, 3 figures, 3 tables)

This paper contains 15 sections, 9 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Representation Learning
  4. Dynamic Contextual Attention
  5. Spatial Attention
  6. Gating
  7. Adaptive Attention Refinement
  8. Regularized Classification Head
  9. Experimental Results
  10. Datasets
  11. Implementation Details
  12. Quantitative Evaluation
  13. Ablation Study
  14. Qualitative Results
  15. Conclusion

Figures (3)

  • Figure 1: Architecture of the Proposed Dynamic Contextual Attention Network (DCAN). Pre-processing applies CLAHE to enhance contrast. EfficientNet$V_2B_0$ extracts spatial features, while the Dynamic Contextual Attention mechanism boosts diagnosis accuracy and explainability by focusing on relevant anatomical regions and filtering out noise. The Regularized Classification Head categorizes images as Normal or Abnormal, and GradCAM++ visually explains the DCAN's decisions.
  • Figure 2: Attention maps generated by different attention mechanisms highlight crucial regions for polyp detection in endoscopic images from both the In-house (top) and HyperKvasir (bottom) datasets. The figure presents both non-CLAHE and CLAHE-enhanced images, showcasing their impact on the Dynamic Contextual Attention (DCA) model. A red-colored bounding box precisely localizes the polyp within the original endoscopic image.
  • Figure 3: Attention maps for polyp detection in abnormal class images post-CLAHE; The maps highlight the contribution of the Spatial and Gated components in conjunction with the Refinement component, demonstrating the importance of covering the entire polyp.