Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DTEA: Dynamic Topology Weaving and Instability-Driven Entropic Attenuation for Medical Image Segmentation

Weixuan Li, Quanjun Li, Guang Yu, Song Yang, Zimeng Li, Chi-Man Pun, Yupeng Liu, Xuhang Chen

TL;DR

The paper addresses the semantic gap in medical image segmentation by introducing DTEA, a skip-connection framework that combines Semantic Topology Reconfiguration (STR) with Entropic Perturbation Gating (EPG). STR builds a dynamic hypergraph over multi-scale features to model high-order cross-resolution dependencies, while EPG uses chaotic perturbation and entropy-based channel gating to suppress high-entropy, noisy channels and sharpen spatial attention. Across Synapse, ISIC 2018, and CVC-ClinicDB, DTEA achieves state-of-the-art Dice scores and robust generalization across backbones, demonstrating superior cross-task performance and resilience to clinical variability. The approach is architecture-agnostic and can be integrated with both CNN and Transformer backbones to enhance multi-scale feature fusion in challenging clinical settings.

Abstract

In medical image segmentation, skip connections are used to merge global context and reduce the semantic gap between encoder and decoder. Current methods often struggle with limited structural representation and insufficient contextual modeling, affecting generalization in complex clinical scenarios. We propose the DTEA model, featuring a new skip connection framework with the Semantic Topology Reconfiguration (STR) and Entropic Perturbation Gating (EPG) modules. STR reorganizes multi-scale semantic features into a dynamic hypergraph to better model cross-resolution anatomical dependencies, enhancing structural and semantic representation. EPG assesses channel stability after perturbation and filters high-entropy channels to emphasize clinically important regions and improve spatial attention. Extensive experiments on three benchmark datasets show our framework achieves superior segmentation accuracy and better generalization across various clinical settings. The code is available at \href{https://github.com/LWX-Research/DTEA}{https://github.com/LWX-Research/DTEA}.

DTEA: Dynamic Topology Weaving and Instability-Driven Entropic Attenuation for Medical Image Segmentation

TL;DR

The paper addresses the semantic gap in medical image segmentation by introducing DTEA, a skip-connection framework that combines Semantic Topology Reconfiguration (STR) with Entropic Perturbation Gating (EPG). STR builds a dynamic hypergraph over multi-scale features to model high-order cross-resolution dependencies, while EPG uses chaotic perturbation and entropy-based channel gating to suppress high-entropy, noisy channels and sharpen spatial attention. Across Synapse, ISIC 2018, and CVC-ClinicDB, DTEA achieves state-of-the-art Dice scores and robust generalization across backbones, demonstrating superior cross-task performance and resilience to clinical variability. The approach is architecture-agnostic and can be integrated with both CNN and Transformer backbones to enhance multi-scale feature fusion in challenging clinical settings.

Abstract

In medical image segmentation, skip connections are used to merge global context and reduce the semantic gap between encoder and decoder. Current methods often struggle with limited structural representation and insufficient contextual modeling, affecting generalization in complex clinical scenarios. We propose the DTEA model, featuring a new skip connection framework with the Semantic Topology Reconfiguration (STR) and Entropic Perturbation Gating (EPG) modules. STR reorganizes multi-scale semantic features into a dynamic hypergraph to better model cross-resolution anatomical dependencies, enhancing structural and semantic representation. EPG assesses channel stability after perturbation and filters high-entropy channels to emphasize clinically important regions and improve spatial attention. Extensive experiments on three benchmark datasets show our framework achieves superior segmentation accuracy and better generalization across various clinical settings. The code is available at \href{https://github.com/LWX-Research/DTEA}{https://github.com/LWX-Research/DTEA}.

Paper Structure

This paper contains 15 sections, 8 equations, 4 figures, 5 tables.

Table of Contents

  1. INTRODUCTION
  2. METHOD
  3. Feature Preprocessing
  4. Semantic Topology Reconfiguration
  5. Entropic Perturbation Gating
  6. Feature Postprocessing
  7. EXPERIMENT RESULTS
  8. Dataset
  9. Implementation Details
  10. Comparison with State-of-the-art Models
  11. Ablation Study
  12. Effectiveness analysis of STR and EPG
  13. $K$-value in EPG
  14. Ablation Study on Backbone
  15. Conclusion

Figures (4)

  • Figure 1: Hypergraph visualization of DTEA. Three patches (Yellow, Blue, and Green) are selected as central nodes to visualize the corresponding hyperedges generated by the STR module. The lesion and non-lesion areas exhibit a clear separation in the hypergraph. Within the lesion region, the hyperedges show strong aggregation, while nodes in the boundary region also display notable similarity and structural correlation.
  • Figure 2: (a) The overall architecture of the proposed DTEA. (b) The skip connection framework. (c) Semantic Topology Reconfiguration (STR). (d) Entropy Perturbation Gating (EPG).
  • Figure 3: Visual comparison of low-entropy and high-entropy channels in EPG. (a) Input image. (b) Channels Feature maps. (c) The entropy maps of the feature maps. (d) The entropy maps of the feature maps after chaotic perturbation.
  • Figure 4: Visual Comparison for other model and proposed DTEA. (a) Input image. (b) Ground Truth. (c) Unet. (d) TransUNet. (e) M$^2$SNet. (f) MADGNet. (g) CFATransUnet. (h) DTEA.