Source-linked AI summary
ET-Net: A Generic Edge-aTtention Guidance Network for Medical Image Segmentation
Zhijie Zhang, Huazhu Fu, Hang Dai, Jianbing Shen, Yanwei Pang, Ling Shao
TL;DR
Medical segmentation methods often prioritize region extraction while overlooking edge information useful for accurate boundaries. ET-Net learns edge-attention features early, transfers them through multi-scale decoding, and outperforms prior methods across four segmentation tasks.
Problem
Existing medical segmentation methods often omit edge information, despite its relevance to accurate boundary segmentation.
Method
ET-Net learns edge-attention representations in early encoding layers and transfers them to multi-scale decoding layers through weighted aggregation.
Results
Across retinal, chest X-Ray, and CT segmentation tasks, ET-Net outperformed prior methods; in lung CT, it achieved 98.68% Acc. and 96.23% mIoU.
Takeaways & Limitations
The experiments support preserving edge-attention representations as a useful component for medical image segmentation accuracy.
Abstract
from arXiv · showhide
Segmentation is a fundamental task in medical image analysis. However, most existing methods focus on primary region extraction and ignore edge information, which is useful for obtaining accurate segmentation. In this paper, we propose a generic medical segmentation method, called Edge-aTtention guidance Network (ET-Net), which embeds edge-attention representations to guide the segmentation network. Specifically, an edge guidance module is utilized to learn the edge-attention representations in the early encoding layers, which are then transferred to the multi-scale decoding layers, fused using a weighted aggregation module. The experimental results on four segmentation tasks (i.e., optic disc/cup and vessel segmentation in retinal images, and lung segmentation in chest X-Ray and CT images) demonstrate that preserving edge-attention representations contributes to the final segmentation accuracy, and our proposed method outperforms current state-of-the-art segmentation methods. The source code of our method is available at https://github.com/ZzzJzzZ/ETNet.
1 Introduction
Medical image segmentation must reconcile edge detection’s local detail with object segmentation’s global semantic information, because each alone can produce incomplete or uncertain boundaries. ET-Net addresses this by learning edge-attention representations early and using them to guide segmentation through edge guidance and weighted aggregation modules.
- Motivation: Edge detection methods rely on local edge representations and lack object-level information, producing trivial regions and discontinuous boundaries, whereas object segmentation methods preserve foreground-background homogeneity and semantic characteristics.The introduction contrasts edge-based methods used for retinal blood-vessel extraction with object segmentation methods [18].
- Background: U-Net has enabled medical-image applications including cell detection, retinal vessel segmentation, and lung-region extraction from chest X-Ray and CT images.These applications illustrate the broad use of U-Net-based segmentation in medical imaging.
- Contribution: ET-Net integrates edge detection and object segmentation in one general medical image segmentation network by embedding edge-attention representations to guide segmentation.The method is designed as a general segmentation approach for medical images.
- Method: The edge guidance module learns and preserves local edge characteristics in early encoding layers, while the weighted aggregation module transfers and combines these representations for segmentation.The supplied architecture description identifies the encoder-decoder network, edge guidance module, and weighted aggregation module.
2 Method
ET-Net extends a ResNet-50 encoder–decoder with an edge guidance module and weighted aggregation module. Edge-attention features learned from early layers guide decoding and are integrated with multi-scale features for segmentation.
- Architecture: ET-Net uses a ResNet-50 encoder–decoder with four resolution-specific encoding blocks, augmented by EGM and WAM modules.Each encoding block combines stacked 1×1−3×3−1×1 convolutions with a shortcut connection.
- Edge Guidance Module: The EGM operates on early encoding layers, learns edge-attention representations, guides decoding, and supervises early convolutions with an edge-detection loss.It is applied to E-Blocks 1 and 2 because low-level features preserve sufficient edge information.
- Edge Guidance Module: The EGM upsamples and concatenates early features, then separately produces decoding guidance features and edge predictions using 1×1 convolutions.It uses Lovász-Softmax loss for edge supervision, helping transmitted edge features guide discriminative high-level feature extraction.
- Weighted Aggregation Module: WAM emphasizes valuable features while aggregating multi-scale outputs and EGM edge-attention representations under edge-guided conditions.A bottom-up pathway builds a hierarchy of feature-map sizes, followed by concatenation and 1×1 convolution; segmentation also uses Lovász-Softmax loss.
- Implementation: Training uses ImageNet-pretrained ResNet-50, synchronized batch normalization with batch size 16, poly learning-rate scheduling, and 512×512 random crops with geometric and color augmentation.Scale augmentation ranges from 0.5 to 2, rotation from −10 to 10 degrees, and color jitter is applied with probability 0.5.
3 Experiments
ET-Net is evaluated across retinal, X-Ray, and CT segmentation tasks against classic and state-of-the-art baselines using generic segmentation metrics. Qualitative results favor ET-Net particularly around object edges, while ablations show benefits from edge-guidance components.
- Experimental setup: The experiments span optic disc/cup and vessel segmentation in retinal images and lung segmentation in X-Ray and CT images.The evaluation uses metrics selected for generic segmentation, including Dice coefficients, mIoU, and Acc. scores.
- Quantitative and qualitative results: ET-Net is compared with FCN, U-Net, M-Net, Multi-task, and pOSAL for optic disc/cup segmentation.pOSAL is described as the state-of-the-art method that achieved first place in the REFUGE challenge.
- Quantitative and qualitative results: ET-Net produces segmentation results closer to ground truth and improves especially in edge regions compared with other methods.Qualitative comparisons cover optic disc/cup and vessel segmentation in retinal fundus images, plus lung segmentation in Chest X-Ray and CT images.
- Ablation study: Adding the edge guidance module raises Drishti-GS performance from 90.11%/95.77%/84.41% to 91.24%/97.17%/86.49% for DiceOC/DiceOD/mIoU.The improvement is reported with only a small addition to computational cost.
- Ablation study: The ablation study evaluates the proposed components on the Drishti-GS dataset using an encoder-decoder base network.The study separately examines the edge guidance module and weighted aggregation module under the stated training settings.
4 Conclusion
ET-Net is proposed as a general medical image segmentation network built on the premise that edge detection and region segmentation mutually benefit each other. Its Edge Guidance and Weighted Aggregation Modules generate edge-aware representations and combine them with emphasized high-level features.
- ET-Net is introduced as a novel network for general medical image segmentation.
- The Edge Guidance Module detects object edges and generates edge-attention representations containing sufficient edge information.
- The Weighted Aggregation Module highlights valuable high-level features and combines them with edge representations.