[Paper Review] ET-Net: A Generic Edge-aTtention Guidance Network for Medical Image Segmentation

Zhijie Zhang, Huazhu Fu+4 more|arXiv (Cornell University)|Jul 25, 2019

Retinal Imaging and AnalysisMedicine20 references21 citations

TL;DR

ET-Net is a generic edge-attention guidance network for medical image segmentation that enhances segmentation accuracy by integrating edge-attention representations through an Edge Guidance Module (EGM) in early encoding layers and a Weighted Aggregation Module (WAM) for multi-scale feature fusion in decoding paths. It achieves state-of-the-art performance across four medical imaging tasks, including retinal and pulmonary segmentation, with mIoU improvements up to 96.23% on CT lung segmentation.

ABSTRACT

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.

Motivation & Objective

• Address the limitation of existing medical segmentation models that overlook edge information, leading to inaccurate boundary predictions.
• Integrate edge detection and object segmentation into a unified deep learning framework to improve boundary localization and segmentation accuracy.
• Develop a generic, transferable architecture that enhances segmentation performance across diverse medical imaging tasks without task-specific retraining.
• Preserve and leverage local edge characteristics through attention mechanisms to guide high-level feature learning in encoder-decoder networks.
• Reduce reliance on post-processing by embedding edge-aware representations directly into the segmentation network architecture.

Proposed method

• Employ an Edge Guidance Module (EGM) in early encoding layers to learn edge-attention representations that capture fine-grained boundary details.
• Integrate the EGM with a ResNet-50-based encoder to extract hierarchical features while preserving edge-specific information.
• Use a Weighted Aggregation Module (WAM) in the decoder path to fuse multi-scale side-outputs and propagate edge-attention features to high-level layers.
• Apply skip connections and concatenation between encoder and decoder blocks to maintain spatial resolution and feature consistency.
• Train the network end-to-end using standard segmentation loss functions, with edge-attention features guiding the final segmentation prediction.
• Utilize residual connections in E-Blocks and D-Blocks to stabilize training and improve gradient flow in deep architectures.

Experimental results

Research questions

• RQ1 Can edge-attention representations significantly improve segmentation accuracy in medical images, especially in boundary localization?
• RQ2 How does the integration of edge guidance in early encoding layers affect the performance of a U-Net-like segmentation network?
• RQ3 To what extent does the weighted aggregation of multi-scale features with edge-attention improve final segmentation outcomes?
• RQ4 Does the proposed ET-Net generalize across diverse medical imaging tasks, including retinal, X-ray, and CT scans?
• RQ5 Can ET-Net outperform existing state-of-the-art methods without requiring task-specific architectural modifications?

Key findings

• ET-Net achieves a mIoU of 87.92% on the Drishti-GS dataset for optic disc and cup segmentation, outperforming the base network by 3.51 percentage points.
• On the DRIVE dataset for retinal vessel segmentation, ET-Net achieves 77.44% mIoU and 95.60% accuracy, surpassing U-Net, M-Net, and FCN.
• For lung segmentation in Chest X-Ray (MC dataset), ET-Net reaches 94.20% mIoU and 98.65% accuracy, exceeding previous SOTA methods.
• On the LUNA CT dataset, ET-Net achieves 96.23% mIoU and 98.68% accuracy, demonstrating strong performance on 3D volumetric imaging.
• Abllation studies confirm that both EGM and WAM contribute significantly to performance gains, with EGM alone improving mIoU by 2.08 points.
• Qualitative results show that ET-Net produces more accurate and continuous boundary predictions, especially in edge regions, compared to baseline models.