[Paper Review] Dual Encoder Fusion U-Net (DEFU-Net) for Cross-manufacturer Chest X-ray Segmentation

Lipei Zhang, Aozhi Liu+2 more|arXiv (Cornell University)|Sep 11, 2020

COVID-19 diagnosis using AI41 references4 citations

TL;DR

This paper proposes DEFU-Net, a dual-path encoder fusion U-Net architecture that combines inception blocks with dilated convolutions and densely connected recurrent convolutional blocks to enhance spatial and contextual feature extraction for cross-manufacturer chest X-ray segmentation. The model achieves state-of-the-art performance on mixed datasets from Montgomery and Shenzhen hospitals, outperforming U-Net, Res-U-Net, BCDU-Net, R2U-Net, and attention-enhanced variants in Dice, IOU, F1-score, and AUC across multiple metrics.

ABSTRACT

A number of methods based on deep learning have been applied to medical image segmentation and have achieved state-of-the-art performance. Due to the importance of chest x-ray data in studying COVID-19, there is a demand for state-of-the-art models capable of precisely segmenting soft tissue on the chest x-rays. The dataset for exploring best segmentation model is from Montgomery and Shenzhen hospital which had opened in 2014. The most famous technique is U-Net which has been used to many medical datasets including the Chest X-rays. However, most variant U-Nets mainly focus on extraction of contextual information and skip connections. There is still a large space for improving extraction of spatial features. In this paper, we propose a dual encoder fusion U-Net framework for Chest X-rays based on Inception Convolutional Neural Network with dilation, Densely Connected Recurrent Convolutional Neural Network, which is named DEFU-Net. The densely connected recurrent path extends the network deeper for facilitating contextual feature extraction. In order to increase the width of network and enrich representation of features, the inception blocks with dilation are adopted. The inception blocks can capture globally and locally spatial information from various receptive fields. At the same time, the two paths are fused by summing features, thus preserving the contextual and spatial information for decoding part. This multi-learning-scale model is benefiting in Chest X-ray dataset from two different manufacturers (Montgomery and Shenzhen hospital). The DEFU-Net achieves the better performance than basic U-Net, residual U-Net, BCDU-Net, R2U-Net and attention R2U-Net. This model has proved the feasibility for mixed dataset and approaches state-of-the-art. The source code for this proposed framework is public https://github.com/uceclz0/DEFU-Net.

Motivation & Objective

• To address the challenge of segmenting soft tissue in chest X-rays across different imaging manufacturers, where domain shifts and variability in image quality affect model generalization.
• To improve feature extraction by simultaneously enhancing spatial perception and contextual understanding in medical image segmentation.
• To develop a robust model capable of handling mixed datasets from diverse sources (Montgomery and Shenzhen hospitals) without performance degradation.
• To outperform existing U-Net variants in segmentation accuracy, especially on edge and small-region pixels prone to misclassification.
• To enable faster convergence and better fitting on heterogeneous data through architectural innovations in the encoder path.

Proposed method

• DEFU-Net employs a dual-path encoder: one path uses inception blocks with dilated convolutions to capture multi-scale spatial features across varying receptive fields.
• The second path implements densely connected recurrent convolutional blocks (DCRC) to deepen the network and enhance feature propagation across layers.
• Features from both encoder paths are fused via element-wise summation before decoding, preserving rich contextual and spatial information for accurate segmentation.
• The network uses skip connections between encoder and decoder paths, similar to U-Net, to maintain spatial resolution and gradient flow.
• The number of feature maps in the bottleneck layer is set to match the fourth encoder level, reducing computational cost by half.
• The model is trained using cross-entropy loss with early stopping and Adam optimization, achieving fast convergence and high performance.

Experimental results

Research questions

• RQ1 Can a dual-path encoder architecture improve segmentation performance on cross-manufacturer chest X-ray datasets compared to standard U-Net variants?
• RQ2 Does the integration of dilated inception blocks enhance multi-scale spatial feature extraction in low-contrast, edge-fuzzy medical images?
• RQ3 To what extent does the densely connected recurrent block improve feature representation depth and context modeling in the encoder?
• RQ4 How does feature fusion via summation of dual-path features affect decoding accuracy, especially near organ borders?
• RQ5 Can the model generalize effectively across datasets from different imaging manufacturers without fine-tuning?

Key findings

• DEFU-Net achieved the highest Dice score of 0.9667 and F1-score of 0.9619 on the test set, outperforming U-Net, Res-U-Net, BCDU-Net, Incep-Res-U-Net, R2U-Net, and Att-R2U-Net.
• On cross-manufacturer generalization, DEFU-Net achieved a Dice score of 0.9227 when training on Shenzhen and testing on Montgomery, and 0.9154 when training on Montgomery and testing on Shenzhen, demonstrating strong robustness.
• The model converged faster than baseline models, reaching an accuracy of 0.9776 after just two epochs, and trained for 145 epochs without overfitting.
• The use of dilated inception blocks significantly improved performance over standard 3×3 convolutions, especially in capturing global and local spatial patterns.
• Visualizations confirmed that DEFU-Net better aligns predictions with ground truth, particularly in edge regions and small lung areas, reducing false positives and false negatives.
• The combination of dilated inception and DCRC blocks led to superior performance in handling domain shifts and image variability across manufacturers.