[Paper Review] LadderNet: Multi-path networks based on U-Net for medical image segmentation

Juntang Zhuang|arXiv (Cornell University)|Oct 17, 2018

Retinal Imaging and Analysis15 references222 citations

TL;DR

LadderNet extends U-Net into a multi-branch chain with shared-weights residual blocks, forming an ensemble of FCNs to improve retinal vessel segmentation. It achieves state-of-the-art metrics on DRIVE and CHASE_DB1 datasets.

ABSTRACT

U-Net has been providing state-of-the-art performance in many medical image segmentation problems. Many modifications have been proposed for U-Net, such as attention U-Net, recurrent residual convolutional U-Net (R2-UNet), and U-Net with residual blocks or blocks with dense connections. However, all these modifications have an encoder-decoder structure with skip connections, and the number of paths for information flow is limited. We propose LadderNet in this paper, which can be viewed as a chain of multiple U-Nets. Instead of only one pair of encoder branch and decoder branch in U-Net, a LadderNet has multiple pairs of encoder-decoder branches, and has skip connections between every pair of adjacent decoder and decoder branches in each level. Inspired by the success of ResNet and R2-UNet, we use modified residual blocks where two convolutional layers in one block share the same weights. A LadderNet has more paths for information flow because of skip connections and residual blocks, and can be viewed as an ensemble of Fully Convolutional Networks (FCN). The equivalence to an ensemble of FCNs improves segmentation accuracy, while the shared weights within each residual block reduce parameter number. Semantic segmentation is essential for retinal disease detection. We tested LadderNet on two benchmark datasets for blood vessel segmentation in retinal images, and achieved superior performance over methods in the literature. The implementation is provided \url{https://github.com/juntang-zhuang/LadderNet}

Motivation & Objective

• Motivate improving information flow in U-Net-like architectures beyond standard encoder-decoder skip connections.
• Introduce LadderNet, a chain of multiple U-Nets with skip connections between adjacent branches.
• Reduce parameters with shared-weights residual blocks while enabling multiple information-flow paths.
• Demonstrate superior segmentation performance on retinal vessel datasets compared with prior methods.

Proposed method

• Propose LadderNet as a multi-branch network forming a chain of U-Nets with skip connections between adjacent decoder branches.
• Use stride-2 convolutions to move to larger receptive fields and stride-2 transposed convolutions to propagate features.
• Double the number of channels at each level across successive columns.
• Implement shared-weights residual blocks where the two convolutions share the same weights and insert dropout between them.
• Treat LadderNet as an ensemble of FCNs due to multiple information-flow paths, while reducing parameters via weight sharing.

Experimental results

Research questions

• RQ1 Can LadderNet improve semantic segmentation performance by increasing information-flow paths beyond traditional U-Net architectures?
• RQ2 Does shared-weights residual block design reduce parameters while maintaining or improving accuracy?
• RQ3 How does LadderNet perform on retinal vessel segmentation compared with existing U-Net variants on DRIVE and CHASE_DB1?
• RQ4 Is LadderNet effectively an ensemble of FCNs, and does this translate to higher AUC and F1-scores?

Key findings

• LadderNet achieves the highest F1-score, accuracy (AC), and AUC across DRIVE and CHASE_DB1 compared with listed methods.
• On DRIVE, LadderNet attains F1-score 0.8202, SE 0.7856, SP 0.9810, AC 0.9561, AUC 0.9793.
• On CHASE_DB1, LadderNet attains F1-score 0.8031, SE 0.7978, SP 0.9818, AC 0.9656, AUC 0.9839.
• LadderNet’s ROC and precision-recall curves indicate top-tier performance with AUC above 0.97 and PR AUC above 0.88 on both datasets.
• The shared-weights residual blocks reduce parameter count while preserving or enhancing segmentation performance.