Source-linked AI summary
LadderNet: Multi-path networks based on U-Net for medical image segmentation
Juntang Zhuang
TL;DR
Medical image segmentation commonly uses U-Net, but its encoder-decoder variants have limited information-flow paths. LadderNet chains multiple U-Nets with shared-weights residual blocks, and it achieves superior performance on retinal blood-vessel segmentation benchmarks while reducing parameter growth.
Problem
U-Net variants retain encoder-decoder structures with limited information-flow paths, motivating a more highly connected architecture for medical image segmentation.
Method
LadderNet chains multiple U-Nets with skip connections between branches and shared-weights residual blocks that reduce the number of parameters.
Results
LadderNet achieves superior performance over previous literature methods on two public retinal blood-vessel segmentation datasets.
Takeaways & Limitations
LadderNet improves the key segmentation problem supporting automatic retinal disease detection and can also be used for tumor or brain-lesion segmentation.
Abstract
from arXiv · showhide
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}
1. INTRODUCTION
U-Net is widely used for medical image segmentation because its encoder-decoder structure and skip connections support efficient information flow. LadderNet addresses the limited number of information-flow paths in U-Net variants by introducing a multibranch design.
- U-Net is widely used in medical image analysis because encoder-decoder structure with skip connections supports efficient information flow and works with limited training data.
- Existing U-Net variants include recurrent convolution, attention modules, and dense blocks, but retain the encoder-decoder structure.
- LadderNet introduces more information-flow paths than prior U-Net variants for semantic segmentation.
- The paper evaluates LadderNet on retinal blood-vessel segmentation and reports superior performance over methods in the literature.
2. METHODS
LadderNet chains multiple U-Nets through multibranch encoder-decoder pathways and inter-U-Net skip connections. Shared-weights residual blocks increase path connectivity while reducing parameter growth.
- LadderNet: Columns 1 and 3 are encoder branches, while columns 2 and 4 are decoder branches across spatial levels A to E.
- LadderNet: LadderNet uses multiple encoder-decoder branches arranged as a chain of U-Nets, with additional U-Nets attachable to form more complex structures.
- LadderNet: Skip connections between adjacent U-Nets occur at levels A-D, and branch features are summed rather than concatenated.
- Connection between LadderNet and U-Net: LadderNet provides multiple information-flow paths that can each be viewed as a variant of a fully convolutional network.
- Shared-weights residual block: Shared-weights residual blocks reuse weights across two convolutional layers, add dropout, and use separate batch-normalization layers to reduce parameters.
3. EXPERIMENTS
The experiments evaluate LadderNet for retinal blood-vessel segmentation on DRIVE and CHASE DB1 using grayscale image patches and standard segmentation metrics. Training uses a five-level, 1.5M-parameter model optimized with cross-entropy and Adam.
- Datasets: DRIVE contains 40 retinal images, with 20 used for training and 20 for testing; training patches measure 48 × 48 pixels.
- Datasets: CHASE DB1 contains 28 retinal images, with 20 used for training and 8 for testing, and training patches measure 48 × 48 pixels.
- Preprocessing: All patches were converted to grayscale, and FOV masks were generated for CHASE DB1 before sampling across entire images.
- Training setup: The selected LadderNet has five levels, dropout rate 0.25, 10 channels at level A, and 1.5M parameters.
- Evaluation approaches: Evaluation uses accuracy, sensitivity, specificity, F1-score, ROC curves, and AUC based on TP, TN, FP, and FN.
4. RESULTS
LadderNet produces predictions visually close to ground truth on DRIVE and CHASE DB1, with strong curve-based and quantitative results across both tasks.
- Above 0.97 ROC AUC and 0.88 precision-recall AUC are reported for both retinal vessel-segmentation tasks.
- LadderNet generates predictions visually very close to the ground truth on the evaluated datasets.
- LadderNet achieves the highest F1-score, accuracy, and AUC for both DRIVE and CHASE DB1.
- The reported comparison emphasizes AC, AUC, and F1-score because they evaluate performance across both categories rather than focusing on one category.
5. DISCUSSION AND CONCLUSION
LadderNet expands U-Net with multiple encoder-decoder pairs and shared-weight residual blocks, achieving superior performance on two public retinal vessel-segmentation datasets.
- LadderNet uses more encoder-decoder pairs than U-Net, with skip connections that create multiple information-flow paths whose number grows exponentially with the pairs.
- Its shared-weights residual block combines residual connections, recurrent convolution, and drop-out regularization while greatly reducing parameter count.
- LadderNet shows superior performance over previous literature methods on two public datasets for retinal blood-vessel segmentation.
- The architecture is presented as applicable to other semantic-segmentation tasks, including tumor segmentation and brain-lesion detection.