[Paper Review] Retinal Vessel Segmentation Based on Conditional Deep Convolutional Generative Adversarial Networks

Yun Jiang, Ning Tan|arXiv (Cornell University)|May 11, 2018

Retinal Imaging and Analysis22 references3 citations

TL;DR

This paper proposes a conditional deep convolutional generative adversarial network (cDC-GAN) with a residual-based generator for retinal vessel segmentation, improving sensitivity and accuracy by using skip connections, channel-halving convolutions, and residual learning. On the DRIVE and STARE datasets, it achieves 96.08% accuracy, 82.74% sensitivity, and 82.08% F-measure, outperforming R2U-Net by 4.82% in sensitivity.

ABSTRACT

The segmentation of retinal vessels is of significance for doctors to diagnose the fundus diseases. However, existing methods have various problems in the segmentation of the retinal vessels, such as insufficient segmentation of retinal vessels, weak anti-noise interference ability, and sensitivity to lesions, etc. Aiming to the shortcomings of existed methods, this paper proposes the use of conditional deep convolutional generative adversarial networks to segment the retinal vessels. We mainly improve the network structure of the generator. The introduction of the residual module at the convolutional layer for residual learning makes the network structure sensitive to changes in the output, as to better adjust the weight of the generator. In order to reduce the number of parameters and calculations, using a small convolution to halve the number of channels in the input signature before using a large convolution kernel. By used skip connection to connect the output of the convolutional layer with the output of the deconvolution layer to avoid low-level information sharing. By verifying the method on the DRIVE and STARE datasets, the segmentation accuracy rate is 96.08% and 97.71%, the sensitivity reaches 82.74% and 85.34% respectively, and the F-measure reaches 82.08% and 85.02% respectively. The sensitivity is 4.82% and 2.4% higher than that of R2U-Net.

Motivation & Objective

• To address limitations in existing retinal vessel segmentation methods, such as poor sensitivity to thin vessels and weak noise robustness.
• To improve segmentation performance on challenging retinal images with lesions and low-contrast vessels.
• To reduce model complexity and computation by optimizing convolutional operations in the generator.
• To enhance feature learning and gradient flow through residual and skip connections in the generator architecture.

Proposed method

• The generator employs residual modules in convolutional layers to enable residual learning and improve sensitivity to output changes.
• A small convolutional layer is used before a large kernel convolution to halve the number of input channels, reducing parameters and computation.
• Skip connections link outputs of convolutional and deconvolutional layers to preserve low-level features and improve feature propagation.
• The conditional GAN framework uses real retinal images as conditions to guide the generator in producing precise vessel segmentation masks.
• The discriminator is trained to distinguish between real and generated segmentation masks, improving the realism and accuracy of predictions.
• The overall architecture is trained end-to-end using adversarial loss, with attention to preserving fine vessel structures.

Experimental results

Research questions

• RQ1 Can a conditional GAN with residual learning improve segmentation accuracy and sensitivity for retinal vessels?
• RQ2 How does the use of skip connections between convolutional and deconvolutional layers affect feature preservation and segmentation quality?
• RQ3 To what extent does channel reduction via small convolutions before large kernels reduce computational cost without degrading performance?
• RQ4 Does the proposed generator structure outperform existing U-Net variants like R2U-Net in sensitivity and F-measure on benchmark datasets?
• RQ5 How robust is the method to noise and lesions in retinal fundus images?

Key findings

• The proposed method achieves 96.08% segmentation accuracy on the DRIVE dataset, significantly improving upon baseline methods.
• Sensitivity reaches 82.74% on DRIVE, which is 4.82% higher than R2U-Net, indicating better detection of thin and faint vessels.
• On the STARE dataset, the method attains 97.71% accuracy and 85.34% sensitivity, demonstrating strong generalization.
• F-measure reaches 82.08% on DRIVE and 85.02% on STARE, confirming a balanced trade-off between precision and recall.
• The model reduces parameter count and computational load through channel-halving convolutions, improving efficiency without sacrificing performance.
• The use of skip connections and residual modules enhances feature learning and gradient flow, leading to more stable and accurate training.