[Paper Review] RV-GAN : Retinal Vessel Segmentation from Fundus Images using Multi-scale Generative Adversarial Networks.

Sharif Amit Kamran, Khondker Fariha Hossain+4 more|arXiv (Cornell University)|Jan 3, 2021

Retinal Imaging and Analysis25 references8 citations

TL;DR

RV-GAN proposes a multi-scale generative adversarial network with two generators and two multi-scale autoencoder discriminators for precise retinal vessel segmentation. By combining reconstruction and weighted feature matching loss, it achieves state-of-the-art performance with AUCs of 0.9887 (DRIVE), 0.9814 (CHASE-DB1), and 0.9887 (STARE), outperforming prior methods in AUC, Mean-IOU, and SSIM.

ABSTRACT

Retinal vessel segmentation contributes significantly to the domain of retinal image analysis for the diagnosis of vision-threatening diseases. With existing techniques the generated segmentation result deteriorates when thresholded with higher confidence value. To alleviate from this, we propose RVGAN, a new multi-scale generative architecture for accurate retinal vessel segmentation. Our architecture uses two generators and two multi-scale autoencoder based discriminators, for better microvessel localization and segmentation. By combining reconstruction and weighted feature matching loss, our adversarial training scheme generates highly accurate pixel-wise segmentation of retinal vessels with threshold >= 0.5. The architecture achieves AUC of 0.9887, 0.9814, and 0.9887 on three publicly available datasets, namely DRIVE, CHASE-DB1, and STARE, respectively. Additionally, RV-GAN outperforms other architectures in two additional relevant metrics, Mean-IOU and SSIM.

Motivation & Objective

• To address the degradation in segmentation accuracy when using high-confidence thresholds in retinal vessel segmentation.
• To improve microvessel localization and pixel-wise precision in retinal fundus images.
• To develop a multi-scale generative architecture that enhances feature representation across varying vessel scales.
• To outperform existing methods in both segmentation accuracy and structural similarity metrics.
• To integrate reconstruction and weighted feature matching losses for improved adversarial training stability and output fidelity.

Proposed method

• The architecture employs two generators to produce multi-scale segmentation predictions for enhanced vessel detail.
• Two multi-scale autoencoder-based discriminators are used to evaluate both local and global image realism and structural consistency.
• A hybrid loss function combines reconstruction loss and weighted feature matching loss to improve pixel-level accuracy and feature-level alignment.
• The adversarial training scheme is optimized to generate high-confidence segmentation outputs at threshold ≥ 0.5.
• The model is trained end-to-end on three public datasets: DRIVE, CHASE-DB1, and STARE.
• Feature matching loss is weighted to emphasize finer vessel structures and reduce false positives.

Experimental results

Research questions

• RQ1 Can a multi-scale GAN architecture improve the accuracy of retinal vessel segmentation at high confidence thresholds?
• RQ2 How does the integration of reconstruction and weighted feature matching loss affect segmentation fidelity in retinal images?
• RQ3 To what extent do multi-scale discriminators enhance microvessel localization compared to single-scale counterparts?
• RQ4 Does the proposed architecture outperform existing state-of-the-art methods in both AUC and structural similarity metrics?
• RQ5 Can the model generalize effectively across diverse retinal fundus image datasets with varying imaging conditions?

Key findings

• RV-GAN achieves an AUC of 0.9887 on the DRIVE dataset, indicating superior classification performance.
• On the CHASE-DB1 dataset, the model attains an AUC of 0.9814, demonstrating strong generalization on small-vessel regions.
• The model records an AUC of 0.9887 on the STARE dataset, confirming high accuracy across diverse retinal anatomy.
• RV-GAN outperforms existing architectures in Mean-IOU, indicating improved overlap between predicted and ground-truth vessel regions.
• The model also achieves higher SSIM values, reflecting better structural preservation in segmentation outputs.
• The combination of reconstruction and weighted feature matching loss significantly enhances pixel-wise segmentation accuracy at threshold ≥ 0.5.