Source-linked AI summary
Retinal Vessel Segmentation in Fundoscopic Images with Generative Adversarial Networks
Jaemin Son, Sang Jun Park, Kyu-Hwan Jung
TL;DR
Retinal vessel segmentation methods can miss fine vessels or produce false positives because pixel-wise CNN objectives do not model natural vascular structure. This paper uses adversarial training, with a generator producing vessel maps and a discriminator comparing them with expert annotations. The method reports state-of-the-art performance on DRIVE and STARE while reducing false positives at fine vessels, though it fails to detect some one-pixel-thin vessels.
Problem
Existing CNN vessel-segmentation methods can produce blurry vessels and false positives because pixel-wise objectives do not accommodate natural vascular structure.
Method
The method uses adversarial training in which a generator produces vessel maps from fundus images and a discriminator distinguishes generated maps from expert annotations.
Results
The method achieves state-of-the-art performance on DRIVE and STARE and reports fewer false positives at fine vessels than existing methods.
Takeaways & Limitations
Adversarial training can improve retinal vessel maps by producing clearer vessel lines with adequate detail and fewer fine-vessel false positives.
Takeaways & Limitations
The method fails to detect very thin vessels spanning only 1 pixel.
Abstract
from arXiv · showhide
Retinal vessel segmentation is an indispensable step for automatic detection of retinal diseases with fundoscopic images. Though many approaches have been proposed, existing methods tend to miss fine vessels or allow false positives at terminal branches. Let alone under-segmentation, over-segmentation is also problematic when quantitative studies need to measure the precise width of vessels. In this paper, we present a method that generates the precise map of retinal vessels using generative adversarial training. Our methods achieve dice coefficient of 0.829 on DRIVE dataset and 0.834 on STARE dataset which is the state-of-the-art performance on both datasets.
1 Introduction
Retinal vessel segmentation supports analysis of ocular and systemic vascular conditions, but existing CNN methods can produce blurry vessels and false positives because pixel-wise objectives do not model natural vascular structure. The paper proposes adversarial training to generate sharper vessel maps that resemble expert annotations.
- Fundus vessel segmentation can support detection of ocular diseases and signs of systemic vascular burden.Abnormal retinal vascular structures are associated with hypertension, diabetes mellitus, retinal vein occlusion, and retinal artery occlusion.
- Earlier approaches relied on signal-processing assumptions, line detection, and hand-crafted features before automatic feature learning improved results.
- CNN-based methods can surpass human experts on multiple datasets but may produce blurry vessels and false positives around minuscule or faint branches.
- Pixel-wise objectives do not actively accommodate the natural vascular structure present in fundus images.
- Adversarial training constrains generated vessel maps to resemble expert annotations, targeting clearer and sharper segmentation outputs.The discriminator distinguishes expert gold-standard maps from generator outputs while the generator seeks outputs the discriminator cannot differentiate.
2 Proposed Methods
The proposed method treats vessel segmentation as conditional image translation: a generator produces vessel probability maps from fundus images, while adversarial and segmentation objectives train outputs toward expert annotations. The framework explores discriminator decision scales and combines adversarial training with supervised segmentation loss.
- Network Structure: The generator maps each fundus image to a same-sized retinal-vessel probability map, while the discriminator judges whether a vessel image is expert-annotated or generated.Probability values range from 0 to 1, representing the likelihood that each pixel belongs to a vessel.
- Network Structure: U-Net-style skip connections transfer low-level edges and blobs from initial convolutional maps to upsampled layers for segmentation.
- Network Structure: The discriminator can make pixel-level, patch-level, or image-level authenticity decisions, and the study evaluates Pixel GAN, Patch GAN, and Image GAN variants.
- Objective Function: The GAN objective trains the discriminator to classify expert pairs as real and generated pairs as fake while the generator produces outputs that resist this discrimination.
- Objective Function: The segmentation loss penalizes distance between gold-standard vessel maps and generator outputs, and λ balances it against the GAN objective.
3 Experiments
Experiments evaluate GAN-based retinal vessel segmentation on DRIVE and STARE using ROC AUC, PR AUC, and Dice, comparing discriminator designs and existing methods. Results show improved segmentation quality, sharper probability maps, and fewer false positives relative to DRIU, with some increased false negatives near uncertain terminal vessels.
- Evaluation: Performance is measured with ROC AUC, PR AUC, and Dice coefficient after thresholding probability maps with Otsu thresholding.For STARE, a field-of-view mask is generated by detecting a central blob because the dataset lacks mask images.
- Discriminator comparison: Image GAN outperforms U-Net and patch GANs, indicating that discriminator-based training improves segmentation quality and that stronger image-level discrimination performs best among tested models.The comparison covers models with different discriminator outputs on both datasets.
- Comparison with existing methods: The method achieves higher ROC AUC and Dice coefficient than DRIU and surpasses the human annotator on DRIVE, while outperforming other compared methods across most operating regimes.The comparison uses ROC and PR curves plus AUC and Dice summaries for DRIVE and STARE.
- Qualitative analysis: Probability maps from the proposed method more closely match the gold standard, whereas DRIU assigns overconfident probabilities to fine vessels and vessel–background boundaries.The overconfident predictions can produce over-segmentation in these regions.
- Qualitative analysis: After Otsu thresholding, DRIU generally produces more false positives, while the proposed method produces more false negatives around terminal vessels and uncertain regions.The latter behavior is described as resembling human annotator uncertainty.
4 Conclusion and Discussion
The paper concludes that adversarial training improves retinal vessel segmentation accuracy and clarity across public datasets. The method reduces false positives and produces clearer vessel lines, but still misses vessels only one pixel wide.
- Conclusion: Adversarial training helps segment retinal vessels more accurately and clearly, and the method outperforms existing methods on ROC AUC, PR AUC, and Dice coefficient.The conclusion attributes the improvement to the presence of a discriminator.
- Discussion: Compared with the best existing method, the proposed approach yields fewer false positives at fine vessels and clearer lines with details resembling human annotation.The qualitative comparison is illustrated using probability maps from DRIVE and STARE.
- Limitations and future work: The method fails to detect very thin vessels spanning only 1 pixel, and added structural priors such as connectivity are suggested as a way to improve performance.This identifies a remaining limitation and a direction for future work.