[Paper Review] Exploring The Limits Of Data Augmentation For Retinal Vessel Segmentation

Enes Sadi Uysal, Mücahit Bilici+3 more|ArXiv.org|May 19, 2021

Retinal Imaging and Analysis22 references12 citations

TL;DR

This paper proposes a U-Net-based retinal vessel segmentation model that achieves state-of-the-art performance on the DRIVE dataset through an iterative, problem-driven data augmentation strategy. By systematically applying augmentations—such as rotation, flipping, gamma correction, and noise injection—targeted at specific image quality issues (e.g., thin vessels, brightness variation, noise), the model achieves an AUC of 0.9855 and accuracy of 0.9712, outperforming more complex architectures with minimal architectural changes.

ABSTRACT

Retinal Vessel Segmentation is important for the diagnosis of various diseases. The research on retinal vessel segmentation focuses mainly on the improvement of the segmentation model which is usually based on U-Net architecture. In our study, we use the U-Net architecture and we rely on heavy data augmentation in order to achieve better performance. The success of the data augmentation relies on successfully addressing the problem of input images. By analyzing input images and performing the augmentation accordingly we show that the performance of the U-Net model can be increased dramatically. Results are reported using the most widely used retina dataset, DRIVE.

Motivation & Objective

• To investigate whether heavy, problem-specific data augmentation can significantly improve retinal vessel segmentation performance without architectural complexity.
• To identify and address specific input image quality issues—such as low contrast, noise, and brightness variation—in fundus images through tailored augmentation strategies.
• To demonstrate that iterative, failure-driven augmentation can outperform standard augmentation protocols in medical image segmentation.
• To achieve state-of-the-art performance on the DRIVE benchmark using only standard U-Net architecture and extensive data augmentation.

Proposed method

• The model uses the standard U-Net architecture with binary cross-entropy loss and Adam optimization (initial learning rate 1e-4).
• A progressive data augmentation pipeline is applied, where new augmentation techniques are added based on model failure analysis from previous iterations.
• Augmentations include rotation, flipping, cropping, zooming, gamma correction, noise injection, elastic distortions, blurring, dropout, and histogram equalization.
• The augmentation strategy specifically targets thin vessel segmentation, image brightness issues, and noise robustness, informed by visual error analysis.
• Training is performed on a single RTX 2080 for 15 epochs with mini-batches of size 3, without pre-trained weights.
• Performance is evaluated using AUC, accuracy, and mean dice score on the DRIVE dataset.

Experimental results

Research questions

• RQ1 Can a standard U-Net model achieve state-of-the-art performance in retinal vessel segmentation through extensive, targeted data augmentation?
• RQ2 Which specific image quality problems in fundus images most significantly hinder segmentation performance, and can they be systematically addressed via augmentation?
• RQ3 Does an iterative, failure-driven data augmentation strategy lead to greater performance gains than standard or random augmentation protocols?
• RQ4 How does the performance of a simple U-Net with heavy augmentation compare to more complex architectures in retinal vessel segmentation?

Key findings

• The proposed method achieves an AUC of 0.9855 and accuracy of 0.9712 on the DRIVE dataset, outperforming most models with more complex architectures.
• The model achieves a mean dice score of 0.8255, indicating strong boundary-level segmentation performance.
• Despite using only a standard U-Net, the model outperforms several advanced architectures, including Residual U-Net and DU-Net, in terms of AUC and accuracy.
• The model is within 0.0032 AUC and 0.0078 accuracy points of RV-GAN, a GAN-based model that requires up to 48 hours of training.
• The most significant performance gains came from augmentations targeting thin vessel segmentation, brightness variation (via gamma correction), and noise robustness.
• Each new augmentation technique was introduced based on failure analysis, demonstrating that iterative, problem-specific augmentation is highly effective.