Optic Disc and Cup Segmentation Methods for Glaucoma Detection with Modification of U-Net Convolutional Neural Network

Artem Sevastopolsky

Abstract

Glaucoma is a leading cause of irreversible blindness, requiring precise segmentation of the optic disc and cup for diagnosis via the cup-to-disc ratio (CDR). Manual segmentation is time-consuming and prone to variability. This paper presents a lightweight, modified U-Net architecture designed for efficient automated segmentation. By reducing filter counts and employing a two-stage pipeline with automated cropping, the proposed method achieves state-of-the-art accuracy with significantly reduced model size (5 MB) and inference time (0.1s per image). These improvements make the system highly suitable for mass screening and mobile health applications.

Introduction

Glaucoma diagnosis relies heavily on examining the optic nerve head in eye fundus images. Specifically, clinicians measure the cup-to-disc ratio (CDR)—the size of the optic cup relative to the optic disc. An enlarged cup often indicates potential glaucoma. However, manual segmentation of these regions is a tedious process, often taking up to eight minutes per eye for a skilled grader.

Comparison of healthy and glaucoma-suspicious eyes with segmented optic disc and cup boundaries.

While automated computer vision methods exist, they frequently rely on heavy neural networks, require manual pre-cropping, or are too computationally expensive for real-time use. There is a critical need for a fast, accurate, and fully automated system that can process raw fundus images to assist in mass screening, particularly in regions with limited access to specialists.

Methodology

The proposed approach utilizes a deep learning pipeline centered on a modified U-Net convolutional neural network architecture.

Pre-processing and Architecture

Images are first enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE) to standardize lighting and contrast. The core model is a "slimmed down" U-Net. Unlike the standard architecture, which increases filter counts at lower resolutions, this modification maintains a constant, lower filter count across all resolutions. This significantly reduces the parameter count without sacrificing performance on medical imaging features. The network is trained using a differentiable extension of the Dice score, optimized via Stochastic Gradient Descent (SGD) with extensive data augmentation.

Two-Stage Pipeline

The system operates in two stages to handle the varying difficulty of the tasks:

  1. Optic Disc Segmentation: The network processes the full fundus image directly.
  2. Optic Cup Segmentation: Because the cup is harder to define, the system uses the optic disc prediction to automatically crop the image, feeding only the relevant region into a second network.
The two-stage segmentation pipeline for the optic cup, featuring automated cropping and CLAHE enhancement.

Datasets and Evaluation

The method was evaluated on three public datasets:

  • DRIONS-DB: 110 images (optic disc only).
  • RIM-ONE v.3: 159 images (disc and cup).
  • DRISHTI-GS: 50 images (disc and cup).

Results and Discussion

The modified U-Net achieved highly competitive results compared to much larger models. For optic disc segmentation, it reached a Dice score of 0.94 on DRIONS-DB and 0.95 on RIM-ONE v.3. Optic cup segmentation, while more challenging due to subtle boundaries, achieved a Dice score of 0.85 on DRISHTI-GS.

The most significant advantage of this model is its efficiency. Prediction takes only 0.1 seconds for the disc and 0.06 seconds for the cup on a standard GPU—over 50 times faster than many ensemble-based state-of-the-art methods. Furthermore, the model size is just 5 MB (compared to 120 MB for models like DRIU), making it ideal for deployment on mobile devices or low-power clinical hardware.

Conclusion

This research demonstrates that "bigger isn't always better" in medical image segmentation. By streamlining the U-Net architecture and automating the cropping process, the author provides a robust, mobile-ready solution for glaucoma screening. Future work will focus on improving feature extraction for the subtle boundaries of the optic cup to further enhance diagnostic reliability.

Summary updated

Abstract

Glaucoma is the second leading cause of blindness all over the world, with approximately 60 million cases reported worldwide in 2010. If undiagnosed in time, glaucoma causes irreversible damage to the optic nerve leading to blindness. The optic nerve head examination, which involves measurement of cup-to-disc ratio, is considered one of the most valuable methods of structural diagnosis of the disease. Estimation of cup-to-disc ratio requires segmentation of optic disc and optic cup on eye fundus images and can be performed by modern computer vision algorithms. This work presents universal approach for automatic optic disc and cup segmentation, which is based on deep learning, namely, modification of U-Net convolutional neural network. Our experiments include comparison with the best known methods on publicly available databases DRIONS-DB, RIM-ONE v.3, DRISHTI-GS. For both optic disc and cup segmentation, our method achieves quality comparable to current state-of-the-art methods, outperforming them in terms of the prediction time.
BibTeX
@misc{sevastopolsky2017optic,
  author = {Artem Sevastopolsky},
  title = {{Optic Disc and Cup Segmentation Methods for Glaucoma Detection with Modification of U-Net Convolutional Neural Network}},
  year = {2017},
  eprint = {1704.00979},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1704.00979},
}
Figure