Source-linked AI summary
Optic Disc and Cup Segmentation Methods for Glaucoma Detection with Modification of U-Net Convolutional Neural Network
Artem Sevastopolsky
TL;DR
Glaucoma requires timely structural assessment, including cup-to-disc ratio estimation from optic disc and cup segmentation. The paper proposes a modified U-Net for universal automatic segmentation and reports quality similar to or better than existing methods, with low prediction time. However, optic-cup recognition remains more challenging than optic-disc recognition.
Problem
Timely glaucoma assessment requires optic disc and cup segmentation to estimate cup-to-disc ratio from fundus images.
Method
The paper uses a modified U-Net convolutional neural network for universal automatic optic disc and cup segmentation.
Results
The method provides results similar to or better than existing methods for both segmentation tasks while achieving low prediction time.
Takeaways & Limitations
The same method achieves high segmentation quality across optic disc and cup tasks, supporting its applicability to image-recognition problems.
Takeaways & Limitations
Optic-cup recognition is more challenging than optic-disc recognition.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Glaucoma damages the optic nerve irreversibly, making early diagnosis important. This section motivates automated optic disc and cup segmentation because clinicians currently spend substantial time measuring structures used to estimate cup-to-disc ratio.
- Clinical motivation: Early diagnosis matters because undetected glaucoma can cause irreversible optic-nerve damage leading to blindness.
- Clinical assessment: Cup-to-disc ratio is a representative structural indicator calculated from the heights of the optic cup and disc.
- Automation need: Optic disc and cup segmentation currently requires about eight minutes per eye for a skilled grader.
- Automation need: Automated analysis could support mass screening and care in countries lacking qualified specialists.
- Paper focus: The paper investigates a segmentation-based pipeline that estimates disc and cup dimensions for more transparent and reliable glaucoma assessment.
2. RELATED WORK
Prior work uses varied neural and image-processing pipelines for optic disc and cup segmentation, but several methods have practical limitations. These include advance cropping requirements, high computational demands, reproducibility problems, and difficult cross-method comparison.
- Existing approaches: Existing methods include fully convolutional networks, multi-scale convolutional models, boosted convolutional filters, and graph-cut post-processing.
- Existing approaches: One method outperformed existing approaches on DRISHTI-GS using Intersection-over-Union and Dice score, but required optic-disc or cup-area cropping beforehand.
- Limitations: Advance bounding boxes for optic disc and cup make that method inapplicable to previously unseen full-fundus images.
- Existing approaches: Another approach removed the need to crop by optic-cup area, making segmentation applicable to previously unseen images.
- Limitations: Some deep-network methods require long training, large parameter files, and substantial GPU memory.
- Limitations: Optic-cup segmentation is more challenging than optic-disc segmentation, and some studies are harder to compare because they use small, unavailable, or class-dependent evaluation settings.
- Limitations: Methods from and were described as complicated and difficult to reproduce, while their CPU prediction times were large.
3. THE PRESENTED APPROACH
The presented approach uses a modified U-Net convolutional neural network for automatic optic disc and optic cup segmentation. It combines preprocessing and task-specific pipelines with a lightweight architecture that can segment full eye fundus images without preliminary cropping.
- The method proposes a universal deep-learning approach for optic disc and optic cup segmentation based on a modified U-Net.
- CLAHE is used as preprocessing, with separate pipelines for optic disc and optic cup segmentation.For optic cup segmentation, images are cropped using the optic-disc bounding box with a margin.
- The network uses contracting and expansive paths that merge features across resolutions to recognize patterns at several scales.The architecture applies convolutional layers, max pooling, upsampling, dropout, and ReLU activation.
- Compared with the original U-Net, the modification uses fewer filters and no increasing filter count at lower resolutions.The authors report that these changes preserve recognition quality while reducing parameters and training time.
- The proposed method segments optic disc and optic cup directly on full eye fundus images without preliminary cropping to the optic-disc area.
4. EXPERIMENTS
The experiments compare the proposed segmentation solution with existing methods on DRIONS-DB, RIM-ONE v.3, and DRISHTI-GS using IoU and Dice scores, cross-validation, visual examples, and computational measurements. The method achieves competitive segmentation quality while emphasizing low prediction and training time, compactness, and implementation simplicity.
- Datasets and evaluation: The experiments compare optic disc and cup segmentation methods on DRIONS-DB, RIM-ONE v.3, and DRISHTI-GS.The datasets contain 110, 159, and 50 images, respectively, with varying cropping and ground-truth segmentation arrangements.
- Datasets and evaluation: Segmentation quality is evaluated with Intersection-over-Union and Dice score, using predicted probability maps and correct binary output maps.The measures are described as independent of image scale, object scale, and class imbalance.
- Visual results: 0.93 IoU and 0.97 Dice are reported for best-case optic-disc and optic-cup examples, while worst cases reach 0.80/0.90 and 0.46/0.64, respectively.The visual comparison is performed on RIM-ONE v.3.
- Efficiency and implementation: The proposed method has competitive quality in most score metrics, the lowest prediction time, and the lowest training time among deep-learning solutions.The reported model occupies 5 MB, compared with about 120 MB for DRIU; prediction-time estimates depend on the hardware used.
5. CONCLUSION
The conclusion presents a modified U-Net method that provides similar or better results than existing methods for optic disc and cup segmentation. It emphasizes high quality, simplicity, and low prediction time, while noting that optic cup segmentation remains more challenging and needs further research.
- The modified U-Net method provides results similar to or better than existing methods for optic disc and cup segmentation on eye fundus images.
- Applying the same method to both tasks achieves high segmentation quality, supporting its applicability to varied image-recognition problems.
- The proposed solution combines simple programming with modern frameworks, low prediction time, and a compact model.The conclusion identifies these as advantages of the method.
- Automatic optic disc segmentation can reach quality competitive with human performance in the reported experiments and visual comparisons.
- Optic cup segmentation is more challenging because its border is much more subtle, leaving room for improvement and further research.