Source-linked AI summary
Segmentation of optic disc, fovea and retinal vasculature using a single convolutional neural network
Jen Hong Tan, U. Rajendra Acharya, Sulatha V. Bhandary, Kuang Chua Chua, Sobha Sivaprasad
TL;DR
The paper addresses limited simultaneous segmentation of retinal vasculature, optic disc, and fovea. It uses a normalized-fundus-image CNN with multiscale three-channel inputs, achieving 92.68% accuracy on the DRIVE testing set. Boundary ambiguity remains a stated limitation, especially for optic disc and fovea.
Problem
Prior studies did not provide a single approach for automatically and simultaneously segmenting vasculature, optic disc, and fovea.
Method
A seven-layer CNN classifies each pixel using normalized images and three distinct input channels representing multiple neighbourhood scales.
Results
92.68% accuracy was achieved on the DRIVE testing set, with reported optic-disc and fovea results of 87.90% and an additional value truncated in the supplied passage.
Takeaways & Limitations
The proposed algorithm can simultaneously segment retinal vasculature, optic disc, and fovea, and is unlikely to segment unavailable features.
Takeaways & Limitations
Pixel membership can be ambiguous along boundaries, particularly for the optic disc and fovea.
Abstract
from arXiv · showhide
We have developed and trained a convolutional neural network to automatically and simultaneously segment optic disc, fovea and blood vessels. Fundus images were normalised before segmentation was performed to enforce consistency in background lighting and contrast. For every effective point in the fundus image, our algorithm extracted three channels of input from the neighbourhood of the point and forward the response across the 7 layer network. In average, our segmentation achieved an accuracy of 92.68 percent on the testing set from Drive database.
1. Introduction
Prior work largely handled retinal vasculature, optic-disc, and foveal analysis separately, with fovea studies usually locating rather than segmenting its region. This paper proposes simultaneous pixel classification with a seven-layer CNN using multiscale, three-channel inputs.
- Prior studies commonly used separate methods for vasculature, optic-disc, and fovea analysis rather than simultaneous segmentation.
- Fovea research primarily located its position, and overlap with ground truth was generally not reported for fovea segmentation.
- Existing approaches often assumed vasculature, optic disc, and fovea were present, preventing them from indicating feature absence.
- A seven-layer CNN simultaneously classifies each pixel as background, blood vessel, optic disc, or fovea.
- A 33 x 33 three-channel input combines a scaled 7 x 7 neighbourhood, a 33 x 33 neighbourhood, and a downscaled 165 x 165 region.
- Small 25 x 25 frames can make foveal and optic-disc pixels resemble background, while very large inputs increase training time and memory requirements.
2. Materials
The study used the DRIVE fundus-image database, divided into equal training and testing sets, with masks and manually delineated optic-disc and fovea ground truths.
- Thirty-three images show no diabetic-retinopathy symptoms, while the remainder show traits of mild early diabetic retinopathy.
- Each image includes a mask to exclude non-fundus regions during analysis.
- The images were captured with a Canon CR 5 non-mydriatic 3CCD camera using a 45-degree field of view and 565 x 584 resolution.
- The CNN was trained on the training set and tested on the testing set.
- Optic-disc and fovea ground truths were delineated by an ophthalmologist with more than 10 years of clinical experience.
3. Methods
The method normalizes fundus images, extracts three multiscale channels around each effective point, and processes them through a seven-layer CNN for simultaneous segmentation.
- 3.1 Normalization: A normalized colour image is produced by adjusting the luminance channel in LUV space before converting back to RGB.The procedure adapts prior green-channel normalization to retain a normalized colour image.
- 3.2 The architecture: Each effective point is represented by three 33 x 33 channels derived from 7 x 7, 33 x 33, and 165 x 165 neighbourhoods.The smallest and largest neighbourhoods are respectively scaled up and down using bi-cubic interpolation.
- 3.2 The architecture: 33 x 33 inputs gave the best accuracy among tested sizes from 9 x 9 to 45 x 45.The 9 x 9 input rarely exceeded 90% accuracy, whereas 45 x 45 reached about 91% at best.
- 3.3 Training: Training used standard backpropagation and stochastic gradient descent with batch size 10, learning rate 0.01, and regularization parameter 0.1.The cost function was a log-likelihood function, and only a portion of available points was used for training.
- 3.4 Testing: Testing used DRIVE effective points sampled at one in four per class across 40 epochs, selecting the best model for complete classification.The testing set contained 4,538,439 effective points, and the selected model classified all available testing points.
4. Results
Testing used a MATLAB-based implementation and achieved 92.68% average accuracy across 4,538,439 effective points. Performance varied across images, and the method simultaneously segmented vasculature, optic disc, and fovea while showing lower specificity linked to optic-disc vessel confusion.
- 92.68% average accuracy was achieved across the complete testing set of 4,538,439 effective points.
- 94.54% was the highest accuracy on one testing image, while 88.85% was the lowest.
- 3750.55 seconds was the average time required to completely segment an image.
- The network simultaneously segmented vasculature, optic disc, and fovea rather than only one or two structures, while performing reasonably well in sensitivity against previous works.
- Lower specificity was attributed partly to optic-disc pixels being misidentified as blood vessels and vice versa.
5. Discussion
The discussion identifies pixel-level ambiguity and neighborhood-size trade-offs, then presents a three-channel CNN as a way to combine local boundary detail with broader retinal patterns. The method segments all three structures, with reported class-specific accuracies and CPU/GPU processing constraints.
- Segmentation challenges: Pixel membership can be ambiguous at boundaries because visually sharp boundaries may not separate classes clearly at pixel level.The paper illustrates this ambiguity using a pixel that could be classified as background or blood vessel.
- Proposed design: A three-channel input uses differently scaled views to focus on nearby pixels for boundary ambiguity while capturing macro patterns around the point of interest.The design addresses the need for local detail in vessel and background discrimination and broader context for optic disc and fovea.
- Results: 87.90% and 88.53% of optic disc and fovea pixels were correctly identified, compared with 75.37% for blood vessels.The paper notes that blood vessels occupy a much larger presence in fundus images.
- Potential applications: Automated segmentation may support research on retinal vascular changes, diabetic retinopathy progression, treatment response, and glaucoma screening.The discussion links vessel measurements with cardiovascular disease, hypertension, diabetic retinopathy, and treatment monitoring, and links optic disc segmentation with glaucoma assessment.
- Computational considerations: 3751 seconds of CPU processing was estimated to decrease to just under 5 minutes on a GPU and become faster with multiple GPUs.The reported CPU timing was for segmentation of a single image.
6. Conclusion
The paper proposes a single seven-layer convolutional neural network with background normalization to simultaneously segment retinal vasculature, optic disc, and fovea. On the DRIVE database testing set, it achieved an average accuracy of 92.68%.
- Conclusion: A single convolutional neural network automatically and simultaneously segments retinal vasculature, optic disc, and fovea.The approach performs pixel-level classification across the three target structures and background.
- Conclusion: The algorithm combines background normalization with a 7-layer convolutional neural network.Normalization is used before segmentation, while the network performs the classification.
- Conclusion: 92.68% average accuracy was achieved on the testing set from the DRIVE database.The reported value is the average segmentation accuracy.