Source-linked AI summary
A Fully Convolutional Neural Network based Structured Prediction Approach Towards the Retinal Vessel Segmentation
Avijit Dasgupta, Sonam Singh
TL;DR
Automatic retinal vessel segmentation matters for computer-aided diagnosis but remains difficult because vessel morphology varies against noisy backgrounds. The paper formulates segmentation as multi-label inference and combines convolutional neural networks with structured prediction, reporting strong performance on DRIVE, including 95.33% accuracy and 0.974 AUC.
Problem
Automatic retinal blood-vessel segmentation from fundus images is important for diagnosis but is challenging because vessels vary morphologically against noisy backgrounds.
Method
The paper uses a fully convolutional neural network architecture capable of structured prediction for retinal vessel segmentation.
Results
The proposed method significantly improves performance over other state-of-the-art results on the DRIVE dataset.
Takeaways & Limitations
The architecture demonstrates state-of-the-art performance for automatic retinal vessel segmentation on the DRIVE database.
Abstract
from arXiv · showhide
Automatic segmentation of retinal blood vessels from fundus images plays an important role in the computer aided diagnosis of retinal diseases. The task of blood vessel segmentation is challenging due to the extreme variations in morphology of the vessels against noisy background. In this paper, we formulate the segmentation task as a multi-label inference task and utilize the implicit advantages of the combination of convolutional neural networks and structured prediction. Our proposed convolutional neural network based model achieves strong performance and significantly outperforms the state-of-the-art for automatic retinal blood vessel segmentation on DRIVE dataset with 95.33% accuracy and 0.974 AUC score.
1. INTRODUCTION
Retinal vessel segmentation supports diagnosis but is difficult to automate because vessels vary in branching patterns and appear against noisy, tortuous backgrounds. Prior work spans unsupervised techniques, handcrafted-feature supervised models, and increasingly deep convolutional approaches, motivating the paper’s fully convolutional architecture.
- Retinal vessel segmentation provides an important cue for diagnosing diseases including diabetes, hypertension, microaneurysms, and arteriosclerosis.
- Manual blood-vessel segmentation is tedious and time-consuming, motivating automatic segmentation from fundus images.
- Automatic segmentation is challenging because vessels exhibit abrupt branching-pattern variations, tortuosity, and noisy backgrounds.
- Earlier methods included unsupervised vessel tracking, adaptive thresholding, morphology-based techniques, and supervised models trained on hand-labeled images.
- Recent supervised approaches increasingly use deep convolutional networks, including single networks, ensembles, and stacked denoising autoencoder architectures.
- The paper proposes a fully convolutional neural network architecture for retinal blood-vessel segmentation.
2. PROPOSED METHODOLOGY
The proposed methodology formulates retinal vessel segmentation as structured multi-label prediction using a fully convolutional CNN. It preprocesses fundus images, predicts neighborhood labels from image patches, and produces structured outputs.
- 2.1. Problem Statement: Each pixel neighborhood is classified into vessels or background through a learned function H(ω|I, N(x, y)).The input is a color fundus image, and the prediction targets every pixel in the neighborhood N(x, y).
- 2.2. Convolutional Neural Networks: The CNN transforms three-dimensional feature volumes through convolutional, pooling, and upsampling operations.Pooling downsamples spatial feature maps, while upsampling increases their spatial dimensions.
- 2.3. Preprocessing and Data Preparation: The preprocessing pipeline extracts the green channel, normalizes it using the data mean and standard deviation, applies contrast and gamma adjustments, and scales intensities to [0, 1].Vessels manifest high contrast in the green channel, and the resulting preprocessed image is denoted by Î.
- 2.4. The Proposed Architecture: A 1 × 28 × 28 patch from the preprocessed image Î enters the first convolutional layer.The CNN learns task-dependent hierarchical features from these patches.
- 2.4. The Proposed Architecture: The architecture uses convolution, 2 × 2 max-pooling, upsampling, and convolutional layers to return a 1 × 28 × 28 structured output.The model uses 32-filter and 64-filter convolutional stages, same-size padding, and a final softmax layer.
- 2.4. The Proposed Architecture: Cross-entropy loss compares predicted and ground-truth vectors whose dimensions match the pixel neighborhood N(x, y).The multi-label formulation predicts a vector rather than a scalar value.
3. RESULTS AND DISCUSSIONS
The proposed method is evaluated on the DRIVE dataset using standard segmentation metrics and is reported to improve substantially over recent state-of-the-art methods. Qualitative visualizations show predictions alongside preprocessed images and ground truth.
- The method is evaluated on the publicly available DRIVE dataset.
- Performance is assessed using Precision, Sensitivity, Specificity, Accuracy, and AUC.Training uses a learning rate of 0.0001, RMSprop with momentum 0.7, 60 epochs, and batch size 32.
- The proposed method demonstrates significant improvement over recent state-of-the-art results.
- Table 1 presents a quantitative comparison on DRIVE between the proposed method and existing state-of-the-art methods.
- Figure 3 visualizes outputs for three randomly selected DRIVE samples using preprocessed images, ground truth, and segmented outputs.Additional visualizations and intermediate results are referenced on the project website.
4. CONCLUSION
The paper proposes a fully convolutional architecture for structured prediction in retinal vessel segmentation and reports state-of-the-art performance on the DRIVE database.
- The proposed fully convolutional architecture performs structured prediction for retinal vessel segmentation.
- The architecture achieves state-of-the-art performance on the DRIVE database.