Source-linked AI summary
Deep supervision with additional labels for retinal vessel segmentation task
Yishuo Zhang, Albert C. S. Chung
TL;DR
Retinal vessel segmentation is important for ophthalmological diagnosis but is difficult under variable imaging conditions and for tiny vessels. The paper combines residual U-net segmentation with edge-aware multi-class labels and deep supervision, achieving strong benchmark performance and efficient runtime. It reports the best AUC performance on DRIVE and CHASEDB1 among the compared methods, with 97.99% AUC on DRIVE.
Problem
Retinal vessel segmentation is important for ophthalmological diagnosis but remains difficult because imaging variability, lesions, and tiny low-contrast vessels hinder accurate detection.
Method
The method uses a residual U-net with separate labels for boundaries, thick vessels, and thin vessels, plus side-output deep supervision.
Results
The method achieves the best AUC performance on DRIVE and CHASEDB1 among the compared methods, including 97.99% AUC on DRIVE.
Takeaways & Limitations
The approach provides fine vessel segmentation, especially for capillaries, while maintaining efficient running time for clinical applications.
Abstract
from arXiv · showhide
Automatic analysis of retinal blood images is of vital importance in diagnosis tasks of retinopathy. Segmenting vessels accurately is a fundamental step in analysing retinal images. However, it is usually difficult due to various imaging conditions, low image contrast and the appearance of pathologies such as micro-aneurysms. In this paper, we propose a novel method with deep neural networks to solve this problem. We utilize U-net with residual connection to detect vessels. To achieve better accuracy, we introduce an edge-aware mechanism, in which we convert the original task into a multi-class task by adding additional labels on boundary areas. In this way, the network will pay more attention to the boundary areas of vessels and achieve a better performance, especially in tiny vessels detecting. Besides, side output layers are applied in order to give deep supervision and therefore help convergence. We train and evaluate our model on three databases: DRIVE, STARE, and CHASEDB1. Experimental results show that our method has a comparable performance with AUC of 97.99% on DRIVE and an efficient running time compared to the state-of-the-art methods.
1 Introduction
Retinal vessel segmentation is important for ophthalmological diagnosis but remains difficult because vessel appearance, image quality, interfering structures, and tiny capillaries vary widely. The paper addresses these challenges with edge-aware labels and deep supervision in U-net.
- Retinal vessel segmentation supports ophthalmological diagnosis, including assessment of diabetic retinopathy severity.
- Variable vessel shape, image resolution and contrast, interfering structures, and low-contrast tiny vessels make segmentation challenging.
- Existing methods still struggle with tiny capillaries and running time, motivating a more effective and efficient approach.
- The method labels boundaries, thick vessels, and thin vessels as separate classes so networks emphasize error-prone regions.
- Deep supervision is added to help networks converge and learn better semantic representations.
2 Proposed Method
The proposed architecture combines residual U-net processing with five-class edge-aware labels and deep supervision. Additional boundary and vessel classes receive stronger supervision, while side outputs provide intermediate semantic guidance for the final prediction.
- 2.1 U-net: Residual U-net blocks downsample and reconstruct feature maps while skip connections preserve details useful for fine boundary prediction.
- 2.2 Additional Label: Training converts vessel segmentation into five classes separating ordinary background, boundary regions, thick vessels, and thin vessels.
- 2.2 Additional Label: Boundary classes receive heavier loss weights, assigning higher penalties to errors in crucial regions without changing supervision elsewhere.
- 2.3 Deep Supervision: Four side-output layers compare intermediate predictions with ground truth, and their outputs are fused to generate the final prediction.
- 2.3 Deep Supervision: The training objective combines fused-output loss, side-output losses, class-balanced cross-entropy, and L-2 regularization.
3 Experiments
The experiments evaluate the method on DRIVE, STARE, and CHASEDB1 using standard pixel-wise metrics and qualitative outputs. The method improves thin-vessel detection over simple U-net, achieves strong cross-dataset AUC performance, and runs efficiently.
- 3 Experiments: Experiments use DRIVE, STARE, and CHASEDB1, with annotated masks from different experts supporting training, testing, and human-observer comparison.DRIVE has 20 training and 20 testing images, while STARE and CHASEDB1 contain 20 and 28 images, respectively.
- 3 Experiments: Evaluation reports Specificity, Sensitivity, Accuracy, and AUC for this unbalanced pixel-wise vessel-classification task.AUC is included because the authors consider it more suitable for measuring an unbalanced situation.
- 3.2 Results: Qualitative analyses show that side outputs can identify vessel boundaries that guide precise final predictions, even when tiny vessels are missed.The authors describe mutual promotion between additional labels and deep supervision.
- 3.2 Results: AUC for thin vessels increases by 9.11% compared with a simple U-net.The comparison attributes the improvement to additional labels and deep supervision, especially for capillary detection.
- 3.3 Comparison: The method achieves the best AUC performance on DRIVE and CHASEDB1 among the compared methods.The authors prefer AUC because dataset errors and class imbalance make it more equitable for comparison.
- 3.3 Comparison: The method processes a 584*565 image in 1.2s, making it computationally efficient relative to several compared methods.The authors attribute the speed to U-net processing patches rather than predicting only each patch’s central pixel.
4 Conclusion
The proposed retinal vessel segmentation network combines multi-class edge-aware labeling with residual U-net connections and deep supervision. It reports good performance and efficient running time compared with other state-of-the-art methods.
- Thick vessels, thin vessels, and boundaries are assigned different classes to create a multi-class segmentation task.
- The method reports good performance and efficient running time compared with other state-of-the-art methods.
- The model uses a U-net with residual connections for retinal vessel segmentation.
- Deep supervision helps the network learn better features and semantic information.