Source-linked AI summary
Disc-aware Ensemble Network for Glaucoma Screening from Fundus Image
Huazhu Fu, Jun Cheng, Yanwu Xu, Changqing Zhang, Damon Wing Kee Wong, Jiang Liu, Xiaochun Cao
TL;DR
Glaucoma causes irreversible vision loss, making early screening important, while existing clinical screening includes optic nerve head assessment. DENet screens fundus images using four disc-aware streams, achieving the best reported performance on SCES and SINDI among compared methods.
Problem
Glaucoma causes irreversible vision loss, and early screening is essential for early treatment to preserve vision and maintain life quality.
Method
DENet integrates four deep streams covering global fundus structure, segmentation-guided disc information, local disc regions, and disc polar transformation.
Results
DENet achieves 0.9183 AUC on the SCES dataset and 0.8173 AUC on the SINDI dataset, the best performances among the compared methods.
Takeaways & Limitations
The ensemble incorporates hierarchical representations and contextual information from the optic disc region for glaucoma screening.
Abstract
from arXiv · showhide
Glaucoma is a chronic eye disease that leads to irreversible vision loss. Most of the existing automatic screening methods firstly segment the main structure, and subsequently calculate the clinical measurement for detection and screening of glaucoma. However, these measurement-based methods rely heavily on the segmentation accuracy, and ignore various visual features. In this paper, we introduce a deep learning technique to gain additional image-relevant information, and screen glaucoma from the fundus image directly. Specifically, a novel Disc-aware Ensemble Network (DENet) for automatic glaucoma screening is proposed, which integrates the deep hierarchical context of the global fundus image and the local optic disc region. Four deep streams on different levels and modules are respectively considered as global image stream, segmentation-guided network, local disc region stream, and disc polar transformation stream. Finally, the output probabilities of different streams are fused as the final screening result. The experiments on two glaucoma datasets (SCES and new SINDI datasets) show our method outperforms other state-of-the-art algorithms.
I. INTRODUCTION
Glaucoma screening is important because vision loss is irreversible, yet existing clinical and measurement-based approaches are limited. The paper proposes DENet, which combines disc-aware, multi-level, and multi-module visual information for direct fundus-image screening.
- Motivation: Glaucoma causes irreversible vision loss, making early screening important for treatment and preserving vision.Many patients are unaware of their condition, while early disease may lack visual symptoms.
- Motivation: Manual optic nerve head assessment is time-consuming and costly, motivating automatic screening methods.Optic nerve head assessment is widely performed by trained glaucoma specialists.
- Limitations of Existing Methods: Measurement-based methods rely heavily on segmentation accuracy and can be affected by pathological regions and low image contrast.They may also capture image properties unrelated to clinically relevant visual cues.
- Proposed Approach: DENet emphasizes optic-disc context while incorporating global and local image information and multiple network modules.The design includes a global image stream, segmentation-guided network, local disc stream, and disc polar transformation stream.
- Contributions: The paper contributes an ensemble network, a segmentation-guided detection network, and the SINDI dataset containing 5,783 eye images.SINDI includes 113 glaucomatous and 5,670 normal eyes.
II. DISC-AWARE ENSEMBLE NETWORK
DENet represents fundus images at global and local optic-disc levels. Its local representation preserves fine disc details, while the architecture combines the resulting streams for screening.
- Multi-level Representation: DENet uses global fundus information for coarse structure representation and local disc information for fine optic-disc representation.The two levels provide complementary spatial contexts.
- Four-stream Architecture: The architecture combines four streams: global image, segmentation-guided, local disc region, and disc polar transformation.The fused output provides the final glaucoma screening result.
A. Global Fundus Image Level
The global fundus image level uses two complementary streams: a ResNet-50 classifier and a U-Net-based segmentation-guided network that detects the optic disc and supports glaucoma screening.
- Global image stream: The ResNet-50 stream learns global fundus representations directly and produces glaucoma screening output through global max pooling and a fully connected layer.Its input is resized to 224×224, and ImageNet-pretrained parameters initialize the network.
- Segmentation-guided network: The segmentation-guided stream localizes the optic disc and generates a screening result from disc-segmentation representations.Its architecture follows U-Net, with encoder and decoder paths connected by feature-map concatenation.
- Training objectives: The global networks use binary cross entropy for glaucoma detection and Dice loss for optic disc segmentation.The two losses are integrated into backpropagation using stochastic gradient descent.
- Training strategy: The segmentation-guided model is trained in two phases: disc segmentation first, followed by classification with CNN parameters frozen.The authors state that disc-segmentation representation adds diversity and that disc-detection accuracy is important for the follow-up stream.
B. Optic Disc Region Level
The optic disc region level adds higher-resolution local representations and a polar-transformed disc stream to capture detailed, more balanced cup-region information.
- Local disc region: The local optic disc region is cropped from the segmentation-guided result to preserve higher-resolution detail for fine representation learning.Two local streams operate on this region, including a standard ResNet-based classifier.
- Polar transformation: Polar transformation maps points from Cartesian coordinates to polar coordinates using the disc center, polar radius, and polar angle.The transformation is illustrated as a correspondence between p(u, v) and p′(θ, r).
- Polar transformation: The polar image dimensions are set by the polar radius R and angular discretization 2π/s, where s is the stride.These parameters determine the transformed representation's height and width.
- Polar transformation: Because the disc and cup are near-concentric, polar transformation enlarges the cup region by interpolation and increases its displayed detail.The transformation also preserves equivariance to data augmentation through pixel-wise mapping.
A. Dataset and Evaluation Criteria
The experiments use ORIGA for training and SCES and SINDI for testing, with sensitivity, specificity, and balanced accuracy as evaluation criteria.
- Datasets: The study uses ORIGA, SCES, and SINDI glaucoma screening datasets.ORIGA contains 650 images, SCES contains 1676 images, and SINDI contains 5783 eye images.
- Dataset split: Only ORIGA provides manually labelled optic disc boundaries, so all 650 ORIGA images train disc segmentation and glaucoma screening, while SCES and SINDI serve as test sets.SINDI includes 113 glaucomatous and 5670 normal eyes.
- Evaluation criteria: Sensitivity, specificity, and balanced accuracy are computed from true-positive, true-negative, false-positive, and false-negative counts.Balanced accuracy is defined from sensitivity and specificity.
- Evaluation criteria: The reported operating threshold is selected by tuning diagnostic thresholds and choosing the one with the highest balanced accuracy.This threshold is then used for final performance reporting.
B. Implementation Details
The implementation trains the four streams separately, uses augmentation and pretrained initialization, and evaluates performance with dataset-level comparison tables and ROC curves.
- Training implementation: The four DENet streams are trained separately to accommodate stream dependencies, new-stream integration, and stream-specific datasets or configurations.The local streams depend on disc detections from the global image stream.
- Performance evaluation: Table II reports performance comparisons among different methods on the evaluated datasets.The supplied passage identifies the table's comparison purpose but provides no individual metric values.
- Performance evaluation: The SCES and SINDI ROC figures show AUC scores and mark the operating point with the highest balanced accuracy.Figure 5 covers SCES, while Figure 6 covers SINDI.
- Training implementation: The ResNet-50 stream is initialized with ImageNet-pretrained parameters and fine-tuned across all layers.This initialization supports adaptation of the backbone during training.
C. Glaucoma Screening
DENet outperforms clinical, handcrafted-feature, superpixel, and deep segmentation baselines on SCES and SINDI, while screening directly from visual features and localizing the disc region.
- DENet is compared with Wavelet, Gabor, GRI, Superpixel, DeepCDR, and IOP as glaucoma-screening baselines.
- 0.9183 AUC on SCES and 0.8173 AUC on SINDI are the best reported performances for DENet, about 2% above DeepCDR.These results indicate that visual features can support glaucoma screening without explicitly extracting clinical parameters such as CDR.
- 0.8265 AUC on SCES and 0.7712 AUC on SINDI are achieved by Superpixel, outperforming the IOP baseline.IOP achieves 0.66 AUC on SCES and 0.6233 AUC on SINDI.
- 0.898 AUC on SCES is reported for another deep system using CNN features from the disc region, whereas DENet also localizes the disc from the whole fundus image.
1) Stream Analysis:
Local optic-disc streams outperform global-image streams, polar transformation improves local representation, and averaging all four streams gives the best scores on both datasets.
- Stream Analysis: Local disc-region streams outperform global-image streams for every DENet stream.The authors relate this pattern to clinical symptoms being concentrated near the optic disc.
- Stream Analysis: Polar transformation improves local-disc representation and outperforms the original disc region on both datasets.
- Stream Analysis: Three-stream combinations perform better than two-stream combinations, while all four streams obtain the best scores on SCES and SINDI.
- Ensemble Analysis: Average, max, and multiple ensemble operations significantly improve performance, and even min outperforms a single stream.
- Ensemble Analysis: Averaging is selected for DENet because it performs best on SCES and is simple and efficient without requiring prior information.
- Running Time: 0.5 seconds is required to produce one final test result, including disc localization and glaucoma screening.This is faster than the cited superpixel method at 10 seconds and DeepCDR at 1 second.
4) High-Sensitivity Performance:
At a sensitivity of 0.95, DENet achieves the highest reported performances on both glaucoma datasets, supporting its evaluation under a high-sensitivity screening requirement.
- High-Sensitivity Performance: 0.95 Sensitivity is used to compare method specificities because real-world screening should avoid missing glaucoma cases.
- High-Sensitivity Performance: DENet achieves the highest performances on both datasets at 0.95 Sensitivity.
- Conclusion: DENet integrates four deep streams whose multiple levels and modules incorporate hierarchical representations while retaining optic-disc context.