Abstract
Glaucoma is one of the three leading causes of blindness in the world and is predicted to affect around 80 million people by 2020. The optic cup (OC) to optic disc (OD) ratio (CDR) in fundus images plays a pivotal role in the screening and diagnosis of glaucoma. Existing methods usually crop the optic disc region first, and subsequently perform segmentation in this region. However, these approaches come up with high complexities due to the separate operations. To remedy this issue, we propose a Region Focus Network (RF-Net) that innovatively integrates detection and multi-class segmentation into a unified architecture for end-to-end joint optic disc and cup segmentation with global optimization. The key idea of our method is designing a novel multi-class mask branch which generates a high-quality segmentation in the detected region for both disc and cup. To bridge the connection between the backbone and multi-class mask branch, a Fusion Feature Pooling (FFP) structure is presented to extract features from each level of the pyramid network and fuse them into a final feature representation for segmentation. Extensive experimental results on the REFUGE-2018 challenge dataset and the Drishti-GS dataset show that the proposed method achieves the best performance, compared with competitive approaches reported in the literature and the official leaderboard.
Introduction
Glaucoma is a leading cause of irreversible blindness worldwide. Clinicians diagnose it primarily by calculating the Cup-to-Disc Ratio (CDR)—the ratio between the vertical diameter of the optic cup (OC) and the optic disc (OD).
Automating this measurement is challenging because the OD and OC occupy only about 10% of a typical fundus image, and the boundary of the optic cup is often extremely faint (low contrast). Most existing methods use a two-step "crop-then-segment" pipeline. However, these separate stages increase complexity and rely heavily on the initial cropping accuracy; if the crop is slightly off, the segmentation fails. The authors aim to replace this disjointed process with a single, globally optimized network called RF-Net.
Methods
The proposed RF-Net draws inspiration from the Mask R-CNN architecture but introduces several key innovations tailored for medical imaging:
- Unified Architecture: It uses a Feature Pyramid Network (FPN) with a MobileNet-v1 backbone for efficiency. The network first proposes a region of interest (the optic disc) and then performs multi-class segmentation within that region.
- Fusion Feature Pooling (FFP): Standard pyramid networks usually pick one level of features to perform pooling. FFP instead extracts and fuses features from multiple levels of the pyramid. This ensures the model captures both fine-grained spatial details (from lower levels) and high-level semantic context (from higher levels), which is critical for identifying subtle cup boundaries.
- Multi-class Mask Branch: Unlike standard Mask R-CNN which predicts binary masks (foreground vs. background), RF-Net’s mask branch predicts three classes simultaneously: Optic Disc, Optic Cup, and Background.
- Hybrid Loss Function: The model is trained using a combination of Weighted Focal Loss (to handle the extreme class imbalance) and Dice Loss (to optimize for the overlap and shape similarity of the regions).
Datasets and Results
The model was evaluated on two major benchmarks: REFUGE-2018 (1,200 images) and Drishti-GS (101 images). RF-Net outperformed several competitive models, including U-Net, M-Net, and standard Mask R-CNN:
- Accuracy: On the REFUGE dataset, it achieved a Dice score of 89.79% for the Optic Cup and 96.06% for the Optic Disc, setting a new state-of-the-art at the time of publication.
- Speed: The network reaches an inference speed of 4 FPS on an NVIDIA Tesla P100. While this is not real-time, it is significantly faster than many competing medical segmentation pipelines that require extensive pre- or post-processing.
- Resolution: The authors found that increasing the internal mask resolution from $28 \times 28$ to $56 \times 56$ significantly reduced "serrated" edges and improved boundary smoothness, which is vital for accurate diameter measurements.
Conclusion
By performing detection and segmentation together, RF-Net "focuses" its representative power on the relevant 10% of the image, effectively filtering out noise from non-optic areas. While the contrast between the optic cup and disc remains a fundamental challenge in low-quality images, the "Region Focus" concept offers a robust framework for medical tasks where small structures need to be segmented within large, noisy images, such as identifying microaneurysms in diabetic retinopathy.
Summary updated