[Paper Review] Automated Segmentation of the Optic Disk and Cup using Dual-Stage Fully Convolutional Networks

Lei Bi, Yuyu Guo+4 more|arXiv (Cornell University)|Feb 13, 2019

Retinal Imaging and AnalysisMedicine18 references3 citations

TL;DR

This paper proposes a dual-stage fully convolutional network (DSFCN) for automated optic disk and cup segmentation in retinal fundus images, leveraging iterative refinement using predictions from previous stages to improve boundary accuracy. The method achieves a Dice score of 0.8488 for the optic cup and 0.9441 for the optic disk, with an AUC of 0.9513 for glaucoma detection, outperforming conventional FCN approaches on the REFUGE challenge dataset.

ABSTRACT

Automated segmentation of the optic cup and disk on retinal fundus images is fundamental for the automated detection / analysis of glaucoma. Traditional segmentation approaches depend heavily upon hand-crafted features and a priori knowledge of the user. As such, these methods are difficult to be adapt to the clinical environment. Recently, deep learning methods based on fully convolutional networks (FCNs) have been successful in resolving segmentation problems. However, the reliance on large annotated training data is problematic when dealing with medical images. If a sufficient amount of annotated training data to cover all possible variations is not available, FCNs do not provide accurate segmentation. In addition, FCNs have a large receptive field in the convolutional layers, and hence produce coarse outputs of boundaries. Hence, we propose a new fully automated method that we refer to as a dual-stage fully convolutional networks (DSFCN). Our approach leverages deep residual architectures and FCNs and learns and infers the location of the optic cup and disk in a step-wise manner with fine-grained details. During training, our approach learns from the training data and the estimated results derived from the previous iteration. The ability to learn from the previous iteration optimizes the learning of the optic cup and the disk boundaries. During testing (prediction), DSFCN uses test (input) images and the estimated probability map derived from previous iterations to gradually improve the segmentation accuracy. Our method achieved an average Dice co-efficient of 0.8488 and 0.9441 for optic cup and disk segmentation and an area under curve (AUC) of 0.9513 for glaucoma detection.

Motivation & Objective

• To address the limitations of traditional hand-crafted feature methods in optic disk and cup segmentation.
• To overcome the challenges of limited annotated medical data and coarse boundary predictions in standard fully convolutional networks (FCNs).
• To develop a fully automated, end-to-end method that enhances segmentation accuracy through iterative refinement of predictions.
• To improve boundary delineation of the optic cup and disk for more reliable glaucoma detection.
• To validate the method on the REFUGE challenge dataset, a benchmark for optic disc and cup segmentation.

Proposed method

• The proposed DSFCN uses a dual-stage architecture where the second stage refines predictions from the first stage using both input images and estimated probability maps from prior iterations.
• The network employs deep residual architectures to improve feature learning and gradient flow during training.
• During training, the model learns from both ground-truth annotations and the model's own previous predictions, enabling progressive refinement of boundary details.
• During inference, the test image and the probability map from the previous stage are used as input to the next stage, allowing iterative improvement of segmentation results.
• The method uses fully convolutional layers to maintain spatial resolution and enable dense pixel-wise prediction across the entire image.
• The network is trained end-to-end with a multi-scale loss function to optimize for both global and local segmentation accuracy.

Experimental results

Research questions

• RQ1 Can a dual-stage FCN architecture improve the accuracy of optic cup and disk segmentation compared to single-stage FCNs?
• RQ2 Does iterative refinement using predictions from prior stages enhance boundary delineation in retinal images?
• RQ3 Can the proposed method generalize well to diverse retinal image variations with limited annotated training data?
• RQ4 How does the DSFCN perform on benchmark datasets such as the REFUGE challenge?
• RQ5 To what extent does the method improve glaucoma detection performance through accurate segmentation?

Key findings

• The DSFCN achieved a Dice coefficient of 0.8488 for optic cup segmentation, indicating strong overlap with ground truth.
• The method attained a Dice score of 0.9441 for optic disk segmentation, demonstrating high precision in disk boundary detection.
• The model achieved an area under the curve (AUC) of 0.9513 for glaucoma detection, reflecting strong diagnostic performance.
• The iterative refinement mechanism significantly improved boundary accuracy compared to standard FCN baselines.
• The method demonstrated robustness on the REFUGE challenge dataset, which includes diverse and challenging retinal images.
• The use of predicted probability maps from previous stages during inference enabled progressive enhancement of segmentation quality.