Joint Optic Disc and Cup Segmentation Based on Multi-Label Deep Network and Polar Transformation

Huazhu Fu, Jun Cheng, Yanwu Xu, Damon Wing Kee Wong, Jiang Liu, Xiaochun Cao

TL;DR

This paper introduces M-Net, a deep learning architecture designed for the joint segmentation of the optic disc (OD) and optic cup (OC) in retinal fundus images. By combining a multi-scale U-Net-style network with a novel polar transformation of the input images, the authors achieve state-of-the-art performance in calculating the Cup-to-Disc Ratio (CDR), a critical metric for early glaucoma screening.

What Problem Does This Paper Address?

Glaucoma is a leading cause of irreversible blindness, and early detection is vital. Clinicians primarily screen for it by assessing the Cup-to-Disc Ratio (CDR)—the ratio of the vertical diameter of the optic cup to the optic disc.

Automating this assessment is challenging for several reasons:

  1. Low Contrast: The boundary of the optic cup is often fuzzy and difficult to distinguish from the surrounding neuroretinal rim.
  2. Pathological Interference: Features like peripapillary atrophy (PPA) can confuse traditional segmentation algorithms.
  3. Suboptimal Modeling: Most existing methods segment the disc and cup separately, ignoring their inherent spatial relationship (the cup is always inside the disc).
  4. Data Imbalance: In a standard fundus image, the optic cup occupies a very small percentage of pixels, making it difficult for standard neural networks to learn effectively.

Methods

The authors propose a two-pronged approach: a specialized network architecture (M-Net) and a geometric preprocessing step (Polar Transformation).

1. The M-Net Architecture

M-Net is an end-to-end multi-label deep network that builds upon the popular U-Net. Its key components include:

  • Multi-scale Input Layer: It feeds an image pyramid into the encoder, allowing the network to learn features at multiple receptive field sizes simultaneously.
  • U-shape Backbone: A standard encoder-decoder structure with skip connections to preserve high-resolution spatial information.
  • Side-output Layers: These act as "early classifiers" at different stages of the network, providing deep supervision and helping the model converge more effectively.
  • Multi-label Dice Loss: Instead of treating the problem as a multi-class task (where a pixel is either disc or cup), they treat it as multi-label. Since the cup is inside the disc, a "cup" pixel is also a "disc" pixel. They use a Dice coefficient-based loss to handle the heavy class imbalance between the small cup and the large background.

2. Polar Transformation

This is perhaps the paper's most clever contribution. Before feeding the image into M-Net, they transform the region around the optic disc from Cartesian coordinates $(u, v)$ to polar coordinates $(\theta, r)$.

  • Spatial Constraint: In polar space, the circular relationship between the cup and disc becomes a layered, horizontal relationship. This makes it easier for the convolutional kernels to learn the "cup-inside-disc" constraint.
  • Data Augmentation: Small shifts in the disc center during transformation act as a natural form of data augmentation.
  • Balancing Proportions: The transformation effectively "stretches" the central cup region, increasing its pixel representation from ~4% in Cartesian crops to ~23% in polar crops, significantly reducing the class imbalance problem.

Datasets

The model was evaluated on two major datasets:

  • ORIGA: 650 fundus images with manual ground truth boundaries. It was split into 325 for training and 325 for testing.
  • SCES (Singapore Chinese Eye Study): 1,676 images used primarily to test the model's generalization capability for glaucoma screening.

Key Insights and Ideas

  • Joint is Better than Separate: By segmenting the OD and OC together in a multi-label framework, the network learns the mutual dependencies between the two structures, leading to more anatomically consistent results.
  • Geometry Matters: The polar transformation isn't just a "trick"; it fundamentally changes the optimization landscape by balancing the classes and simplifying the boundary geometry into linear layers.
  • Deep Supervision Works: The inclusion of side-outputs (L1 through L4 in the architecture) helps the network learn discriminative features at every level of the hierarchy, which is particularly useful when training data is relatively scarce (only 325 training images).

Key Results

  • Segmentation Accuracy: On the ORIGA dataset, M-Net with Polar Transformation achieved an overlapping error ($E$) of 0.071 for the disc and 0.230 for the cup, outperforming traditional methods like ASM and Superpixel classification.
  • Glaucoma Screening: The system achieved an Area Under the Curve (AUC) of 0.8508 on ORIGA and 0.8998 on SCES for glaucoma screening based on the calculated CDR.
  • Efficiency: Once trained, the system takes only 0.5 seconds to process an image on a GPU, making it highly suitable for large-scale population screening.

Limitations and Future Work

  • Failure Modes: The authors note that the system can still fail on images with extreme blur or very low contrast where even the disc boundary is nearly invisible to the human eye.
  • Preprocessing Dependency: The polar transformation relies on an initial disc center localization. While the paper uses an existing automated method for this, any significant error in center detection could propagate through the pipeline.
  • Future Directions: The authors suggest that more powerful backbone networks or additional image enhancement pre-processing could further improve performance in "failed" cases.

Summary updated

Abstract

Glaucoma is a chronic eye disease that leads to irreversible vision loss. The cup to disc ratio (CDR) plays an important role in the screening and diagnosis of glaucoma. Thus, the accurate and automatic segmentation of optic disc (OD) and optic cup (OC) from fundus images is a fundamental task. Most existing methods segment them separately, and rely on hand-crafted visual feature from fundus images. In this paper, we propose a deep learning architecture, named M-Net, which solves the OD and OC segmentation jointly in a one-stage multi-label system. The proposed M-Net mainly consists of multi-scale input layer, U-shape convolutional network, side-output layer, and multi-label loss function. The multi-scale input layer constructs an image pyramid to achieve multiple level receptive field sizes. The U-shape convolutional network is employed as the main body network structure to learn the rich hierarchical representation, while the side-output layer acts as an early classifier that produces a companion local prediction map for different scale layers. Finally, a multi-label loss function is proposed to generate the final segmentation map. For improving the segmentation performance further, we also introduce the polar transformation, which provides the representation of the original image in the polar coordinate system. The experiments show that our M-Net system achieves state-of-the-art OD and OC segmentation result on ORIGA data set. Simultaneously, the proposed method also obtains the satisfactory glaucoma screening performances with calculated CDR value on both ORIGA and SCES datasets.
BibTeX
@article{fu2018joint,
  author = {Huazhu Fu and Jun Cheng and Yanwu Xu and Damon Wing Kee Wong and Jiang Liu and Xiaochun Cao},
  title = {{Joint Optic Disc and Cup Segmentation Based on Multi-Label Deep Network and Polar Transformation}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2018},
  url = {https://openreview.net/forum?id=6DBtgVyAded},
}

Loading related content...

Figure