AGE challenge: Angle Closure Glaucoma Evaluation in Anterior Segment Optical Coherence Tomography

Huazhu Fu, Fei Li, Xu Sun, Xingxing Cao, Jingan Liao, José Ignacio Orlando, Xing Tao, Yuexiang Li, Shihao Zhang, Mingkui Tan, Chenglang Yuan, Cheng Bian, Ruitao Xie, Jiongcheng Li, Xiaomeng Li, Jing Wang, Le Geng, Panming Li, Yanwu Xu

Introduction

Glaucoma remains a leading cause of irreversible blindness worldwide, with Primary Angle Closure Glaucoma (PACG) representing its most aggressive form. Diagnosis typically hinges on evaluating the anatomical "angle" where the iris meets the cornea. In a closed angle, fluid drainage is obstructed, leading to increased intraocular pressure and subsequent optic nerve damage.

While Anterior Segment Optical Coherence Tomography (AS-OCT) offers high-resolution cross-sectional imaging of this region, clinical assessment is currently hindered by subjectivity and the time-consuming nature of manual landmark identification. Specifically, identifying the Scleral Spur (SS)—a critical anatomical landmark for calculating drainage metrics—requires expert intervention. The AGE Challenge was established to provide the first large-scale, standardized public benchmark to automate these diagnostic tasks using artificial intelligence.

Methodology and Tasks

The AGE Challenge structured the problem into two primary computational tasks designed to mirror the clinical workflow:

  1. Scleral Spur Localization: Predicting the precise $(x, y)$ coordinates of the scleral spur.
  2. Angle Closure Classification: Categorizing images into "open" or "closed" angle states.
The AGE Challenge pipeline: AS-OCT images are processed to localize the scleral spur, which then informs the classification of the angle as open or closed.

Advanced Deep Learning Strategies

Participating teams utilized several state-of-the-art deep learning techniques to achieve high precision:

  • Heat Map Regression: Rather than direct coordinate regression, top-performing models transformed landmarks into Gaussian "heat maps." This approach allows the network to learn the probability distribution (or "glow") of a landmark, making it significantly more robust to image noise.
  • Coarse-to-Fine Frameworks: To handle the high resolution of OCT scans (often exceeding 2000 pixels), teams employed two-stage architectures. A "coarse" network first identifies the general region of interest, after which a "fine" network zooms in to pinpoint the landmark with micron-level accuracy.
  • Multi-Scale Ensembles: Architectures such as ResNet, EfficientNet, and SE-Net were often combined. By analyzing images at multiple scales, these ensembles captured both broad global structures and minute local anatomical details.
A visualization of a Gaussian heat map used for landmark detection, illustrating how the model identifies the 'glow' of the scleral spur.

The AGE Dataset

The challenge released the AGE Dataset, the largest public collection of its kind, comprising 4,800 AS-OCT images from 199 patients. To ensure a high-quality "gold standard," every image was annotated by four ophthalmologists, with final verification by a senior specialist. Clinical ground truth for classification was established via gonioscopy, the current manual diagnostic reference standard. The dataset was split equally into training, validation, and testing sets (1,600 images each).

Key Insights and Results

The challenge yielded several critical findings for the field of ophthalmic AI:

  • Superiority of Heat Maps: The transition from coordinate regression to pixel-wise probability (heat maps) proved essential for faster convergence and higher localization accuracy.
  • AI vs. Manual Metrics: AI models frequently outperformed classification based solely on traditional clinical measurements like Angle Opening Distance (AOD). This suggests that deep learning models identify subtle morphological patterns and visual textures that manual measurements overlook.
  • Exceptional Performance: The top-performing team, RedScarf, achieved a localization error of only 9.4 pixels (~10µm)—roughly one-seventh the thickness of a human hair. In the classification task, the leading teams achieved a perfect 100% accuracy (AUC of 1.0) on the onsite test set.

Limitations and Future Directions

While the results are promising, the authors identified key areas for future development. The current dataset is limited to images from a single device type (Casia SS-1000), necessitating future benchmarks that include diverse hardware to prevent manufacturer bias. Additionally, real-world screening will require models to handle more "borderline" cases and complex anatomical variations like plateau iris. The ultimate goal remains the development of fully automated, "one-click" systems capable of providing both diagnosis and comprehensive anatomical measurements in a clinical setting.

Summary updated

BibTeX
@article{fu2020age,
  author = {Huazhu Fu and Fei Li and Xu Sun and Xingxing Cao and Jingan Liao and José Ignacio Orlando and Xing Tao and Yuexiang Li and Shihao Zhang and Mingkui Tan and Chenglang Yuan and Cheng Bian and Ruitao Xie and Jiongcheng Li and Xiaomeng Li and Jing Wang and Le Geng and Panming Li and Yanwu Xu},
  title = {{AGE challenge: Angle Closure Glaucoma Evaluation in Anterior Segment Optical Coherence Tomography}},
  journal = {Medical Image Analysis},
  year = {2020},
  doi = {10.1016/j.media.2020.101798},
}
Figure