Human Iris Image Analysis for the Classification of Fuchs’ Crypts and Peripupillary Rings

Hao Wang, Shiaofen Fang, Frankie Wilke, Mats Larsson, Susan WalshLecture Notes in Networks and Systems 2022 · DOI · OpenAlex

Introduction

The human iris is a complex biological structure containing unique patterns such as Fuchs’ crypts (oval-shaped pits in the iris stroma) and peripupillary pigmented rings (circular melanin deposits around the pupil). These features are highly heritable and can serve as biomarkers for certain diseases. However, genetic research into these patterns currently relies on "manual phenotyping," where human experts grade thousands of images. This process is labor-intensive, subjective, and prone to inter-rater inconsistency. To enable large-scale genome-wide association studies (GWAS), an objective, automated method for quantifying these biological traits is required.

Methodology

Researchers developed two distinct pipelines to address the different nature of these iris features, combining deep learning for structural detection with classical statistical methods for pigmentation analysis.

1. Fuchs’ Crypt Frequency Analysis

Structural features like crypts vary significantly in size and clarity. The classification process involves a two-stage model:

  • Detection & Segmentation: A Mask R-CNN model (Detectron2) identifies and outlines individual crypts, distinguishing between three levels of clarity.
  • Classification: The counts of detected crypts are fed into a Linear Support Vector Machine (SVM) to determine the final frequency category (1, 2, or 3) for the entire iris.

To handle the limited size of medical datasets, the team utilized elastic transformations. This augmentation technique mimics natural structural variations in iris tissue, improving the model's ability to generalize across different biological presentations.

Examples of Fuchs’ crypts (labeled 1-3) across various iris colors and textures.

2. Peripupillary Pigmented Ring Classification

Pigmentation features are better handled through color-space analysis rather than object detection. The iris is segmented into concentric rings, and images are converted from RGB to the HSV (Hue-Saturation-Value) color space. This allows the model to isolate pigmentation (Hue and Saturation) from lighting intensity (Value). A K-Nearest Neighbors (KNN) classifier then analyzes 2D histograms of these color values to categorize the rings based on their "color fingerprints."

Iris images demonstrating different grades of peripupillary pigmented rings.

Results and Performance

The automated system demonstrated high reliability, matching the performance of trained human experts across both phenotypes.

Crypt Frequency: The Mask R-CNN + SVM pipeline achieved 80.0% accuracy. Precision and recall were highest for Category 1 (no marked crypts) and Category 3 (numerous crypts), while Category 2 (the intermediate state) proved the most challenging for both humans and AI. Notably, the model's performance is remarkably close to the 84.4% agreement rate between two independent human experts, suggesting the model has reached a performance ceiling dictated by the inherent ambiguity of these biological features.

Pigmented Rings: The histogram-based KNN approach reached 86.6% accuracy. The researchers tested seven different distance metrics for the KNN classifier, finding that color-space fingerprints provided a robust way to distinguish varying levels of melanin deposit regardless of overall iris color.

The study highlights that different biological features require specialized machine learning architectures: structural features benefit from object detection and segmentation, while pigmentation is more effectively analyzed via color-space histograms. Furthermore, the use of elastic deformation proved to be a superior augmentation strategy for medical imaging of the iris, as it more accurately simulates the biological variability of tissue hypoplasia compared to standard rotations or flips.

Conclusion and Future Work

While the current models achieve human-level consistency, challenges remain. Specular reflections from external light sources can occasionally obscure pigmented rings, leading to misclassification. While the current dataset of 1,500 high-resolution images provided a strong foundation, expanding the library could further improve the precision of the crypt detection stage. By providing an objective and scalable method for iris pattern analysis, this research facilitates more extensive genetic studies into the biological significance of these ocular traits.

Summary updated

BibTeX
@inproceedings{wang2022human,
  author = {Hao Wang and Shiaofen Fang and Frankie Wilke and Mats Larsson and Susan Walsh},
  title = {{Human Iris Image Analysis for the Classification of Fuchs’ Crypts and Peripupillary Rings}},
  booktitle = {Lecture Notes in Networks and Systems},
  year = {2022},
  doi = {10.1007/978-3-031-16078-3\_55},
}
Figure