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

Abstract

The human iris is one of the most important identifiable features that contain many complex patterns. In this work, we attempted to automatically classify irises with machine learning models based on several different iris patterns in order to assist genetic research related to pigmentation and structural tissue differences within the human iris. Specifically, two main iris patterns that are commonly observed in the general population were analyzed: the Fuchs’ crypts and the peripupillary pigmented ring. A two-stage machine learning model was proposed to classify the iris crypt frequency, in which a Mask R-CNN model was first built to identify the number of crypts of each size level in the iris, followed by a SVM model to determine the final category. Another KNN model, which used the area-refined histogram features, was applied to classify the iris based on the peripupillary pigmented ring. The labels used in the images were generated independently by two trained expert raters. The performance of these models was evaluated on a test set with overall accuracies of the models estimated at 80.0% and 86.6% for crypts and pigmented ring, respectively. These optimized objective models were therefore concordant with the inter-rater reliability scores produced by expert human raters.