Dynamic analysis of iris changes and a deep learning system for automated angle-closure classification based on...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Primary angle-closure disease (PACD) is a significant cause of vision loss.
• Dynamic iris changes during pupillary response are implicated in PACD.
• Anterior segment optical coherence tomography (AS-OCT) provides detailed imaging of ocular structures.

Purpose of the Study:

• To investigate the association between dynamic iris changes and PACD using AS-OCT videos.
• To develop and validate an automated deep learning system for PACD screening.
• To analyze dynamic clinical parameters of iris motion in relation to PACD.

Main Methods:

• Analysis of 369 AS-OCT videos (19,940 frames) from 159 PACD subjects and 210 controls.
• Correlation of iris constriction dynamics (pupil diameter) with PACD under expert ophthalmologist guidance.
• Development of a temporal network for discriminative feature learning from videos, with fivefold cross-validation.

Main Results:

• Angle-closure eyes exhibited significantly lower mean velocity of pupil constriction (VPC) and acceleration of pupil constriction (APC) compared to normal eyes.
• The automated system achieved high areas under the curve (AUCs) for angle-closure classification, with aligned AS-OCT videos yielding the best performance (up to 0.919).

Conclusions:

• Iris in angle-closure eyes demonstrates reduced stretching in response to illumination compared to normal eyes.
• Dynamic iris motion features are valuable for improving the accuracy of angle-closure classification.
• The developed deep learning system shows potential for effective automated screening of PACD.