A human-in-the-loop deep learning paradigm for synergic visual evaluation in children

Area of Science:

• Ophthalmology
• Pediatrics
• Computer Science
• Artificial Intelligence

Background:

• Visual development in early childhood is critical for overall development.
• Early detection of visual abnormalities in children is essential but challenging.
• Current methods for assessing pediatric visual acuity have limitations.

Purpose of the Study:

• To develop and evaluate a human-in-the-loop deep learning (DL) paradigm for assessing children's visual acuity.
• To improve the accuracy and efficiency of vision examinations in children.
• To integrate software, hardware, and human expertise for enhanced pediatric vision assessment.

Main Methods:

• A novel human-in-the-loop DL paradigm was developed, combining traditional vision examination with DL.
• The system involved two rounds: a human round and a DL round, enabling mutual supervision.
• DL-based object localization and image identification were utilized, incorporating physician experience and captured videos.

Main Results:

• The human-in-the-loop DL paradigm achieved a final accuracy of 75.54% in evaluating children's visual acuity.
• This approach demonstrated superior performance compared to a standalone automatic DL method.
• The paradigm showed potential for more accurate vision evaluation than human assessment alone.

Conclusions:

• The developed human-in-the-loop DL paradigm significantly enhances the accuracy of pediatric visual acuity assessment.
• This integrated approach offers a promising solution for overcoming challenges in examining young children's vision.
• The system facilitates automatic vision evaluation, with potential for application using wearable devices.