Enhancing Ishihara and educational images using machine learning: toward accessible learning for colorblind...

Area of Science:

• Ophthalmology and Computer Vision
• Medical Imaging and Machine Learning

Background:

• Color Vision Deficiency (CVD) affects over 300 million globally, with protanopia and deuteranopia causing red-green confusion.
• Existing diagnostic tools and educational materials pose challenges for individuals with CVD.

Purpose of the Study:

• To classify normal and simulated CVD Ishihara plate images using machine learning.
• To generate enhanced versions of Ishihara plates and educational diagrams for improved CVD perception.
• To validate the effectiveness of enhanced images through feedback from diagnosed individuals.

Main Methods:

• Simulated protanopia and deuteranopia using sRGB to LMS cone modeling on 1,400 Ishihara plates.
• Image enhancement via a daltonization function, optimizing enhancement strength (α).
• Machine learning models (ResNet-50, EfficientNet-B0, DenseNet-201, PCA, OvA classifiers, random forest, gradient boosting, neural network) for classification and enhancement.

Main Results:

• Optimal enhancement parameters (α=0.54 for deuteranopia, 0.64 for protanopia) achieved significant contrast gains (69.6% and 64.3%) with minimal color distortion (ΔE≈4.9).
• Machine learning models achieved high accuracy (OvA: 99.7%, MLP: 100%).
• User validation showed a significant perceptual improvement, with previously unreadable digits/symbols becoming fully visible and high user ratings (4/5).

Conclusions:

• The OvA technique combined with daltonization effectively enhances Ishihara plates and educational images for red-green CVD.
• This approach offers real-time assistance for improving accessibility of visual information for individuals with CVD.
• The study provides a validated method to improve educational resources and diagnostic tools for a significant portion of the CVD population.