Deep learning-based pupil model predicts time and spectral dependent light responses.

Area of Science:

• Neuroscience
• Computational Biology
• Vision Science

Background:

• Current pupil models estimate static diameter, failing to capture dynamic responses to light.
• Existing models do not account for temporal receptor weighting or spectral adaptation in the pupil control pathway.

Purpose of the Study:

• To develop a deep learning-based pupil model for predicting pupil diameter over time.
• To reconstruct the pupil's temporal course using photometric, colorimetric, or receptor-based stimulus data.

Main Methods:

• A deep learning approach merging feed-forward neural networks with a biomechanical differential equation.
• Model training and validation using polychromatic and chromatic light spectra at controlled luminance levels.

Main Results:

• The proposed model accurately predicts the temporal pupil light response.
• Achieved a mean absolute error below 0.1 mm for diverse spectral stimuli.
• Demonstrated predictive capability across a range of color temperatures and wavelengths.

Conclusions:

• The deep learning pupil model offers a non-parametric, self-learning approach to pupil response prediction.
• This model advances the generalized description of pupil behavior under varying light conditions.
• Enables accurate temporal prediction of pupil diameter, overcoming limitations of static models.