Contrast Response Function Estimation with Nonparametric Bayesian Active Learning

Area of Science:

• Ophthalmology and Vision Science
• Machine Learning in Healthcare

Background:

• Estimating Contrast Sensitivity Functions (CSFs) is crucial for understanding visual function but is often time-consuming.
• Current clinical methods compromise accuracy for speed or rely on strong assumptions about CSF shape.
• There is a need for more accurate and efficient methods for CSF estimation in research and clinical settings.

Approach:

• Developed the Machine Learning Contrast Response Function (MLCRF) estimator, reframing CSF estimation as a classification problem.
• Utilized machine learning classifiers to quantify the probability of success in contrast detection/discrimination tasks.
• Evaluated the MLCSF estimator using simulated data and human contrast response data, employing Bayesian active learning for stimulus selection.

Key Points:

• The MLCSF estimator, particularly with Bayesian active learning, achieved rapid convergence, requiring only tens of stimuli for accurate estimates.
• MLCSF demonstrated efficiencies comparable to conventional parametric estimators like quickCSF.
• MLCSF consistently achieved higher accuracy than existing methods while allowing adjustable trade-offs between accuracy and efficiency.

Conclusions:

• Machine learning classifiers offer a powerful approach to balance accuracy and efficiency in CSF estimation.
• The MLCSF estimator shows significant potential for improving both research and clinical applications in visual function assessment.
• Further exploration of MLCSF's adjustable accuracy-efficiency balance is warranted.