Evaluating Deep Human-in-the-Loop Optimization for Retinal Implants Using Sighted Participants.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Computer Vision

Background:

• Personalizing visual prostheses is crucial for restoring sight.
• Human-in-the-loop optimization (HILO) shows promise for adaptive device calibration.
• Previous HILO studies were limited to simulations, lacking human validation.

Purpose of the Study:

• To evaluate the efficacy of HILO with human participants for optimizing visual prosthesis stimulation.
• To compare HILO performance against baseline methods under various conditions.
• To identify differences between human and simulated decision-making in optimization.

Main Methods:

• Sighted participants viewed simulated prosthetic vision.
• Participants iteratively refined a deep stimulus encoder (DSE) using HILO.
• HILO was tested under standard, threshold-misspecified, and out-of-distribution conditions.

Main Results:

• Participants consistently preferred HILO-optimized stimuli over naive encoders and DSE alone.
• HILO demonstrated superior performance across all tested conditions.
• Significant discrepancies were noted between human and simulated optimization decision-making.

Conclusions:

• HILO is a viable and effective method for personalizing visual prostheses.
• Human validation is essential for refining optimization strategies in realistic scenarios.
• This work supports the clinical relevance of HILO for individual visual prosthesis adaptation.