Adaptive vision transformer for enhanced perception in visual prostheses.

Area of Science:

• Biomedical Engineering
• Computer Science
• Neuroscience

Background:

• Prosthetic vision offers hope for visually impaired individuals, but current technology provides suboptimal visual perceptions.
• Physical constraints and patient anatomy cause phosphene distortions, limiting the functionality of visual prosthetics.
• Personalized approaches integrating deep learning show promise for improving prosthetic vision.

Purpose of the Study:

• To develop a novel deep learning model for prosthetic vision that reduces phosphene distortions.
• To enhance the usefulness of perceptions generated by visual prosthetics for daily tasks.
• To establish a new benchmark for prosthetic vision systems using complex image datasets.

Main Methods:

• A novel neural network architecture incorporating a vision transformer was developed.
• The model analyzes visual input and patient-specific parameters to optimize stimulation.
• Geometric transformations were implemented to correct implant field-of-view distortions.

Main Results:

• The proposed model demonstrated superior performance compared to baseline methods on the MNIST dataset.
• The model achieved new benchmarks for complex image datasets like ImageNet, CIFAR-10, and COCO.
• Optimized stimulation parameters and geometric corrections significantly reduced perceptual distortions.

Conclusions:

• The developed deep learning model significantly improves prosthetic vision by reducing distortions.
• This personalized approach enhances the quality of visual perception for visually impaired users.
• The findings pave the way for more effective and functional visual prosthetic systems.