Convolutional neural network classifies visual stimuli from cortical response recorded with wide-field imaging in...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Optic nerve stimulation offers a less invasive alternative for visual neuroprostheses compared to cortical implants.
• Optimizing electrical stimulation parameters requires understanding cortical activation patterns linked to visual stimuli.
• Decoding visual stimuli across large cortical areas is crucial for translational applications in humans.

Purpose of the Study:

• To develop an algorithm for automatically associating cortical activation patterns with specific visual stimuli.
• To enable closed-loop stimulation for visual neuroprostheses by using cortical responses as feedback.
• To create a translational decoding method applicable to future human studies.

Main Methods:

• Utilized wide-field calcium imaging to record primary visual cortex responses in three mice.
• Developed a decoding algorithm based on a convolutional neural network (CNN).
• Trained and fine-tuned the CNN on datasets of visual stimuli and corresponding cortical activity.

Main Results:

• Achieved a best classification accuracy of 75.38% ± 4.77% by pre-training on MNIST and fine-tuning on experimental data.
• Demonstrated generalization capabilities, with cross-dataset accuracies of 64.14% ± 10.81% and 51.53% ± 6.48%.
• Identified optimal CNN training strategies for improved decoding performance.

Conclusions:

• Combining wide-field calcium imaging with CNNs provides a viable method for decoding visual stimuli from cortical responses.
• This approach offers a potential alternative to existing decoding techniques for visual neuroprosthetics.
• The developed algorithm can serve as reliable feedback for future optic nerve stimulation experiments.