Decoding behavior from global cerebrovascular activity using neural networks

Area of Science:

• Neuroscience
• Machine Learning
• Bioimaging

Background:

• Functional Ultrasound (fUS) offers high-resolution brain vascular activity imaging in behaving animals.
• Current fUS data utilization is limited by a lack of effective visualization and interpretation tools.

Purpose of the Study:

• To develop and validate a machine learning approach for decoding animal behavior from fUS data.
• To demonstrate the potential of neural networks in interpreting complex fUS signals for behavioral classification.

Main Methods:

• Training neural networks on Functional Ultrasound datasets to classify animal behavior.
• Utilizing single 2D fUS images for behavior determination after model training.
• Analyzing network weights in latent space to understand feature importance for behavior classification.

Main Results:

• Neural networks reliably determine animal behavior (e.g., movement, sleep/wake states) from fUS data, even from single images.
• The trained model demonstrated transferability to new recordings and potentially other animals without retraining.
• Analysis of network weights provided insights into the data features critical for behavior classification.

Conclusions:

• Machine learning, specifically neural networks, can effectively decode behavior from fUS brain imaging data.
• This method enhances the utility of fUS data, enabling real-time interpretation of brain activity.
• The approach offers a powerful new tool for neuroscientific research, facilitating behavior-brain activity correlation studies.