Machine Learning Approaches to Classify Anatomical Regions in Rodent Brain from High Density Recordings

Area of Science:

• Neuroscience
• Computational Biology
• Medical Technology

Background:

• Intraoperative identification of functional brain regions is crucial for neurosurgery and preclinical research.
• Current methods rely on expert neurophysiologists, highlighting a need for advanced navigation tools.
• Accurate localization of neural signals is essential for precise implant placement and minimizing animal use.

Purpose of the Study:

• To correlate extracellular neural signals with their precise brain region of origin using deep learning.
• To evaluate the efficacy of various neural network architectures for this classification task.
• To establish a foundation for improved in situ brain navigation and reduced animal experimentation.

Main Methods:

• Utilized multisite electrophysiological data sets from the brain.
• Applied and evaluated deep learning models including Bidirectional Long Short-Term Memory (BLSTM), Bidirectional Gated Recurrent Unit (BGRU), Quantile Regression Neural Network (QRNN), and Convolutional Neural Network (CNN).
• Correlated recorded high-density neural signals with their specific brain locations.

Main Results:

• All evaluated neural network architectures successfully distinguished between cortical and thalamic brain regions based on neural signals.
• The Bidirectional Gated Recurrent Unit (BGRU) network achieved the highest accuracy at 88.6%.
• Demonstrated the potential for detailed classification with minimized complex preprocessing.

Conclusions:

• Deep learning models can effectively link neural signals to their brain region of origin.
• The BGRU architecture shows significant promise for real-time intraoperative brain navigation.
• This approach can enhance surgical precision and contribute to the ethical reduction of animal models in preclinical research.