Deep learning networks reflect cytoarchitectonic features used in brain mapping.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• Cortical cytoarchitecture, the distribution of neurons, defines human brain structural maps.
• Current mapping methods are limited in throughput; deep learning offers a potential solution.
• The extent to which deep learning models adhere to cytoarchitectural principles remains unclear.

Purpose of the Study:

• To investigate if deep convolutional neural networks (CNNs) trained for cytoarchitectural brain mapping internally reflect traditional cytoarchitectural features.
• To compare the internal structure of these CNNs with the grey level index (GLI) profile approach.

Main Methods:

• A 10-block deep CNN architecture was trained to segment the primary and secondary visual cortex.
• Filter activations within the CNNs were analyzed to identify resemblances to traditional cytoarchitectural features.
• CNN features were compared against the GLI profile approach.

Main Results:

• The CNNs demonstrated resemblances to cellular, laminar, and cortical area-specific cytoarchitectural features.
• Learned filter activations reflected the distinct cytoarchitecture of segmented visual cortical areas, particularly their laminar organization.
• CNN performance compared favorably to statistical criteria of the GLI profile approach.

Conclusions:

• Deep convolutional neural networks successfully incorporate relevant cytoarchitectural features.
• These findings support the use of deep learning models as a valid tool for high-throughput cytoarchitectural mapping.
• CNNs offer a promising avenue for advancing structural brain mapping techniques.