Neuron Counting for Macaque Mesoscopic Brain Connectivity Research.

Area of Science:

• Neuroscience
• Computational Biology
• Data Science

Background:

• Precise quantification and localization of tracer-labeled neurons are crucial for understanding brain connectivity and building a macaque brain connectome atlas.
• Existing methods face challenges in dataset development and accuracy.

Purpose of the Study:

• To introduce the Macaque Fluorescently Labeled Neurons (MFN) dataset for neuron quantification and localization.
• To develop and validate an advanced AI model for accurate cell counting and localization in macaque brains.

Main Methods:

• Created the MFN dataset from retrograde tracing in three rhesus macaques, including 1,600 images and 33,411 neuron annotations.
• Developed a Dense Convolutional Attention U-Net (DAUNet) model integrating Dense Convolutional blocks and a multi-scale attention module.
• Validated DAUNet on the MFN dataset and four additional public datasets.

Main Results:

• DAUNet achieved a Mean Absolute Error of 0.97 for cell counting and an F1-score of 96.29% for cell localization on the MFN dataset.
• The model outperformed several benchmark models in cell counting and localization tasks.
• DAUNet demonstrated robust generalization capabilities across multiple datasets.

Conclusions:

• The MFN dataset and DAUNet model provide valuable resources for advancing mesoscopic brain connectivity research in macaques.
• The developed model successfully quantified labeled neurons and mapped connectivity patterns in a macaque brain.
• This work facilitates the construction of a macaque brain connectome atlas.