Interpretable machine learning approach for neuron-centric analysis of human cortical cytoarchitecture
Andrija Štajduhar1,2, Tomislav Lipić3, Sven Lončarić4
1School of Public Health "Andrija Štampar", School of Medicine, University of Zagreb, 10000, Zagreb, Croatia. andrija.stajduhar@hiim.hr.
Scientific Reports
|April 5, 2023
Summary
This study introduces a novel data science method for quantitative histology, focusing on individual neurons rather than pixels to map human cortex organization. This approach enhances understanding of neuronal phenotypes and cortical layer relationships.
Area of Science:
- Neuroscience
- Histology
- Data Science
Background:
- The human cerebral cortex's complexity is key to its function.
- Current histological studies often focus on image-level analysis, limiting neuron-specific insights.
Purpose of the Study:
- To develop a quantitative histology methodology for neuron-level analysis of the cerebral cortex.
- To create interpretable machine learning models for mapping neuronal phenotypes to cortical layers.
Main Methods:
- Automatic segmentation of neurons across entire histological sections.
- Extraction of extensive features reflecting neuronal phenotype and neighborhood properties.
- Application of an interpretable machine learning pipeline for phenotype-to-layer mapping.
Main Results:
- Development of a principled veridical data science methodology.
- Creation of a unique dataset with expert-annotated cortical layers for validation.
- Demonstration of high interpretability in mapping neuronal features to cortical layers.
Conclusions:
- The neuron-level approach provides a deeper understanding of human cortex organization.
- This methodology can help formulate new scientific hypotheses and manage data uncertainty.
- Offers a shift from pixel-wise image content to neuron-centric investigations.
