Feature selection for the classification of traced neurons

Area of Science:

• Neuroscience
• Computational Biology
• Bioinformatics

Background:

• Numerous computational tools exist for analyzing traced neuron properties.
• Automated neuron classification requires effective feature selection to enhance accuracy.
• Identifying and selecting relevant neuronal descriptors is crucial for reliable classification.

Purpose of the Study:

• To compare the efficacy of various feature selection techniques for neuron classification.
• To identify the most important L-measure features for distinguishing neuron types.
• To improve the quality of automated neuron classification through optimized feature selection.

Main Methods:

• Utilized a dataset of 318 traced neurons (192 GABAergic interneurons, 126 pyramidal cells).
• Extracted features using the L-measure software.
• Applied and evaluated filter, wrapper, embedded, and ensemble feature selection methods, assessing their stability.
• Tested selected feature subsets with classifiers like Random Forest, SVM, and Naïve Bayes.

Main Results:

• Compared filter, embedded, wrapper, and ensemble feature selection methods.
• Evaluated the performance of selected feature subsets across various supervised classification algorithms.
• Identified specific L-measure features (EucDistanceSD, PathDistanceSD, Branch_pathlengthAve, Branch_pathlengthSD, EucDistanceAve) consistently selected.

Conclusions:

• The identified L-measure features are critical for accurate classification of GABAergic interneurons and pyramidal cells.
• Feature selection significantly improves the performance of automated neuron classification systems.
• The study provides evidence for the importance of specific morphological descriptors in neuronal classification.