Decoding neural activity of odor perception with dimensionality reduction and machine learning on fMRI data

Neuroimaging methods have been widely used to explore the functional neuroanatomy of olfaction. Understanding how odors are perceived by humans and their implications for our behavior and choices goes beyond scientific interest and also offers interesting commercial opportunities. The objective of this work is to verify whether a large olfaction-related dataset made of 5718 brain activation maps is free from noise and whether relevant information about odor perception can be extracted from it. Two research directions have been conducted. On one hand, a classifier capable of predicting, based on a given activation map, whether the perceived odor is pleasant or unpleasant has been implemented. We found that a pipeline, consisting of dimensionality reduction, inverse linear system, and supervised learning steps, enabled effectively the training and testing of classifiers on the dataset. Ultimately, the Random Forest and SVM classifier emerged as the most accurate methods. On the other hand, a research study was conducted to identify brain activation areas for each olfactory family using unsupervised learning. The results of two methods were compared to manual clustering for a specific family before determining the most effective approach to be generalized across all 13 olfactory families. We found that the coordinate-based approach, although less conventional, can be combined with agglomerative clustering to yield better results than more traditional methods in identifying activation zones for each olfactory family.