Machine and Deep learning approaches for resting state fMRI data of people who were exposed to anesthesia at an early age.
Recent studies showed promising results when state-of-the-art machine learning methods, namely SVM, have been applied to analyse fMRI data. Also, the functional connectivity is another kind of emerging field in medicine which helps analyzing the spatiotemporal relations of brain hemodynamics between different regions of brain. In this part of the study, we aimed to engage these two emerging fields to come up with a robust classifier that is able to distinguish the resting state fMRIs of test subjects, who were exposed anesthesia in their early age, from the resting state fMRIs of control subjects. As a result, we obtained a set of classifiers which predict whether a given fMRI scan belongs to the exposed group or the control group, with a maximum test accuracy value of 75% (p<0.01). ...learn more
Project status: Published/In Market
Overview / Usage
It is important for us to know whether our brain is vulnerable to the anesthetics that are used in surgery or dental operations. It is shown in rodents that anesthesia causes widespread cell death, permanent neuronal deletion, neurocognitive, long-term memory and learning impairments (Lin et. al). Since anesthetics act as potent modulators of excitatory and inhibitory neurotransmissions, it may hinder the proper development of the immature brain (Backeljauw 2015). This study helps us discover if there is a significant change in the functional connectivity on the brain after evaluating time-series data of T2 weighted brain images i.e. fMRI
Methodology / Approach
The MRI data is taken from researchers at Stanford University. First, the raw data was registered on MSDL atlas, then the time series of the fMRI data was extracted. Later, a connectivity matrix consisting of 39 brain regions was formed. By comparing the activations of each pair of these regions via Pearson correlation, we assigned colors to the matrix. Using this matrix as well as time series data, we have performed machine learning approaches ( Naive Bayes, SVM, Random Forest, Gradient Boosting, Decision Tree), dimension reduction techniques i.e t-SNE, recurrent and convolutional neural networks in combination with dense layers in order to classify the exposed data from unexposed.
Technologies Used
Softwares: Python 3.6
Packages: Nilearn, Scikit learn, Tensorflow, Keras,numpy,scipy,matplotlib
We have performed our analysis on a local computer so far. However, in order to test further datasets and try newer models, we are currently updating our code for Intel Xeon Clusters.
Other links
Collaborators
There are no people to show.