PET PIPELINE AUTOMATION AND STRUCTURAL MRI EXPLORATION
PhD project: cholinergic PET imaging of people with REM sleep behaviour disorder (prodromal stage of Parkinson's / Lewy body dementia). Taking place at the Montreal Neurological Institute. Part of the ongoing Montreal RBD cohort study that tracks disease progression across time.
My lab's particular focus is a new PET radiotracer (FEOBV) that shows great promise in quantifying brain cholinergic systems.
I want to attempt to see what can be done to garner more attention to this radiotracer.
At a birds-eye-view, I have two intentions:
- Grasp the current pipeline and replicate it on existing data.
- Learn about software and statistical techniques that I can implement in the lab and bring to the longitudinal study.
PROJECT 1: LEARN CURRENT PREPROCESSING PIPELINE
This part I'm less excited about. Why?
- I'm at BrainHack because I like analyzing data, less so preprocessing it. (But, I need to get over this hump.)
- The instructions I've been passed down are a little rough.
Goals
- Understand my lab's preprocessing pipeline.
- Successfully replicate it.
- Become comfortable working with minctools, CBrain, CIVET
- If time allows, run data through ANTS, dartel
Deliverable 1
Write a script that automates the lab's pipeline steps.
PROJECT 2: STRUCTURAL MRI ANALYSIS - ML CLASSIFIER.
I plan to then switch gears to machine learning on structural MRI data of individuals with Alzheimer's disease (AD). I would like to create a classifier to determine whether scan comes from an individual with AD or a healthy control.
Goals
- Download a subset of the PREVENT-AD or OASIS datasets
- Feature extraction
- Basic morphology: cortical thickness, brain volume, etc.
- Put these features into workable matrices using numpy and pandas
- Dimensionality reduction via PCA
- Enter remaining features into model
- Model type: SVM, random forest?
- Learn about cross-validation techniques thanks to break-out session
- Nilearn to implement the above
- Matplotlib plots along the way to visualize correlation matrices, model error, ROC curves, etc.
If extra time allows, delve into:
- Longitudinal data
- Model to predict onset of disease conversion
- Naive bayes classifiers?
- Survival analysis?
- Cox hazards functions?
Deliverable 2
A Jupyter notebook walking through each of the above steps, with plots saved inline.
Project medium
I will present these jupyter notebooks in a lab meeting and provide all documents on Github.
During my PhD, I hope to apply these ML techniques to my PET imaging, and to get involved in the cohort study.