Skip to content

Home

Jump to bottom
Bliss Chapman edited this page Mar 11, 2019 · 2 revisions

Idea

  • Can we generate a distribution of possible images of your brain in 5/10 years from a brain image of you now?
  • Could we use such a model to capture and characterize neuroanatomical variability in the aging process?
  • How do our methods perform on various image modalities designed to capture different dimensions of neuroanatomical structure/function? (fMRI, DTI, PET, etc)

Anticipated Challenges

  1. We do not have many direct mappings between images of young brains and images of old brains.
    1. IDEA: Can we build generative models at each age point (likely fully convolutional VAEs) and find a way to bridge the gap between the generative models? See Latent Translation paper below.
    2. What semantic attributes do we KNOW to be aligned between old and young brains? Can we use these as softer constraints on the learned latent space in combination with aligning direct examples?
  2. Brain imagery is methodologically messy to work with and usually datasets are not setup for deep learning out of the box.
  3. Evaluation is tricky. How do we know what variability our model is capturing? What metrics do we use to validate generative model performance? How do we know the model is showing me an interesting distribution for MY brain image and not just the general distribution of old brain images?
    1. Surrogate metrics like brain age (see papers below…seems straightforward)
    2. Analysis of regions of interest between populations -> mean and std of volume between real data and fake data in general and then specific to subsets of the population with known aging dynamics represented in the dataset
    3. Analysis of how well-understood factors like race/sleep per night (no idea if these are good examples of factors… will need to discuss, do some reading, and look at whats in the data) are represented in our model compared to the general population? For example, in the real data if attribute X of a person increases the likelihood that they develop Alzheimer’s compared to the general population, is that also true in the model generated distribution of possible future brains.

Datasets

https://brain-development.org/ixi-dataset/ https://www.sciencedirect.com/science/article/pii/S1053811916000331

Papers

Latent Translation: Crossing Modalities by Bridging Generative Models https://arxiv.org/pdf/1902.08261v1.pdf

Brain age predicts mortality https://www.nature.com/articles/mp201762

Brain Age Prediction Based on Resting-State Functional Connectivity Patterns Using Convolutional Neural Networks https://arxiv.org/abs/1801.04013

Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker https://arxiv.org/pdf/1612.02572.pdf

E. C. Robinson, A. Hammers, A. Ericsson, A. D. Edwards, D. Rueckert. Identifying population differences in whole-brain structural networks: A machine learning approach. NeuroImage, 50(3): 910-919, 2010.

Clone this wiki locally