Diagnosing Autism using individual brain regions from rs-fMRIs

One Sentence Summary

• This repository outlines the construction of a deep linear neural network used to classify patients with autism using regional brain activity from functional MRIs gathered at rest (rs-fMRIs); patient data will be gathered from the ABIDE dataset.

Overview

• The task was to use several patients’ rs-fMRIs, convert them individually into time series of 48 features using the Harvard-Oxford Cortical Structural Atlas, and train deep linear neural networks for several pertinent regions to produce a regression result; namely, these results would be a probability of whether each individual patient has Autism Spectrum Disorder (ASD).
• Only the orbitofronal cortex was considered for a small selection of 80 patients from NYU.
• Summary of the performance achieved: Results from training a model on the OFC time series indicate an accuracy of up to 81%, but through many simulations appear no better than random.

Motivation and Background

• I was assigned this project as part of my undergraduate and graduate research under Dr. Pedro Maia at UTA. It sounded like an interesting project to work on, especially since I also have ASD.

• Previous works:

  • Identification of autism spectrum disorder using deep learning and the ABIDE dataset - PubMed
  • Disease prediction using graph convolutional networks: Application to Autism Spectrum Disorder and Alzheimer’s disease - ScienceDirect
  • Enhancing studies of the connectome in autism using the autism brain imaging data exchange II | Scientific Data

• This appears to be a previous work that I plan on extending using several different models.

Summary of Work Done

Data

• Source: downloaded through Nilearn, also available through this link via the Neuroimaging Informatics Tools and Resources Clearinghouse (NITRC).
• Type: NII or NII.GZ files
  • Input: time series for individual brain regions, as outlined in the Preprocessing / Clean-up section
• Size: ~100 GB, although it is ~100 MB per file
• Instances: 80 patients through NYU; 64 used for training , 16 for testing
  • Selection size of about 8 GB

Preprocessing / Clean up

• In order to fully process the data, I had to run it through a masker, which is the name Nilearn gives to its class of objects that can transform raw Nifti objects into Numpy arrays. Then, I stored these 2D Numpy arrays as CSVs so that they could be called later.

Data Visualization

The activity in the orbitofrontal cortex of one patient. The x-axis is the number of each individual time step, and the y-axis represents the unnormalized blood flow in that brain region.

Three cross-sections of one slice of an rs-fMRI. Views for y = -44, x = 0, and z = -2.

Problem Formulation

• Input: time series gathered from masker, as detailed in preprocessing
• Models:
  • Running a DNN through sklearn.
• The metric was accuracy, considering that this is just a binary classification; precision and recall were also to be considered.

Training

• I used Google Colab to load the data, code the model, and save the model for further use.
• I trained 1000 simulations of one MLP model.
  • Hidden Layers: (30,15,5,)
  • Solver: Stochastic Gradient Descent
  • Activation: ReLU

Conclusions

• Results from 1000 simulations of the classifier model showed a normal distribution around 50% accuracy, with a maximum accuracy of 81.25%.
• Precision and recall each showed similar results, with a maximum of 100% during many runs.
• The study was limited by a small sample size due to hardware limitations, and any sample size larger than 100 needs at least 32 GB of RAM to run, especially during the smoothing process.
• The dataset takes up over 100 GB of disk space storage, but using 80 patients took up only 8 GB that was able to be loaded into Google Colab, the working environment in which this study was done.

Future Work

• Patients in this study were from only one center, and using patients from multiple brain scan centers as provided in the dataset may contribute to a higher accuracy.
• The utilization of the subcortical atlas could explore regions further embedded in the brain, such as the amygdala, which may contribute to a more successful model.

How to reproduce results

Overview of files in repository

• Nilearn_fMRI.ipynb: The current notebook I am working in.
• TS_NYU: The folder containing 80 CSVs of pre-processed patient data.
• Discover2023: The folder containing how the project was presented at a symposium.
• NYU_0051002_OFC_standardFALSE.png: An example of the change from mean blood flow plotted against time step iteration.
• README.md: YOU ARE HERE
• UTA-DataScience-Logo.png: Link to the logo displayed at the top.

Software Setup

• One of the prominent packages involved in this project is Nilearn, a package of Python that specailizes in providing statistical methods to analyze brain volumes and scans.
• These packages may be installed via pip or pip3 in your local Python terminal or within a Jupyter notebook.
  • Running the command !pip install nilearn within your notebook will install these easily within your notebook.

Data

• Data can be downloaded through a website or through the function nilearn.datasets.fetch_abide_pcp()`
• Preprocessing and use of the masker can be done by copying the functions within the notebook, specifically nilearn.datasets.fetch_atlas_harvard_oxford() with your choice of atlas style as described in the documentation.

Citations

• Identification of autism spectrum disorder using deep learning and the ABIDE dataset - PubMed
• Disease prediction using graph convolutional networks: Application to Autism Spectrum Disorder and Alzheimer’s disease - ScienceDirect
• Enhancing studies of the connectome in autism using the autism brain imaging data exchange II | Scientific Data
• ABIDE Dataset