Building Pipelines for Processing Neurological MRI Images & Data

Uday Suresh
Sep 19, 2017 · 4 min read

At UCSF Mission Bay, in a collaborative project between Peder Larson’s Lab in the Department of Radiology and Biomedical Imaging and Roland Henry’s Lab in the Department of Neurology, I’ve been working on building a series of biologically computational pipelines. Primarily, I’ve been supervised by Anisha Keshavan in Professor Henry’s Lab, and she’s provided me with an immense amount of guidance — essentially teaching me from scratch how to meaningfully utilize a computer to extract information about neurological data.

The main tool I’ve been using is Nipype — an open source interface built under NiPy in Python that bundles up a lot of existing neuroimaging software for building pipelines and processing workflows. The workflows that I created were modeled after and belong in the subset of workflows that exist as a part of the collection of such pipelines for MRI data processing within UCSF’s PBR (Pipelines for Brain Repository). Each workflow in PBR is targeted towards processing MRI images in a particular way, and mine are specifically for handling Ultrashort Echo Time (UTE) pulse sequences. These ultrashort echo time sequences offer a unique diagnostic window into MRI imaging as they are able to detect tissue components that run-of-the-mill 3 Tesla and 7 Tesla standard MRI imaging can't detect.

Nipype’s workflow framework from the docs

Nipype is useful as it is able to connect various processing nodes to build a cascading workflow to move imaging data through. It is easily possible to link the various algorithms from different packages, as shown here. This preprocessing workflow for basic alignment of MRI data relies on FSL and FreeSurfer.

Visualization of node connection in a preliminary preprocessing workflow

Building a pipeline for UTE image processing involves a lot of guessing and checking. The integral idea is to adjust image characteristics to threshold and quantify the myelin degradation within the brain’s white matter. Our study specifically is concerned with imaging the demyelination that occurs in the brain through the onset of Multiple Sclerosis (MS) and uses this UTE signal to probe nervous system failure. This requires a series of Nipype nodes that will best quantify the demyelination occurring in these lesions and plaques that appear within the grooves of the brain as the myelin sheath degrades.

Most of my attempts involved trying to aptly align the image of the brain before adjusting contrasts and filters by applying various masks on the scan. NIfTI (Neuroimaging Informatics Technology Initiative) files containing scan data are taken and situated before being dissected to focus in on the visible white matter. A binary brain mask of the image is created using FSL’s program BET such that from the noisy neuroimaging scan data the brain image is exclusively extracted.

Slices of MRI scan images

To expand upon the brain image extraction, an anatomical registration tool called ANTs (Advances Normalization Tools) is used to map the identifiable features in the image to various structures of the brain. This medical image registration and segmentation toolkit allows us to hone in on the white matter and portions of the brain that are susceptible to demyelination.

Images are processed in batches throughout a range of prospective slices of the brain that are of interest. The nodes in the workflow are connected to eventually reach a DataSink for interactive outputs, chaining the processing nodes to an eventual dump in a directory. The core framework for Nipype workflows are as follows, with the nodes resembling the preprocessing pipeline visually mapped out prior:

In this particular workflow, to create some possible image contrasting, I’ve adjusted the bet.inputs.frac such that the fractional intensity threshold is varied as BET runs its skull stripping algorithm to create a brain mask. These sibling images will be run in parallel through the entirety of the pipeline, and processed by the DataSink neatly such that they will be bundled up and placed in the directory of choice.

My work in chasing the perfect datasets that flaunt the UTE signal’s ability to masterfully probe the brain’s white matter for lesions is still in progress, but the pipelines that I’ve built have — through trial and error — provided me with doses of quantitative results of demyelination. I’m particularly partial to my research for its ability to be more than just interesting and challenging, but also applicable and compelling. Neurological imaging is more than just a series of biophysical quandaries when the results have the potential to shape the medical treatment of MS patients.

Brain layers

This image registration and segmentation with multi-echo datasets — beyond just the UTE signal — over small timescales is the overarching direction of the research that I’m conducting as a research associate at the Larson Lab.

Contact me:

A part of the UCSF Larson Lab

Thank you to Professor Peder Larson, Professor Roland Henry, and Anisha Keshavan for instruction & guidance

Uday Suresh

Written by

Bioengineering + Creative Writing @ UC Berkeley ’18

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade