Comparing longitudinal registration tools for 2D MRI

[brieuclehmann]

Comparing longitudinal registration tools for 2D MRI

Project Description

While research MRI anatomical images are usually 3D (e.g. FLASH), clinical scans are typically 2D acquisitions with thick slices. In this project, we take up the challenge of longitudinal registration with 2D scans as would typically be acquired in a long term clinical trail (e.g. for multiple sclerosis). Longitudinal brain imaging can be particularly useful in the analysis of volumetric changes or lesion burden, and shows great promise for the development of novel biomarkers.

Registration is a key step in the pipeline that affects all further downstream analysis of neuroimaging data. Although using cross-sectional tools to process longitudinal data is unbiased, this ignores the common information across scans. Longitudinal processing aims to reduce the within-subject variability. Both SPM and FreeSurfer offer tools for longitudinal registration of scans across multiple (more than two) time points and, as with most image processing tools, these have naturally been developed with research-quality data in mind. As researchers are increasingly gaining access to clinical data, however, it would be timely to determine how current longitudinal processing tools perform on lower-quality 2D MRI scans.

Using the publicly-available OASIS dataset, we would like to investigate the performance of the SPM and FreeSurfer longitudinal registration tools. The OASIS-3 (Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer’s Disease) dataset consists of images from c.1000 subjects, many of which are accompanied by volumetric segmentation files produced through FreeSurfer. With these files as a 'gold-standard', we will average slices from 3D acquisitions to simulate 2D acquisitions and assess the accuracy of each processing tool.

Skills required to participate

Any of the following:

• Experience in programming (mainly Matlab, C or C++)

• Experience with FreeSurfer or SPM12

• Experience with structural image analysis

Integration

Contributions towards any of the following milestones would be very welcome!

Milestones

1. Downsample OASIS T1 3D data to lower-resolution 2D images

2. Isolate the longitudinal registration codebase from FreeSurfer

3. Longitudinal registration of 2D images in SPM and FreeSurfer

4. Assessment of segmentation performance to original 3D images

Preparation material

The OASIS project
Chapter 27 (Longitudinal registration) of the SPM 12 manual
FSL longitudinal processing

Papers

• M. Reuter, N.J. Schmansky, H.D. Rosas, B. Fisch (2012): Within-subject template estimation for unbiased longitudinal image analysis
• J. Ashburner, G.R. Ridgeway (2013): Symmetric diffeomorphic modeling of longitudinal structural MRI

Communication

Join the chat in our mattermost channel :)

[ltetrel]

really interesting, I still don't get this part "we will average slices from 3D acquisitions to simulate 2D acquisitions" but we can discuss about it tomorrow :)