Locating Seed Automatically in PCC for rs-fMRI Data Analysis by Using Unsupervised Machine Learning
Poster No:
1927
Submission Type:
Abstract Submission
Authors:
Mingyi Li1, Katherine Koenig2, Mark Lowe2
Institutions:
1Cleveland Clinic, Cleveland, OH, 2The Cleveland Clinic, Cleveland, OH
First Author:
Co-Author(s):
Introduction:
When using seed-based methods to analyze resting state fMRI (rs-fMRI) data, the selection of the seed has large influence on the results. In previous study, we presented a method to automatically produce seed locations from voxel clusters generated by self-organizing map (SOM) whose input was feature vectors formed by combining anatomical and rs-fMRI data[1]. The limitation of the work was bad seed search ROI derived from FreeSurfer brain parcellation in some subjects. To solve the problem, we attempted to propagate template searching ROI to individual subjects by using nonlinear registration. However, the improvement was not obvious compared to FreeSurfer method. Later, we found out SynthSeg, a deep learning based brain parcellation tool, produced obviously better results in brain gray matter parcellation in one of our datasets [2]. In this study, we tested seed searching ROI generated by using SynthSeg [3].
Methods:
Data acquisition: Nineteen subjects consisting six healthy controls and thirteen patients were scanned in an IRB-approved protocol at 3T Siemens scanner (Erlangen, Germany) using a bitebar to reduce head motion, in a 12-ch receive head coil. Scans included T1-MPRAGE (voxel size=1x1x1.2mm, matrix size=256x256x120, TE/TR/TI=1.75/1900/900ms), and rs-fMRI(voxel size=2x2x4mm, matrix size=128x128x31, TR/TE/FA=2800/29/80, 132volumes).
Data processing: Each rs-fMRI dataset was motion-corrected, low-pass filtered and spatially filtered. T1 image was parcellated into ROIs by using SynthSeg. The ROIs were registered to rs-fMRI image space by using "align_epi_anat.py" tool in AFNI[4]. The ROI covering the left/right PCC was used as seed searching region for automatic seed generation method (Left side and right side are processed separately). From rs-fMRI data, the global connectivity between each voxel in the seed searching ROI and all other brain cortex voxels were computed and then the connectivity distribution was fitted into a Gaussian distribution [5]. The feature vector was formed by counting the number of voxels whose connectivity value was outside three standard deviations, in the parcellated cortex ROIs. The above feature vector forming step is shown in figure 1. Then the feature vectors of all the voxels in the searching region was feed into a size 4x5 SOM classifier in Matlab. For SOM cluster is further divided into spatially aggregated sub-clusters by using k-means clustering. At last, voxels with weak correlation to other voxels in the same cluster were discarded. The remaining voxels formed the seed clusters. The connectivity map of a seed cluster was the average of the connectivity maps computed for all the voxels in the seed cluster.
Using each subject's rs-fMRI data, seeds in PCC were also manually located by experts through the "instacorr" method in AFNI.
Seeds comparison: Seeds acquired through the automatic method were compared to those picked through the manual method.
Data processing: Each rs-fMRI dataset was motion-corrected, low-pass filtered and spatially filtered. T1 image was parcellated into ROIs by using SynthSeg. The ROIs were registered to rs-fMRI image space by using "align_epi_anat.py" tool in AFNI[4]. The ROI covering the left/right PCC was used as seed searching region for automatic seed generation method (Left side and right side are processed separately). From rs-fMRI data, the global connectivity between each voxel in the seed searching ROI and all other brain cortex voxels were computed and then the connectivity distribution was fitted into a Gaussian distribution [5]. The feature vector was formed by counting the number of voxels whose connectivity value was outside three standard deviations, in the parcellated cortex ROIs. The above feature vector forming step is shown in figure 1. Then the feature vectors of all the voxels in the searching region was feed into a size 4x5 SOM classifier in Matlab. For SOM cluster is further divided into spatially aggregated sub-clusters by using k-means clustering. At last, voxels with weak correlation to other voxels in the same cluster were discarded. The remaining voxels formed the seed clusters. The connectivity map of a seed cluster was the average of the connectivity maps computed for all the voxels in the seed cluster.
Using each subject's rs-fMRI data, seeds in PCC were also manually located by experts through the "instacorr" method in AFNI.
Seeds comparison: Seeds acquired through the automatic method were compared to those picked through the manual method.
·Figure 1. Generating feature vector by combining rs-fMRI connectivity and T1 parcellation. Panel A: Z-map, Panel B: Z-score distribution, Panel C: Brain parcellation, Panel D: feature vectors.
Results:
In four out of six control subjects, there was one automatically generated seed overlapping with manually picked seed. There was no matching for the remaining two control subjects. In nine out of thirteen patients, there was one automatically generated seed overlapping with manually picked seed. There was no matching for the remaining four control subjects.
SynthSeg apparently did a better job than FreeSurfer on generating seed searching region. For the two unmatched control and three unmatched patients, the manually located seed was partially inside the seed searching ROI of the automatic method. For one unmatched patient, the manually selected seed was outside searching ROI.
One matched patient case is shown in figure 2.
SynthSeg apparently did a better job than FreeSurfer on generating seed searching region. For the two unmatched control and three unmatched patients, the manually located seed was partially inside the seed searching ROI of the automatic method. For one unmatched patient, the manually selected seed was outside searching ROI.
One matched patient case is shown in figure 2.
·Figure 2. Matched seeds. Top row shows manually picked seed and bottom row shows automatically generated seed.
Conclusions:
The performance of the automatic seed generation method was universal across control and patients.
We will test out region grow technique on generating seed searching ROI.
We will test out the method in a much larger dataset in an ongoing study.
We will test out region grow technique on generating seed searching ROI.
We will test out the method in a much larger dataset in an ongoing study.
Modeling and Analysis Methods:
Methods Development 1
Task-Independent and Resting-State Analysis 2
Keywords:
Computing
Data analysis
Development
FUNCTIONAL MRI
Machine Learning
MRI
1|2Indicates the priority used for review
Provide references using author date format
[2] M. Li, et al (2023), Comparison of 7T MRI Brain Image Parcellation Results Between FreeSurfer and SynthSeg, 29th Annual Meeting of the Organization on Human Brain Mapping
[3] Billot B., et al. (2022), Robust Segmentation of Brain MRI in the Wild with Hierarchical CNNs and no Retraining, MICCAI
[4] Cox RW. (2012), AFNI: What a long strange trip it’s been, NeuroImage, 62(2):743-747
[5] Lowe MJ et al. (1998), Functional Connectivity in Single and Multislice Echoplanar Imaging Using Resting-State Fluctuations, NeuroImage, 7(1):119-132.