-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathReadme_plus
135 lines (101 loc) · 4.79 KB
/
Readme_plus
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
SLIC: a whole brain parcellation toolbox
Copyright (C) 2016 Jing Wang
1. About fMRI data --------------------------------------------------------
We first download the Beijing_Zang dataset, which is part of the 1000
Functional Connectomes Project on NITRC.
https://www.nitrc.org/frs/?group_id=296
(Zang, Y.F.; n = 198 [76M/122F]; ages: 18-26; TR = 2; # slices = 33; #
timepoints = 225)
The subjects were preprocessed by DPARSFA (http://rfmri.org/DPARSF). The
preprocessed fMRI data from 40 subjects are shared in this demo. They are
chosen from the first 41 subjects. Sub12220 was excluded since there are
empty time courses within the graymatter mask in this subject. You could
download the preprocessed fMRI data from the following link. The total
size is 2.82 GB.
https://www.nitrc.org/frs/?group_id=1030
Alternativly, you might use your own database, or preprocess the
Beijing_Zang dataset in your own manner. The only requirement of this
demo is that it requires the preprocessed resting-state fMRI data of 40
subjects.
2. About preprocessing of the fMRI data -----------------------------------
The preprocessing procedure is detailed in our paper. Some utilities could
be found in the following file, including the script of preparing the fMRI
data for DPARSFA, configuration of DPARSFA, and the script of exporting
the preprocessed data.
https://www.nitrc.org/frs/download.php/9507/DPARSF_configuration.zip
3. About software ---------------------------------------------------------
Matlab R2012b/R2014b
SPM8 V4667
DPARSF_V2.3_130615
REST_V1.8_130615
Matlab R2016b
SPM12 v6225
DPABI_V2.2_161201
Both combinations are all right. There might be errors and warnings when
using DPARSF. They could be overcome by experiences, Google, or contacting
the author of DPARSF (http://rfmri.org/DPARSF).
This demo also depends on the NIFTI toolbox:
NIfTI_20140122.
4. About hardware ---------------------------------------------------------
The experiments were run on a Dell(R) workstation with twenty 2.80 GHz
Intel(R) Xeon(R) processors and 256 GB memory. With this server, the
number of parallel workers could be set to 20~40. You might adjust this
number by the function parc_parpool according to the accessible
computational resources. The out of memory problem might occurs when
running this toolbox in a personal computer since the dimensionality
of fMRI data is rather high.
5. About redundancy in eigendecomposition ---------------------------------
To parcellate the brain into K clusters, K indicator vectors corresponding
to the K smallest nontrivial eigenvalues (>10^-4) should be computed. To
compute K indicator vectors, we should calcualte more than K eigenvectors
since the number of nontrivial eigenvalues is unpredictable in prior.
Assume the number of redundant eigenvectors is delta. The delta value
should be large enough to cover the nontrivial eigenvalues. On the other
hand, when delta is too large, it would consume a lot of time to do
eigendecomposition. In practice, this value is determined by
trial-and-error.
A clue is to check the number of empty rows and columns (nEmpty) in the
weight matrices. If nEmpty is zero or very small, then the delta value
could be very small. For example, delta = 10~20 could cover most cases.
If nEmpty is large, e.g., for the individual subject level weight matrices
generated by SS3, the delta value should be set to be a very large value,
e.g., nEmpty + 600. The smallest (or close to the smallest) delta values
I found in practice are listed below.
SS1
sub_eigen, 1
mean_eigen, 1
MSC_twolevel_eigen, 2
SLIC_twolevel_eigen, 16
SS2
sub_eigen, 1
mean_eigen, 1
MSC_twolevel_eigen, 1
SLIC_twolevel_eigen, 1
SS3
sub_eigen, nEmpty+600
mean_eigen, 2
MSC_twolevel_eigen, 1
SLIC_twolevel_eigen, 1
6. About different weighting functions and different sparsifying schemes --
It could be adjusted by the parameter "opt" in sub_weight.
opt=1, correlation with SS1
opt=2, Gaussian with SS1
opt=3, correlation with SS2
opt=4, Gaussian with SS2
opt=5, correlation with SS3
opt=6, Gaussian with SS3
7. Others -----------------------------------------------------------------
(1) GSR. Include GSR in the preprocessing procedure.
(2) Overclustering. Fix the cluster number in the individual subject level to
be 1000.
(3) Random parcellation. For the individual subject level weight matrics
generated by correlation with SS1, set the nonzero elements to be one.
Reference:
Jing Wang, Haixian Wang. A supervoxel-based method for groupwise whole
brain parcellation with resting-state fMRI data. Frontiers in Human
Neuroscience. DOI: 10.3389/fnhum.2016.00659
Contact information:
Jing Wang
wangjing0@seu.edu.cn
yuzhounh@163.com
2016-12-13 22:06:10