1. ICA vs. SPM for Similarity Retrieval of fMRI Images

2. Overview of Sara’s and Rosalia’s Method

3. Similarity between Components
ICA (Sara’s Method)
Extract Spatial Feature Vectors
fastICA
Similarity between Components
fMRI raw Data
108 components
IC maps
Each component associates with time-course

4. SPM (Rosalia’s Method)
SPM Z Score with FDR=0.05
Get SPM clusters (connected Component analysis)
Compute Feature Vectors
Similarity Between Two Brains

5. How to Compare these two methods?
Similarity between components => similarity between brains? ?
Bai etc. (2007)
Maximum Weight Bipartite Matching used fMRI brain image retrieval based on ICA components.

6. Optimal Assignment and Bipartite Matching
Given two sets s1 and s2, is defined as the cost between each element i in S1 and each element j in S2.
An optimal assignment is a permutation p = (p1, , pn) of the integers (1, , n) that minimizes S1 S2

7. Example of brains with two components
0.7 1 1
=0.3
=0.2
0.8 2 2
Brain2
Brain1
Minimum Cost = C12+C21 = = 0.5
Similarity Score =0.5

8. Apply Optimal Assignment
Step 1. Convert Similarity Matrix to a cost Matrix
Step 2. Find the minimum cost between two brains using Bipartite Matching.

9. The comparison SPM data: FaceUpVsFixation
ICA data: raw fMRI data of 108 scans
Run Rosalia’s method to get a similarity Matrix M1
Run Sara’s method and bipartite matching to get similarity Matrix M2

10. The results Correlation between M1 and M2: 0.7132???
Not a very strong correlation

11. Possible Reasons Convert Similarity Matrix to Cost Matrix?
Raw fMRI data has more than two cognitive tasks, while SPM contrast map has only two cognitive tasks.
Is Bipartite Matching a good measure for similarity between two brains? C2
Fix
FaceUp
HouseUp C1

12. Reference
B. Bai, P. Kantor, A. Shokoufandeh, D. Silver. fMRI brain image retrieval based on ICA components. enc, pp ,  Eighth Mexican International Conference on Current Trends in Computer Science (ENC 2007),  2007
Y. Cheng, V. Wu, R. Collins, A. Hanson, and E. Riseman. Maximum-weight bipartite matching technique and its application in image feature matching. In Proc. SPIE Visual Comm. And Image Processing, Orlando, FL, 1996.

13. Component Similarity Measure
Each feature is weighted according to its mean and std:
if < threshold
Similarity Score:

14. SPM (Rosalia’s Method)
Preprocess t-contrast maps of particular cognitive tasks to get the activated voxels.
Cluster the resulting voxels to distinct regions.
Similarity Measure between brains
Q-to-T Score =
T-to-Q Score =
Similarity Score =

15. Example The cost between S1 and S2 can be represented by a nxn matrix
1 2 3
C11+C22+C33
1 3 2
C11+C23+C32
3 2 1
C13+C22+C31
2 3 1
C12+C23+C31
2 1 3
C12+C21+C33
3 1 2
C13+C21+C32 S2
C11
C12
C13
C21
C22
C23
C31
C32
C33 S1

16. Apply Optimal Assignment
Step 1. Convert Similarity Matrix (sM)to a cost Matrix (cM)
Replace infinite value in sM to the maximum value of all the finite values.
Take top N percent of the similarity scores, set others to be zero.
Set the non-zero scores to be 1-SM/maximum.
set the zeros to be infinite value.
Step 2. Using Kuhn-Munkres Algorithm (Hungarian method) to find the minimum cost between two brains.