Title: Detecting%20sparse%20signals%20and%20sparse%20connectivity%20in%20scale-space,%20with%20applications%20to%20the%20'bubbles'%20task%20in%20an%20fMRI%20experiment
1Detecting sparse signals and sparse connectivity
in scale-space, with applications to the
'bubbles' task in an fMRI experiment
- Keith Worsley, Nicholas Chamandy, McGill
- Jonathan Taylor, Stanford and Université de
Montréal - Robert Adler, Technion
- Philippe Schyns, Fraser Smith, Glasgow
- Frédéric Gosselin, Université de Montréal
- Arnaud Charil, Montreal Neurological Institute
2Astrophysics
3Sloan Digital Sky Survey, data release 6, Aug. 07
4What is bubbles?
5Nature (2005)
6Subject is shown one of 40 faces chosen at
random
Happy
Sad
Fearful
Neutral
7 but face is only revealed through random
bubbles
- First trial Sad expression
- Subject is asked the expression
Neutral
- Response
Incorrect
75 random bubble centres
Smoothed by a Gaussian bubble
What the subject sees
Sad
8Your turn
Subject response Fearful CORRECT
9Your turn
Subject response Happy INCORRECT (Fearful)
10Your turn
Subject response Happy CORRECT
11Your turn
Subject response Fearful CORRECT
12Your turn
Subject response Sad CORRECT
13Your turn
Subject response Happy CORRECT
14Your turn
Subject response Neutral CORRECT
15Your turn
Subject response Happy CORRECT
16Your turn
Subject response Happy INCORRECT (Fearful)
17Bubbles analysis
- E.g. Fearful (3000/4750 trials)
Trial 1 2 3 4
5 6 7 750
Sum
Correct trials
Thresholded at proportion of correct
trials0.68, scaled to 0,1
Use this as a bubble mask
Proportion of correct bubbles (sum correct
bubbles) /(sum all bubbles)
18Results
- Mask average face
- But are these features real or just noise?
- Need statistics
Happy Sad
Fearful Neutral
19Statistical analysis
- Correlate bubbles with response (correct 1,
incorrect 0), separately for each expression - Equivalent to 2-sample Z-statistic for correct
vs. incorrect bubbles, e.g. Fearful - Very similar to the proportion of correct bubbles
ZN(0,1) statistic
Trial 1 2 3 4
5 6 7 750
Response 0 1 1 0
1 1 1 1
20Results
- Thresholded at Z1.64 (P0.05)
- Multiple comparisons correction?
- Need random field theory
ZN(0,1) statistic
Average face Happy Sad
Fearful Neutral
21Euler Characteristic Heuristic
Euler characteristic (EC) blobs - holes (in
2D) Excursion set Xt s Z(s) t, e.g. for
neutral face
EC 0 0 -7 -11
13 14 9 1 0
30
Heuristic At high thresholds t, the holes
disappear, EC 1 or 0, E(EC) P(max Z
t).
Observed
Expected
20
10
EC(Xt)
0
- Exact expression for E(EC) for all thresholds,
- E(EC) P(max Z t) is extremely accurate.
-10
-20
-4
-3
-2
-1
0
1
2
3
4
Threshold, t
22The result
Lipschitz-Killing curvatures of S (Resels(S)c)
EC densities of Z above t
filter
white noise
Z(s)
FWHM
23Results, corrected for search
- Random field theory threshold Z3.92 (P0.05)
-
- 3.82 3.80 3.81
3.80 - Saddle-point approx (2007) Z? (P0.05)
- Bonferroni Z4.87 (P0.05) nothing
ZN(0,1) statistic
Average face Happy Sad
Fearful Neutral
24 Theorem (1981,1995)
25Steiner-Weyl Tube Formula (1930)
Morse Theory method (1981, 1995)
- Put a tube of radius r about the search
- region ?S
- EC has a point-set representation
r
Tube(?S,r)
?S
- Find volume, expand as a power series
- in r, pull off coefficients
- For a Gaussian random field
26Tube(?S,r)
r
?S
Steiner-Weyl Volume of Tubes Formula (1930)
Lipschitz-Killing curvatures are just intrinisic
volumes or Minkowski functionals in the
(Riemannian) metric of the variance of the
derivative of the process
27S
S
S
Edge length ?
Lipschitz-Killing curvature of triangles
Lipschitz-Killing curvature of union of triangles
28Non-isotropic data?
ZN(0,1)
s2
s1
- Can we warp the data to isotropy? i.e. multiply
edge lengths by ?? - Globally no, but locally yes, but we may need
extra dimensions. - Nash Embedding Theorem dimensions D
D(D1)/2 D2 dimensions 5. - Better idea replace Euclidean distance by the
variogram - d(s1, s2)2 Var(Z(s1)
- Z(s2)).
29(No Transcript)
30Non-isotropic data
ZN(0,1)
ZN(0,1)
s2
s1
Edge length ?(s)
Lipschitz-Killing curvature of triangles
Lipschitz-Killing curvature of union of triangles
31We need independent identically distributed
random fields e.g. residuals from a linear model
Lipschitz-Killing curvature of triangles
Lipschitz-Killing curvature of union of triangles
32Scale space smooth Z(s) with range of filter
widths w continuous wavelet transform adds an
extra dimension to the random field Z(s,w)
Scale space, no signal
34
8
22.7
6
4
15.2
2
10.2
0
-2
6.8
-60
-40
-20
0
20
40
60
w FWHM (mm, on log scale)
One 15mm signal
34
8
22.7
6
4
15.2
2
10.2
0
-2
6.8
Z(s,w)
-60
-40
-20
0
20
40
60
s (mm)
15mm signal is best detected with a 15mm
smoothing filter
33Matched Filter Theorem ( Gauss-Markov Theorem)
to best detect signal white noise, filter
should match signal
10mm and 23mm signals
34
8
22.7
6
4
15.2
2
10.2
0
-2
6.8
-60
-40
-20
0
20
40
60
w FWHM (mm, on log scale)
Two 10mm signals 20mm apart
34
8
22.7
6
4
15.2
2
10.2
0
-2
6.8
Z(s,w)
-60
-40
-20
0
20
40
60
s (mm)
But if the signals are too close together they
are detected as a single signal half way between
them
34Scale space can even separate two signals at the
same location!
8mm and 150mm signals at the same location
10
5
0
-60
-40
-20
0
20
40
60
170
20
76
15
34
w FWHM (mm, on log scale)
10
15.2
5
6.8
Z(s,w)
-60
-40
-20
0
20
40
60
s (mm)
35Scale space Lipschitz-Killing curvatures
36Rotation spaceTry all rotated elliptical
filters
Unsmoothed data
Threshold Z5.25 (P0.05)
Maximum filter
37Beautiful symmetry
Steiner-Weyl Tube Formula (1930)
Taylor Gaussian Tube Formula (2003)
- Put a tube of radius r about the search region
?S and rejection region Rt
Z2N(0,1)
Rt
r
Tube(Rt,r)
Tube(?S,r)
r
?S
Z1N(0,1)
t
t-r
- Find volume or probability, expand as a power
series in r, pull off coefficients
38Z2N(0,1)
Rejection region Rt
Tube(Rt,r)
r
Z1N(0,1)
t
t-r
Taylors Gaussian Tube Formula (2003)
39EC densities for some standard test statistics
- Using Morse theory method (1981, 1995)
- T, ?2, F (1994)
- Scale space (1995, 2001)
- Hotellings T2 (1999)
- Correlation (1999)
- Roys maximum root, maximum canonical correlation
(2007) - Wilks Lambda (2007) (approximation only)
- Using Gaussian Kinematic Formula
- T, ?2, F are now one line
- Likelihood ratio tests for cone alternatives (e.g
chi-bar, beta-bar) and nonnegative least-squares
(2007)
40Accuracy of the P-value approximation
The expected EC gives all the polynomial terms in
the expansion for the P-value.
41Bubbles task in fMRI scanner
- Correlate bubbles with BOLD at every voxel
- Calculate Z for each pair (bubble pixel, fMRI
voxel) - a 5D image of Z statistics
Trial 1 2 3 4
5 6 7 3000
fMRI
42Thresholding? Cross correlation random field
- Correlation between 2 fields at 2 different
locations, - searched over all pairs of locations, one in S,
one in T - Bubbles data P0.05, n3000, c0.113, T6.22
Cao Worsley, Annals of Applied Probability
(1999)
43Discussion modeling
- The random response is Y1 (correct) or 0
(incorrect), or YfMRI - The regressors are Xjbubble mask at pixel j, j1
240x38091200 (!) - Logistic regression or ordinary regression
- logit(E(Y)) or E(Y) b0X1b1X91200b91200
- But there are only n3000 observations (trials)
- Instead, since regressors are independent, fit
them one at a time - logit(E(Y)) or E(Y) b0Xjbj
- However the regressors (bubbles) are random with
a simple known distribution, so turn the problem
around and condition on Y - E(Xj) c0Ycj
- Equivalent to conditional logistic regression
(Cox, 1962) which gives exact inference for b1
conditional on sufficient statistics for b0 - Cox also suggested using saddle-point
approximations to improve accuracy of inference
- Interactions? logit(E(Y)) or E(Y)b0X1b1X91200
b91200X1X2b1,2