Title: Segmentation of Multi spectral Magnetic Resonance Image using Penalized Fuzzy competitive Learning N
1Segmentation of Multi spectral Magnetic Resonance
Image using Penalized Fuzzy competitive Learning
Network
- Jzau-Sheng Lin, Kuo-Sheng Cheng Chi-Wu Mao
2Some Background
- Segmentation of Images
- Multi Spectral Images
- MR Scanner
- Magnetic Resonance Imaging
3Segmentation
- Segmentation of an image is a process of finding
a region which is connected to a set of pixels
that have one or more similar properties that are
different from other surrounding pixels. - Can be done by Thresholding, Clustering , etc.
4Multi Spectral Images
- More than one image of the same object or scene
taken in different bands of optic or infra
wavelength. - Ex
-
- Green Red
- Photo IR False IR
5MRI- Magnetic Resonance Imaging
- Its a scanning technique
- MRI does not use radiation
- It combines the use of a large magnet and radio
waves. - The hydrogen atoms in the patient's body react to
the magnetic field, and a computer analyzes the
results and makes pictures of the inside of your
body.
6MR Scanner
7MR Scanner
8MR Images
9Critical Parameters in MRI
- Spin-Spin Relaxation time (T2)
- Spin-Lattice Relaxation time (T1)
- Proton density (PD)
- Spin- is a fundamental property of nature like
electric charge or mass. Protons, electrons and
neutrons have spin(1/2).
10Spin-Lattice Relaxation time T1
- External Magnetic field is along Z axis.
- If we move a magnetic field from its equilibrium
then its net Magnetization vector is 0. - It tries to return back to its original position,
the time it takes to do so is called Spin Lattice
Relaxation time T1
11Spin-Spin Relaxation time T2
- If magnetization is along XY
- then time taken to return it to equilibrium is
called spin-spin relaxation time or T2 - T2 images are used to distinguish healthy from
pathological tissues or for classification of
tissues
12The Problem
- To segment the multi spectral magnetic resonance
images. - Get a good segmentation method for the
classification of the tissues.
13Artificial Neural Networks
- Powerful technique in pattern recognition and
decision making - ANNs are Supervised ANN
- Unsupervised ANN
- This paper uses an unsupervised scheme based on
least squares criterion.
14Methods
- Fuzzy Clustering Algorithms
- Fuzzy C-Means, Penalized C-Means
- Architecture between conventional competitive and
penalized fuzzy CNN - Experimental results obtained from PFCM
15FUZZY CLUSTERING TECHNIQUES
16Fuzzy Clustering Techniques
- Overview on Clustering
- Basic Notions
- -Data Sets
- -Clusters Cluster Prototypes
- Overview on Clustering Methods
- Fuzzy C-Means Objective Function
- Fuzzy C-Means Algorithm
- Penalized Fuzzy C-Means Algorithm
17Clustering Techniques
- Features
- - Unsupervised Methods
- - Organization of Data based on Similarities
- - Useful in Absence of Priori Knowledge
- -Applications
- - Image Processing Segmentation
- - Speech Recognition
- - Data Compression
- - Modeling Identification
- - Pattern Recognition
-
18Basic Notions in Clustering
- The Data Set
-
- Clusters Prototypes
-
19Overview on Clustering Methods
- Crisp or Hard Clustering
- Soft or Fuzzy Clustering
20Fuzzy Clustering Medical Science ?
Easy Versatile Valuable Supportive Tool
21Most widely used
FCM Algorithm
Well known
Powerful
22Goal of the FCM
Detect similarity between members of a
collection
Minimize the criteria in the least squared
error sense
23Basis of the FCM Algorithm
- Fuzzy C Means Objective Function / Functional
24The Fuzzy C-Means Functional
- c n
- JFCM 1/2 ?j1 ? i1 (?i,j)m xi-wj2
- where,
- xi i th sample, i1,2,3,.n
- c number of clusters, c? 2
- n number of data sample vectors
- ?i,j grade of membership of xi in cluster j
- wj cluster centroids,w1,w2,w3wjwc
- m fuzziness parameter
25Parameters of the FCM algorithm
Termination Criterion, ?
26Number of clusters,c
- Most Important
-
- Assumed
-
- Or Chosen Equal to the Number of Groups Existing
in the Data. -
- - Other parameters
- have less influence
- on the resulting
- partition
- - When clustering real data if no priori
information available - - In case of brain images 4 clusters to
represent the CSF, the white matter, the gray
matter, and the skull bone
27Fuzziness Parameter, m
- Influences Fuzziness of Resulting Partition
- Reduces the Noise Sensitivity
- Effect for ? i,j Depends on m
- - larger the value of m, higher is the
dependence - Usually, m2
28Termination Criterion, ?
- Terminaton condition
- for the FCM algorithm
- Usual choice for ?
- is 0.001.
- Choosing ? 0.01 drastically
- reduces the computing times
29The FCM Algorithm
- Step 1
- Given a data set,
- Select the number of clusters
- Initialise the cluster centroids, wj
- ( 2 ? j ? c )
- Initialise the fuzzification parameter, m
- (1 ? m ? ?)
- Initialise the termination tolerance, ? gt0
- Initialise the fuzzy membership matrix U to U(0),
- U(0) ? U
30 The FCM Algorithm
- Step 2
- Calculate the Euclidean distance, di,j
- between each training sample xi and
- the class centroid wj
- Calculate the membership matrix Uµi,j,
- using the equation,
-
- (1/(d i,j )2)1/ (m-1)
- µi,j c
- ?j1 (1/(d i,j )2)1/(m-1)
-
31The FCM Algorithm
Step 3 Update the class centroids
n ? i1( ? i,j )m xi
wj c ?
i1( ? i,j )m
32The FCM Algorithm
Step 4 Compute ? max (
U(t1) - U(t) ). If ? gt ? , then go to Step
2 otherwise go to Step 5.
33The FCM Algorithm
Step 5 Find the results for the final class
centroids.
34Penalised FCM Algorithm
More Effective
Generalised type of FCM
35Penalized Fuzzy C-Means Functional
- c n
- J PFCM ½ ?j1 ? i1 (?i,j)m xi-wj2
- c n
- - ½ ? ?j1 ? i1 (?i,j)mln? j
-
- J FCM - ½ ? ?j1 ? i1 (?i,j)mln? j
The penalty term
36where, ? j proportional constant of class
j given by, n ? i1( ? i,j )m
? j c n
? i1 ? i1( ? i,j )m ? a constant, ? ?
0 J PFCM J FCM when ? 0
37 and, wj same as for J FCM
c ((d i,j )2 - ?ln? j )1/ (m-1)
-1 µi,j ?l1 ((d
i,j )2 - ?ln? j )1/(m-1)
38The PFCM Algorithm
- Step 1
- Given a data set,
- randomly set the following
- - cluster centroids wj ( 2 ? j ? c )
- - fuzzification parameter, m, (1 ? m ? ?)
- - the termination tolerance, ? gt0
- Give a fuzzy c-partition U(0).
-
39The PFCM Algorithm
- Step 2
- Using the equations,
- compute the ? j (t) and wj(t) using U (t-1)
- Calculate the membership marix U ? i,j
- with ? j (t) and wj(t)
40The PFCM Algorithm
- Step 3
- Compute ? max ( U(t1) - U(t) )
- If ? gt ? , then go to Step 2
- otherwise go to Step
41The PFCM Algorithm
- Step 4
- Find the results for the final class centroids
42Neural Networks
Weights adapt thru unsupervised learning
rules
- Supervised
- Fixed Weight
- Unsupervised
- Competitive Neural Networks
- Learning rules
- Supervised (error based)
- unsupervised (output based)
No Teacher guidance
43Basic Competitive Learning Network
- One layer of input neurons and one layer of
output neurons - An input pattern x is a sample point in the
n-dimensional real or binary vector space. - Output nodes - Binary-valued (1 or 0) local
representations - output neurons number of classes
44Basic Competitive Learning Network
- comprises the
- Training Rules
- feed forward excitatory network(s)
- lateral inhibitory network(s).
- Minimal learning Model-Fixed output
nodes(clusters)
Hebbian rules
Winner take all (WTA)
45Competitive learning network
- Input is multi spectral magnetic spin echo images
X255, 240, 245 - Hebbian Learning Rules
-
- Single layer of neurons
- Fully connected to output nodes
46Competitive Learning Network
X1
? 1,1
X2
w1
X3
w2
X4
w3
X5
w4
.
wc
Xn-1
? n,c
Xn
47Conventional Competitive Learning Neural Network
- Least squares criterion
- Hebbian learning rules
- Modifies the winning unit to move them closer to
the input - Similar to c-means clustering, cluster centroids
in multidimensional pattern space
48Parameters of FCM
- nj-number of pixels in a class cj
-
- wj be the mean of the class (i.e. centroids)
- wj?xj?cj xi/nj
- w0 be the global center of mass of X
- wj?xj?cj xi/nj
-
49Scatter Functions
- JT JWJB
- c n
- JT ? ? xi-w0 2
- j1 xj?cj
- c
- JW 1/2 ? ? xi-wj 2
- j1 xi?cj
- c
- JB 1/2 ? ? wi-w0 2
- j1 xi?cj
JT- Total Scatter Function
JW- Within Class
JB- Between-class
50Criterion
- Minimization of JW gt Maximization of JB
- So our criteria function could be
- c
- JW 1/2 ? ? xi-wj 2
- j1 xi?cj
- i.e. we want to minimize the sum squared error
vector for each class xi-wj
51Cont.
- JW represents the least sum of squares error
between n samples within class. - We get c class centers w1,w2,wc
- Objective function for the learning network is
modified to - c n
- JC 1/2 ? ? ? i,j xi-wj 2
- j1 i1
- ? i,j 1 if xi belongs to cj and ? i,j 0 for
all other clusters
52Winner-take-all-neuron
- The neuron that wins is called winner-take-all-neu
ron. - ? i,j indicates whether the input sample xi
activates neuron j to win. - ? i,j 1 if xi - wj lt xi - wk , for
all k - 0 otherwise
53Gradient (Steepest) Descent
- Gradient descent on the objective function Jc
gives - n
- ?wj -??Jc -?? (xi-wj) ? i,j
- ?Wj i1
- The update rule is the sum over all samples, its
is usually used incrementally, i.e., sample after
sample - ?wj ? (xi-wj) ? i,j valid for all j and
- wj(t1) wj(t) ?wj(t)
- ? - learning-rate parameter
54Algorithm
- Initialize the cluster centroids wj (2 lt j lt c)
- Initialize learning rate ?, and states of input
samples U? i,j - Update neuron states with competitive learning
- Compute synaptic weights
- Repeat 3, 4 for all input samples, record no of
neurons with changes state. If no neuron state is
changed go to 6 - Output final classification
55Penalized Fuzzy Competitive Learning Neural
Network
- Same as the conventional learning network.
- Unsupervised network with penalized fuzzy
reasoning - Objective function
- c n
c n - JC,PFCM 1/2 ? ? ? i,j xi-wj 2 -1/2 ? ? ? ?
i,j ln aj - j1 i1
j1 i1 - c n
- JFCM -1/2 ? ? ? ? i,j ln aj
- j1 i1
-
aj -is a proportionality constant
? - is a constant gt0
56Why Penalty?
- c n
- JC,PFCM JFCM -1/2 ? ? ? ? i,j ln aj
- j1 i1
- since aj is the weighted average of ? over the
entire data aj lt 1 ? ln aj lt 0
- Hence we are actually adding a penalty which is
log of a proportionality constant times the
membership of each sample summed over the entire
data. - This adds a term to the evaluation function,
which increases it by a small constant, hence can
give better weights.
57Gradient Descent
- Gradient descent on the objective function Jc
gives - ?wj
- n
- -??(Jc,PFCM)-???(JFCM) -1 ?m(? i,j)m-1(ln aj) ?
? i,j - ?Wj i1 ?wj 2
?wj - Simplifying this by obtaining the derivative of ?
i,j and replacing in the above equation, the
gradient descent on the objective function can be
updated as - n
- ?wj ? ? ? i,jm (xi-wj)1- m (1-? i,j )
- i1 m-1
- ? - learning-rate parameter
58Gradient descent
- ?wj ?? i,j (xi-wj)1- m (1-? i,j )
- m-1
- ? - learning-rate parameter
59Algorithm
- Initialize cluster centroids wj,
- fuzzification parameter m,
- learning rate ?,
- constant ?
- and the value egt0.
- Give a fuzzy c-partition U(0).
60Algorithm cont.
- Find aj (t) ,wj (t) with U(t-1). Calculate U?
i,j . - Sequentially update weights of a neuron using
competitive learning - Compute ?max( U(t1) -U(t)). If ? gt e, goto 2
else goto 5. - Execute a defuzzification process and output
final results
61Defuzzification process
- A pixel is assigned to the cluster when its
membership grade in that cluster is greater than
0.5. - If none of the grades satisfy then the class
with maximum grade is chosen, provided that sum
of two largest gt0.5
62Experimental Results Discussion
- Data Input
- Parameter Values Used
- Acquisition Parameters
- Network Associated Parameters
- Discussion
- Results
- Performance Evaluation
- Conclusion
63Data Input Used
- T2 weighted MR images
- Recorded from a man of 32 years
- Acquired on a Siemens 1.5 Tesla
Magnetom MR scanner - 5 mm Slice Thickness
- 1mm Interslice Space
- 256 X 256 Pixels Matrix Size
- 8-bit Gray Levels with Spin Echo
- Sequences
64Data UsedAcquired Brain Images
TR1/TE12500ms/75ms
TR2/TE22500ms/100ms
TR3/TE31500ms/59ms
65Data UsedExtracted Peritoneal Cavity Images
TR 2500 ms TE 158 ms
TR 2500 ms TE 130 ms
TR 2500ms TE 144 ms
66Parameter Values Used
- for Fig.2a-2c,
- TR1/TE12500 ms/75 ms
- TR2/TE22500 ms/100 ms
- TR3/TE31500 ms/59 ms
- for Fig.3a-3c ,
- TE130,144,158
- TR2500 ms
- Fuzzification parameter, m 1.5
- Learning rate,? 0.3
- The constant, ? 1.2
Acquisition Parameters Network
Associated Parameters
67Discussion
- Pixel Vectors
- Initial Centroid Values
- Iterations
- Errors Due to Noise
- Postclassification Filtering
68Results
Segmented Images with 4 Clusters with 5
Clusters
69(No Transcript)
70(No Transcript)
71Performance Evaluation
-
- Goals Achieved
-
- Successful in the Identification of Most
Important - Regions in the Image
-
- Results Produced are in Acceptable Visual
Agreement with Human Expert Opinion
72Results Independent of the Initial Cluster
Centroids
No Operator Intervention
More Powerful Performance
Conclusion
Stable Results
More Efficient Mechanism
73QUESTIONS COMMENTS ?