Lecture Notes in Computer Science
Download 12.42 Mb. Pdf ko'rish
|
A (1)
Mode-1 Matrix I 1
2
3
2
3
3
3
1
(2)
Mode-2 Matrix I 1
2
3
3
1
1
1
2
(3)
Mode-3 Matrix Fig. 1. Example of Unfolding the third order tensor to the three mode-n matrices 824 R. Xu and Y.-W. Chen fold the third order to their mode-n matrices. The mode-n product of a ten- sor
A ∈ R I 1 ×I 2 ×...×I N and a matrix U ∈ R J
×I n , denoted as A × n U, is an (I 1 × I 2 × ... × I n −1
n × I
n+1 × ... × I N ) tensor. Entries of the new ten- sor is defined by ( A ×
n U) i 1 i 2 ...i n −1 j n i n+1 ...i
N def
= i n a i 1 i 2 ...i n −1 i n i n+1 ...i N u j n i n . These entries can also be calculated by matrix product, B (n) = U
· A (n)
, B (n) is mode-n matrix of the tensor B = A ×
n U. The mode-n product has two properties. One can be expressed by ( A ×
n U) × m V = (
A × m V) × n U = A × n U × m V; and the other is ( A × n U) × n V = A × n (V · U). The inner product of two tensors A, B ∈ R I 1 ×I 2 ×...×I N is defined by A, B def
= i 1 i 2 ... i N a i 1 i 2 ...i
N b i 1 i 2 ...i N . The Frobenius-norm of a tensor A is defined by A def = A, A . The Frobenius-norm of a tensor can also be calculated from its mode-n matrix, A = A
(n) = tr(A (n) · A
T (n)
), tr() is the trace of a matrix. 3 Appearance Models Built by Generalized 3D-PCA An medical volume, such as MR volume, can be seen as a third order tensor. The three modes have their own physical meanings (coronal, sagittal and transversal axis respectively). Supposing there are a series of medical volumes collected from the same organs but different patients, there must be some common statistical information of their appearance (the voxel’s intensity) for these volumes. In this paper, we propose the multilinear algebra based algorithm named by G3D-PCA to build such appearance models for 3D medical volumes. We generalize the G3D-PCA for the appearance models building as follows. Given a series of the third order tensors with zero-mean 4 A i ∈ R
I 1 ×I 2 ×I 3 , i = 1, 2, ..., M , M i=1
A i = 0, we hope to find three matrices with orthogonal columns U ∈ R
I 1 ×J 1 , V
∈ R I 2 ×J 2 , W ∈ R I 3 ×J 3 , (J 1 < I 1 , J 2 < I 2 , J 3 < I 3 ) which is able to reconstruct A i from the corresponding dimension-reduced tensors B i ∈ R J 1 ×J 2 ×J 3 with least error. Each reconstructed tensor ˆ A i can be expressed by Eq. 1. The matrices U, V, W contain the statistical intensity information of the volume set A i
subspaces. ˆ A i = B i × 1 U × 2 V × 3 W (1)
We need to minimize the cost function S, shown by Eq. 2, to find the three matrices. S = M
A i − ˆ A i 2 = M i=1 A i − B i × 1 U × 2 V × 3 W 2 (2) 4 If the tensors do not have zero-mean, we can subtract the mean-value from each tensor to obtain a new series of tensors A i which have zero-mean, i.e. A i = A i − 1 M È M i=1 A i . Appearance Models for Medical Volumes with Few Samples 825
Algorithm 1. Iteration Algorithm to Compute the Matrices U opt
, V opt
, W opt
IN: a series of third order tensors, A i ∈ R I 1 × I 2 × I 3 , i = 1, 2, ..., M . OUT: Matrices U opt
∈ R I 1 ×J 1 , V opt ∈ R
I 2 ×J 2 , W
opt ∈ R
I 3 ×J 3 , (J
1 < I 1 , J 2 < I 2 , J 3
3 ) with orthogonal column vectors. 1. Initial values: k = 0 and U 0 , V 0 , W
0 whose columns are determined as the first J 1 , J 2 , J
3 leading eigenvectors of the matrices È M
(A i(1)
· A T i(1) ), È M i=1 (A i(2) · A T i(2) ), È M i=1 (A i(3) · A T i(3) ) respectively. 2. Iterate until convergence – Maximize S = È M i=1 C i × 1 U T 2 , C i = A i × 2 V T k × 3 W T k Solution: U whose columns are determined as the first J 1 leading eigenvec- tors of È M i=1 (C i(1) · C T i(1) ) Set U
k+1 = U.
– Maximize S = È M i=1 C i × 2 V T 2 , C i = A i × 1 U T k+1 × 3 W T k Solution: V whose columns are determined as the first J 2 leading eigenvec- tors of È M i=1 (C i(2) · C T i(2) ) Set V
k+1 = V.
– Maximize S = È M i=1 C i × 3 W T 2 , C i = A i × 1 U T k+1 × 2 V T k+1 Solution: W whose columns are determined as the first J 3 leading eigenvec- tors of È M i=1 (C i(3) · C T i(3) ) Set W
k+1 = W.
k = k + 1 3. Set U
opt = U
k , V
opt = V
k , W
opt = W
k . In Eq. 2, only the tensors A i are known. However, supposing the three ma- trices are known, the answer of B i to minimize Eq. 2 is merely the result of the traditional linear least-square problem. Theorem 1 can be obtained. Theorem 1. Given fixed matrices U, V, W, the tensors B i that minimize the cost function, Eq. 2, are given by B i
A i × 1 U T × 2 V T × 3 W T . With the help of Theorem 1, Theorem 2 can be obtained. Theorem 2. If the tensors B i
B i = A i × 1 U T × 2 V T × 3 W T , minimization of the cost function Eq. 2 is equal to maximization of the following cost function, S = M i=1
A i × 1 U T × 2 V T × 3 W T 2 . There is no close-form solution to simultaneously resolve the matrices U, V, W for the cost function S ; however if two of them are fixed, we can find the explicit solution of the other matrix which is able to maximize S . Lemma 1. Given the fixed matrices, V, W, if the columns of the matrix U are selected as the first J 1 leading eigenvectors of the matrix M i=1
(C i(1)
· C T i(1) ), C i(1) is the mode-1 matrix of the tensor C i = A i × 2 V T × 3 W T , the cost function S can be maximized. 826 R. Xu and Y.-W. Chen I 1
2
3
1
1
2
2
1
2
3
1 2 3 I 3
3 Fig. 2. Illustration of reconstructing a volume by the principal components B i and the three orthogonal bases of mode subspaces U opt
, V opt
, W opt
Lemma 2. Given the fixed matrices, U, W, if the columns of the matrix V are selected as the first J 2 leading eigenvectors of the matrix M i=1
(C i(2)
· C T i(2) ), C i(2) is the mode-2 matrix of the tensor C i = A i × 1 U T × 3 W T , the cost function S can be maximized. Lemma 3. Given the fixed matrices, U, V, if the columns of the matrix W are selected as the first J 3 leading eigenvectors of the matrix M i=1
(C i(3)
· C T i(3) ), C i(3) is the mode-3 matrix of the tensor C i = A i × 1 U T × 2 V T , the cost function S can be maximized. According to Lemma 1, Lemma 2 and Lemma 3, we can use an iteration algo- rithm to get the optimal matrices, U opt
, V opt
, W opt
, which are able to maximize the cost function S . This algorithm is summarized as Algorithm 1. In Algo- rithm 1, we terminate the iteration when the values of the cost function is not significantly changed in two consecutive times. Usually, the convergence is fast. According to our experiences, two or three times iterations are enough. Using the calculated matrices U opt , V
opt , W
opt , each of the volume A i
approximated by ˆ A i with least errors, where ˆ A i = B i × 1 U opt × 2 V opt
× 3 W opt and
B i = A i × 1 U T opt × 2 V T opt × 3 W T opt
. The approximation can be illustrated by Fig. 2. In 3D PCA, the three matrices U opt , V
opt , W
opt construct the bases on the third order tensor space; and the components of the tensor B i are the principal components. Table 1. Comparison of G3D-PCA and the two PCA-based methods, given the training volumes
A i , where A i ∈ R I 1 ×I 2 ×I 3 , i = 1, 2, ..., M Methods
Dimension of Maximal Number of Covariance Matrices Available Bases PCA–First Way (I 1 · I 2 · I
3 ) × (I 1 · I
2 · I
3 ) I 1 · I
2 · I
3 PCA–Second Way M × M
M − 1
Mode-1 : I 1 × I 1 G3D-PCA
Mode-2 : I 2 × I 2 I 1 · I 2 · I 3 Mode-3 : I 3 × I
3 Appearance Models for Medical Volumes with Few Samples 827
Compared to the PCA-based methods, the G3D-PCA has advantages in two aspects. Given M training volumes with the dimension of I 1 × I
2 × I
3 , the
covariance matrix for the first way to implement PCA has the a huge dimension of (I
1 ·I 2 ·I 3 ) ×(I 1 ·I 2 ·I 3 ). So, this method is suffered from huge computing cost. Different from the PCA-based method, G3D-PCA is to find three orthogonal bases in the three mode subspaces rather than find one orthogonal bases in the very long 1D vector space. The three bases are respectively obtained by calculating the leading eigenvectors from three covariance matrices of the mode subspaces. Since the three matrices have the dimension of I 1 × I
1 , I
2 × I
2 , I
3 × I
3 respectively, the calculation of eigenvectors is efficient. So G3D-PCA overcomes the problem of huge computing cost. In the other aspect, the G3D-PCA can obtain enough bases to represent the untrained volumes compared to the PCA- based method implemented by the second way. In theory, the maximal number of the tensor bases constructed from the bases of the three mode subspaces is I 1 ·I
· I 3 . So the number of the available bases is not limited on the number of training samples and this make the G3D-PCA has good performance on generalization even when few samples are used for training. Table 1 compares the differences between the G3D-PCA and the two PCA-based methods. 4 Experiments We use eighteen MR T1-weighted volumes of Vanderbilt [11] database to build the appearance models of the human brain. These eighteen volumes are collected from different patients, and their dimensions are 256 × 256 × 26. We choose one volume as the template and align the other seventeen volumes onto the template by similarity-transformation based registration. A 3D similarity-transformation has seven parameters, three for translations, three for rotation angles and one for scaling factor. Fig. 3 gives examples for three volumes of the registered data. Fig. 3. Example of registered volumes For these volumes, the first way to implement PCA is suffered from the huge burden of computing cost. For the 256 × 256 × 26 volumes, the unfolded vectors have the huge dimensions of 1703936, so the covariance matrix in the unfoled vector space has an extremely huge dimension of 1703936 × 1703936. Suppos- ing float type is used to store the covariance matrix in a computer, we need to allocate the memory of 10816GB. This is impossible for the current desktop PCs.
828 R. Xu and Y.-W. Chen The second way to implement PCA is suffered from the problem of bad perfor- mance for generalization. This can be demonstrated by the leave-one-out exper- iment, where the seventeen volumes are used to train the appearance model and the left one is reconstructed by the trained models for testing. Fig. 4 gives the result of leave-one-out experiment when the PCA is implemented in the second way. It should be noticed that all the 16 available bases are used in the recon- struction in this experiment. It can be seen that the quality of the reconstructed slices is not satisfied. The reconstructed slices are blurred and the tumor part on the 17-th can not be visualized from the reconstructed result. This problem is because the available bases are too few to represent the untrained vector in the unfolded vector space with large dimension. 7th-Slice 13th-Slice 17th-Slice RMSE: 0.197 Original Slices Reconstructed Slices
Fig. 4. Result of leave-one-out testing for the PCA-based method implemented by the second way The G3D-PCA is not suffered from the two problems. The covariance matrices for the three mode subspaces have the dimension of 256 ×256, 256×256 and 26× 26 respectively, so the eigenvectors of them can be calculated efficiently. In the other aspect, it is able to obtain enough bases to represent an untrained volumes in the space of third order tensor by G3D-PCA. This can also be demonstrated by the leave-one-out experiment. Fig. 5 gives the result of leave-one-out experiment for the G3D-PCA. The training samples and testing volume are the same as the experiment for PCA. The untrained volume is tried to be represented by three tensor bases with different dimensions of 50 · 50 · 15 = 37500, 75 · 75 · 20 = 112500, 100 · 100 · 20 = 200000. Since the number of the maximal available bases is 256 · 256 · 26 = 1703936, the compressing rates for the three cases are
50 ·50·15
256 ·256·26
≈ 2.2%, 75 ·75·20 256 ·256·26
≈ 6.6% and 100
·100·20 256
·256·26 ≈ 11.7% respectively. It can be seen that the quality of the reconstructed results become better and
Appearance Models for Medical Volumes with Few Samples 829
Original Slices 7th-Slice 13th-Slice 17th-Slice Dimension of the Bases: 15 50
RMSE: 7.69-E2 Dimension of the Bases: 20 75 75 RMSE: 5.36-E2 Dimension of the Bases: 20 100
100 RMSE: 4.23-E2 Fig. 5. Result of leave-one-out testing for the proposed G3D-PCA method better with the increase of the dimensions of the bases. The root mean square error (RMSE) is also calculated between the reconstructed volumes and the original volumes. Compared to the PCA-based result, it can be seen that the reconstructed results based on G3D-PCA is much better. The reconstructed results are clearer. Additionally, the tumor region is also reconstructed well. This experiment illustrates that the appearance models for medical volumes based on G3D-PCA has good performance on generalization even when the models are trained from few samples. 5 Conclusion We propose a method named as G3D-PCA based on multilinear algebra to build the appearance models for 3D medical volumes with large dimensions in this |
ma'muriyatiga murojaat qiling