5.2 Methods
5.2.3 Joint Probabilistic Model of Position, Orientation and
The problem of joint registration and clustering of hybrid point sets comprising, multiple data features such as positions, orientations and scalar measures, is for- mulated as one of maximum likelihood estimation, using a hybrid mixture model (HMM). In the context of Application 1, the proposed HMM is used to approxi- mate the joint PDF of spatial positions (of voxel centroids), fibre orientations, and fractional anisotropy, derived from DTIs. By assuming voxel positions, fibre ori- entations, and FA values to be independent and identically distributed (i.i.d), for each subject and across multiple subjects, the joint PDF can be approximated as a
product of the individual conditional densities (Bishop,2006) for position, orien- tation and FA. Consequently, by considering all data points dki = Dk, from all K subjects, to be i.i.d. the conditional probability of an observation being sampled from an M -component hybrid mixture model is given by equation5.1a. Θp rep- resents the set of model parameters associated with the Student’s t-distributions S, used to model the distribution of voxel spatial positions; Θn represents the parameters of Watson distributions W (modelling fibre orientations); Θf denotes the set of parameters of the Gaussian distributions N (modelling FA); and πj = Π represents the set of mixture coefficients, of the hybrid mixture model. Here and throughout, subscript j = 1...M denotes mixture components and the choice of distributions indicated earlier will be justified later in this Section. Using equa- tion (5.1a) the log-likelihood function is formulated as shown in equation (5.1b), which defines the cost function to be optimised with respect to the mixture model parameters {Θp, Θn, Θf, Π} = Ψ, to jointly register and cluster the hybrid point set data Dk = D. p(dki|Θp, Θn, Θf) = M X j=1 πjS(xki|Θp)W(nki|Θn)N (fki|Θf) (5.1a) ln p(D|Ψ) = K X k=1 Nk X i ln p(dki|Θp, Θn, Θf) (5.1b) Pkijt = πjp(dki|Θ t p, Θtn, Θtf) M P j=1 πjp(dki|Θtp, Θtn, Θtf) (5.1c) Q(Ψt+1|Ψt) = K X k=1 Nk X i=1 M X j=1 Pkijt hln πj + Q(Θt+1pj |Θ t pj) + Q(Θ t+1 nj |Θ t nj) + Q(Θ t+1 fj |Θ t fj) i (5.1d) A tractable approach to maximising equation 5.1b is achieved using the expectation-maximisation (EM) framework (Dempster, Laird, and Rubin, 1977), which iteratively alternates between: the expectation (E)-step, which evaluates the mixture component membership probabilities as shown in equation5.1c(i.e. posterior probabilities Pt
kij), given an estimate of the model parameters Ψt, at the tth EM-iteration; and the maximisation (M)-step, which uses the computed pos- terior probabilities Pt
kij to maximise the conditional expectation of the complete- data-log-likelihood function Q (refer to equation5.1d), with respect to each model
parameter, resulting in revised estimates Ψt+1. As shown in equation5.1d, Q for the hybrid mixture model can be expressed as a sum of contributions from each distribution and corresponding data feature (i.e. position, orientation and FA), denoted: Q(Θt+1
p |Θtp), Q(Θt+1n |Θtn), Q(Θt+1f |Θ t
f), respectively. The complete algo- rithm for the proposed hybrid mixture model, to jointly register and cluster a group D of hybrid point sets, is summarized in Algorithm3. Subsequent sections discuss each probability distribution and estimation of their associated parame- ters, within the proposed framework, in more detail.
Algorithm 3Hybrid Mixture Model: HMM
Inputs: Group of hybrid point sets Dk=1..K, number of mixture components M, max.iterations
Outputs: Set of HMM parameters {Θp, Θn, Θf} = Ψ, soft correspondences
1: INITIALIZATION
2: Initialize M, σ2
p, σf2 using K-means clustering.
3: All πj = 1/M and νj = 3.0, κj = 1.0
4: procedure STAGE 1 EM: GR O U P-W I S E R I G I D
R E G I S T R A T I O N(Dk, Θp, Θn, Θf, Π, Tk) .EM initialized
5: while Iteration < max.iterations do
6: Compute Pkij .E-step
7: Update Rk, sk,tk .M-step 8: Update Θp, Θn, Θf .M-step 9: end while 10: return M, σ2, Υ, Π, T k 11: end procedure
12: Estimated mean template M, mixture coefficients Π and similarity transfor- mations {Tk}k=1...K initialise Stage 2.
13: procedure STAGE 2 EM: GR O U P-WI S E NO N-RI G I D
RE G I S T R A T I O N(Dk, Θp, Θn, Θf, Π, Tk, Wk) .EM non-rigid initialized 14: while Iteration < max.iterations do
15: Compute Pkij .E-step
16: Update Wk .M-step
17: Update σ2
p, νj, Θn, Θf .M-step
18: mpj remain fixed.
19: end while
20: Soft correpondences established using, Pkij, Θp, Θn, Θf, estimated follow- ing convergence.
21: returnSoft correspondences, Θp, Θn, Θf, Π, Wk
22: end procedure
Algorithm 3 is described in the context of Application 1. It was also em- ployed for Application 2, with minor modifications such as: employing Fisher distributions (in place of Watson distributions) to model surface normal vectors, and omitting the GMM component in the HMM (used to model scalar-valued quantities).