Algorithm 1.

Kalman multiatlas segmentation.

1:	Affine register Ji to Ji−1 for i = 1, …, N − 1 by minimizing the MSE, and record the transformations ${Ã}_i^{i-1}$. This can be done offline before I is available
2:	Initialize $P_0^{+}$, Q, and R
3:	When I is available, register I0 to I and record the transformations $A_0({\mathit{\alpha }}_0^{+})$
4:	for i = 1, 2, …, N − 1 do
5:	Predict transformation: ${\mathit{\alpha }}_i^{-}=F_i{\mathit{\alpha }}_{i-1}^{+}$
6:	Covariance matrix: $P_i^{-}=F_{i-1}{P}_{i-1}^{+}{F}_{i-1}^{\top }+Q$
7:	Get observation: affine register Ii to I by minimizing the dissimilarity measure and record the transformation βi
8:	Kalman gain: $K=P_i^{-}{(P_i^{-}+R_i)}^{-1}$
9:	Update state: ${\mathit{\alpha }}_i^{+}={\mathit{\alpha }}_i^{-}+K({\mathit{\beta }}_i-{\mathit{\alpha }}_i^{-})$
10:	Update covariance matrix: $P_i^{+}=(Id-K)P_i^{-}$
11:	end for