State of the Art of Level Set Methods in Segmentation and Registration of Medical Imaging Modalities - Biomedical Image Analysis - page 78

active contours, threshold level set, canny-edge level set and Laplacian level
set methods.

2.4.3 Applications to Clinical Studies

2.4.3.1 Robust Adaptive Segmentation of 3D Medical Images

with Level Sets

This work was published by Baillard et al. in [30].
Method: The proposed method uses a 3D level set algorithm with the intro-
duction of an adaptive adjustment of the time step and the external propagation
force at each iteration. A region-based force is derived from intensity proba-
bility density functions over the data. Assumptions are made on the input data
which is modeled as a mixture of distributions. Mixture of Gaussian distribu-
tions for MRI and Gaussian and Rayleigh distributions for ultrasound data are
validated through two experiments. Each distribution defines a class c k through
a parameter vector that contains the distribution parameters and the proba-
bility p k that a voxel belongs to class c k . The parameters vector is estimated
from the data using the stochastic expectation-maximization (SEM) algorithm
[92], which is a stochastic version of the EM algorithm that utilizes probabilistic
learning stage. Advantages of the SEM over the EM algorithm include: (1) Only
an overestimation of the number of classes is required, (2) it is less dependent
on the initialization. The stopping criterion for the deformation process is based
on the stabilization of the average segmented volume size.
Experiments: Experiments were performed on brain MRI volumes. The
statistical model was initialized with seven classes.
1. A first experiment used simulated brain MRIs from the MNI group [93].
Brain MRI volumes of size (181 × 217 × 181) simulating WM, GM and CSF were
generated under noiseless conditions and three different combinations of noise
and inhomogeneities. The segmentation method was applied to extract together
GM and WM volumes. Initialization was performed by defining a large cube of
size (100 × 70 × 70) inside the data volume. Gaussian distribution parameters
for WM + GM were automatically estimated prior to segmentation. Quantita-
tive validation was performed using overlapping measurements [94] between
the result and the known ground truth on these phantom data sets. The mea-
sures included estimation of the number of true-positive (TP) true-negative (TN),