Page 131 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 3 Learning cardiac anatomy 103

Figure 3.3. Schematic visualization of the learning-based boundary deformation
step.

3.2.1.5 Nonrigid parametric deformation estimation
Given a bounding box computed with marginal space deep
learning (MSDL), one can use this information to estimate the
nonrigid deformation of the object. An initial estimation of the
shape is obtained by rigidly transforming the mean object shape
according to the estimated pose. This initial estimate is then itera-
tively refined using an active shape model based on deep learned
image features. In more detail, a boundary classifier is trained to
decide whether there is a boundary point at a given position and
under a given orientation. To solve this problem, Zheng et al. [31]
proposed steerable features combined with the PBT [203].
An effective alternative is to use cascaded SADNNs to learn
adaptive, sparse feature sampling patterns around the boundary.
These classifiers are trained to answer whether there is a bound-

ary point at a given position T = t x ,t y ,t z and orientation R =
φ x ,φ y ,φ z on the warping shape. The orientation is defined by the

normal for the respective shape point. The training is performed
with positive samples on the ground-truth boundary of training
examples (aligned with the shape normal) and negative samples at
different higher distances from the boundary. The sparse adaptive
patterns are essential in efficiently applying this classifier under
arbitrary orientations.
Fig. 3.3 shows the iterative algorithm proposed for segmenta-
tion. After each boundary estimation step, the deformed shape is
constrained to a subspace of shapes. Statistical shape modeling is
used for the constraint. The boundary estimation step and shape
constraint enforcement are applied in an iterative manner until
convergence.