Medical Image Processing - Medical Image Processing: Advanced Fuzzy Set Theoretic Techniques - page 27

images of many subjects. It is called atlas registration. Normally, the registra-
tion of a patient image to an image of a normal patient image is referred to
as atlas registration.
Methods based on neural networks, genetic algorithm and fuzzy sets are
being used in medical image registration. In different computational aspects
of registration, a neural network such as Hopfield network and self-organizing
map is used. Genetic algorithm is used to find the exact or approximate solu-
tions to optimization and search procedure.
Majority of image registration methods involve the following steps:
Feature detection : In this step, features of distinctive objects such as
boundary, edges, contour and corner are manually or automatically
detected.
Feature matching : In this step, matching between the features detected in
the sensed image and reference image is computed. Various feature
descriptors and similarity measures along with spatial relationships
among the features are used for such purpose. In feature matching,
due to different imaging conditions, features cannot be similar. So,
the choice of feature descriptors and similarity measures is consid-
ered critically.
Model transformation : In this step, parameters that are used in the map-
ping function or in aligning the sensed image with the reference
image are estimated. The parameters are computed by means of fea-
ture correspondence. While mapping, prior information about the
acquisition process is also required.
Image transformation : The mapping function that is constructed is used
to transform the sensed image to register the images.
Fuzzy sets are widely used that considers the uncertainty in the data. It is
a mathematical tool which is used in the registration process to find reg-
istration transformation or to preprocess the image to remove noise and
enhance the features using image gradient or gamma correction before
segmenting/clustering the regions of the images [7,9,10] that are to be regis-
tered. The basic principle in most of the image registration algorithms is to
find the alignment between the two images by maximizing the fuzzy simi-
larity. Again, in a registration problem as, for example, in a feature-based
registration problem, fuzzy c means clustering (FCM) is used to detect the
regions defined by the users to find a match between the segments/regions
in reference and target images. FCM is used in feature space where features
of each pixel are considered. In fuzzy image registration, different fuzzy
methods are used for obtaining better results and these methods can be
found in many topics. Again these fuzzy methods can be extended using
intuitionistic fuzzy set and Type II fuzzy set theories for better registration
than fuzzy methods where more uncertainties are considered.