|
Image fusion is the task of aligning two sets of images acquired from
the same object but at different time and most likely at different places wit
h different devices so that two different types of information of one object c
an be blended into a whole. Multimodality fusion in medical images resolves th
e problem of insufficient information provided by single modality. That is, on
ly either anatomical or functional information can be provided by a single ima
ging modality. Thus, it is urgently needed to integrate the anatomical (e.g.,
MR and CT) and functional images (e.g., SPECT and PET) of the human brain in t
he diagnosis of brain diseases. In this study, our focus of attention is on th
e three-dimensional fusion of the magnetic resonance (MR) and single photon em
ission computed tomography (SPECT) brain images. In essence, image fusion is a
problem of image registration. Assuming the human brain does not change its s
hape, registration of these two image sets can be modeled as a rigid body moti
on problem. We are required to find out three translation parameters (in x, y,
and z directions respectively) and three rotation parameters (about x-, y-, a
nd z-axes respectively) of the motion and then match the object to a fixed pos
ition. In multimodality fusion, the obstacles stem from the difficulty to extr
act feature pairs with exact correspondence. Most previously works adopted glo
bal features for registration. In particular, principle axes and brain surface
are widely used. In this work, a hybrid method composed of homologous feature
registration and surface fitting is developed. The centers of the two eyeball
s and the interhemispheric fissure are used as the feature sets for registrati
on. Therefore, we can solve the three translation parameters and two rotation
parameters based on a weighted least square registration criterion and derive
the partial transformation matrix. We then reslice the volume data. The remain
ing rotation parameter is determined by minimizing the surface distance in the
resliced 2-D images. After obtaining all these parameters, the transformation
matrix is determined, and the registration can be well done. Our approach is
very computationally efficient and achieves satisfactory result.
|