342
M. Abderrahim et al.
Clinical Datasets. To compare the performance of the studied methods, two
datasets were investigated. The first dataset, from the Retrospective Image Registration Evaluation (RIRE) project [17], consists of eight 3D triplets of PET,
MRI and CT images of brain. The MRI voxel size is of 1.25, 1.28 and 4 mm in
the x, y and z directions, respectively. The PET voxel size is (2.59 mm, 8 mm,
8 mm) in (x, y, z). MR images have been obtained using a Siemens SP 1.5 T
scanner, and the PET ones with a Siemens/CTI ECAT 933/0816 scanner. The
CT voxel size is equal to (0.65 mm, 0.65 mm, 4.0 mm) in (x, y, z). CT images
have been acquired using a Siemens Somatom Plus scanner. The second dataset
is provided by the Center for Addiction and Mental Health of Canada (CAMH).
It includes a collection of nine 3D images. For fixed MRI images, voxel dimensions along the x, y, and z axes are 0.86, 0.86, and 3 mm, respectively. These
images are captured by a Signa 1.5-T scanner from General Electric Medical
System. PET images are captured by a Scanditronix PET scanning system, GE
2048-15B, with x, y and z voxel dimensions equal to 2 mm, 2 mm and 6.5 mm,
respectively.
Evaluation Metrics. To quantify the accuracy of the studied methods, we
measured Normalized Cross-Correlation Coefficient (NCCC) (1) and Normalized
Mutual Information (NMI) (2) scores. NCCC evaluates the degree of similarity
between two medical images. In fact, cross correlation is less sensitive to linear
changes in amplitude and illumination in the images to be compared. A high
value of NCCC shows the high accuracy of the registration. Furthermore, NMI,
which is a measure of the quality of the registration, is defined in terms of the
entropy H of the image. It measures the proximity between the fixed source
image I f and the moving one I m . The more the value of normalized mutual
information is, the more the accuracy of the registration process is.
N CCC =
x=1
X
y=1
Y
(Im(x,y)−Im)(If (x,y)−I f )
x=1
X
y=1
Y
(Im(x,y)−Im)
2 (If (x,y)−I f )
2
,
(1)
NMI =
2(H(I f )+H(Im))
H(I f )+H(Im)+H(I f |Im)+H(Im|I f ) ,
(2)
where, H( ) and H ( | ) denote marginal and conditional entropies, respectively.
4 Results
We compare qualitatively and quantitatively the studied hybrid method against
the other aforementioned softwares for MRI/MRI, MRI/CT, and MRI/PET
images.
Qualitative Evaluation. Figures 4 and 5 show some samples of 3D slices before
and after mono- and multi-modal registrations. For the multimodal case, PET
and CT refer to the moving image and the MRI image is the fixed one. Obtained
results prove the performance of the Hybrid Method (HM) comparatively to
SPM, ITK-Snap and 3D Slicer (Fig. 4). Monomodal registration is similar to
M. Abderrahim et al.
Clinical Datasets. To compare the performance of the studied methods, two
datasets were investigated. The first dataset, from the Retrospective Image Registration Evaluation (RIRE) project [17], consists of eight 3D triplets of PET,
MRI and CT images of brain. The MRI voxel size is of 1.25, 1.28 and 4 mm in
the x, y and z directions, respectively. The PET voxel size is (2.59 mm, 8 mm,
8 mm) in (x, y, z). MR images have been obtained using a Siemens SP 1.5 T
scanner, and the PET ones with a Siemens/CTI ECAT 933/0816 scanner. The
CT voxel size is equal to (0.65 mm, 0.65 mm, 4.0 mm) in (x, y, z). CT images
have been acquired using a Siemens Somatom Plus scanner. The second dataset
is provided by the Center for Addiction and Mental Health of Canada (CAMH).
It includes a collection of nine 3D images. For fixed MRI images, voxel dimensions along the x, y, and z axes are 0.86, 0.86, and 3 mm, respectively. These
images are captured by a Signa 1.5-T scanner from General Electric Medical
System. PET images are captured by a Scanditronix PET scanning system, GE
2048-15B, with x, y and z voxel dimensions equal to 2 mm, 2 mm and 6.5 mm,
respectively.
Evaluation Metrics. To quantify the accuracy of the studied methods, we
measured Normalized Cross-Correlation Coefficient (NCCC) (1) and Normalized
Mutual Information (NMI) (2) scores. NCCC evaluates the degree of similarity
between two medical images. In fact, cross correlation is less sensitive to linear
changes in amplitude and illumination in the images to be compared. A high
value of NCCC shows the high accuracy of the registration. Furthermore, NMI,
which is a measure of the quality of the registration, is defined in terms of the
entropy H of the image. It measures the proximity between the fixed source
image I f and the moving one I m . The more the value of normalized mutual
information is, the more the accuracy of the registration process is.
N CCC =
x=1
X
y=1
Y
(Im(x,y)−Im)(If (x,y)−I f )
x=1
X
y=1
Y
(Im(x,y)−Im)
2 (If (x,y)−I f )
2
,
(1)
NMI =
2(H(I f )+H(Im))
H(I f )+H(Im)+H(I f |Im)+H(Im|I f ) ,
(2)
where, H( ) and H ( | ) denote marginal and conditional entropies, respectively.
4 Results
We compare qualitatively and quantitatively the studied hybrid method against
the other aforementioned softwares for MRI/MRI, MRI/CT, and MRI/PET
images.
Qualitative Evaluation. Figures 4 and 5 show some samples of 3D slices before
and after mono- and multi-modal registrations. For the multimodal case, PET
and CT refer to the moving image and the MRI image is the fixed one. Obtained
results prove the performance of the Hybrid Method (HM) comparatively to
SPM, ITK-Snap and 3D Slicer (Fig. 4). Monomodal registration is similar to