ScienceDirect - Nuclear Medicine and Biology ...

Login:  
Register
Athens/Institution Login Not Registered? 
User Name: Password:
Remember me on this computerForgotten password?

Home Browse Search My Settings Alerts Help
Quick Search   Title, abstract, keywords   Author e.g.  j s smith
  Journal/book title   Volume   Issue   Page         

Nuclear Medicine and Biology
Volume 34, Issue 4, May 2007, Pages 447-457
SummaryPlusFull Text + LinksPDF (529 K)   View thumbnail images |View full size images
               

View Record in Scopus
Cited By in Scopus (0)

doi:10.1016/j.nucmedbio.2007.02.008    
Copyright © 2007 Elsevier Inc. All rights reserved.
An automatic MRI/SPECT registration algorithm using image intensity and anatomical feature as matching characters: application on the evaluation of Parkinson‘s disease
Jiann-Der Leea,,, Chung-Hsien Huanga, Yi-Hsin Wengb,c, Kun-Ju Lind and Chin-Tu Chene,f
aDepartment of Electrical Engineering, Chang Gung University, Tao-Yuan 333, Taiwan, ROC
bMovement Disorder Session, Department of Neurology, Chang Gung Memorial Hospital and University, Taipei, Taiwan, ROC
cNeuroscience Research Center, Chang Gung Memorial Hospital and University, Taipei, Taiwan, ROC
dMolecular Image Center and Nuclear Medicine Department, Chang Gung Memorial Hospital, Lin-Kou, Taiwan, ROC
eDepartment of Radiology and Committee on Medical Physics, The University of Chicago, Chicago, IL, USA
fDivision of Medical Engineering Research, National Health Research Institutes, Zhunan, Miaoli, Taiwan, ROC
Received 24 May 2006;  revised 21 December 2006;  accepted 16 February 2007.  Available online 22 May 2007.
Abstract
Single-photon emission computed tomography (SPECT) of dopamine transporters with 99mTc-TRODAT-1 has recently been proposed to offer valuable information in assessing the functionality of dopaminergic systems. Magnetic resonance imaging (MRI) and SPECT imaging are important in the noninvasive examination of dopamine concentration in vivo. Therefore, this investigation presents an automated MRI/SPECT image registration algorithm based on a new similarity metric. This similarity metric combines anatomical features that are characterized by specific binding, the mean count per voxel in putamens and caudate nuclei, and the distribution of image intensity that is characterized by normalized mutual information (NMI). A preprocess, a novel two-cluster SPECT normalization algorithm, is also presented for MRI/SPECT registration. Clinical MRI/SPECT data from 18 healthy subjects and 13 Parkinson‘s disease (PD) patients are involved to validate the performance of the proposed algorithms. An appropriate color map, such as “rainbow,” for image display enables the two-cluster SPECT normalization algorithm to provide clinically meaningful visual contrast. The proposed registration scheme reduces target registration error from >7 mm for conventional registration algorithm based on NMI to approximately 4 mm. The error in the specific/nonspecific 99mTc-TRODAT-1 binding ratio, which is employed as a quantitative measure of TRODAT receptor binding, is also reduced from 0.45±0.22 to 0.08±0.06 among healthy subjects and from 0.28±0.18 to 0.12±0.09 among PD patients.
Keywords: Image registration;Normalized mutual information; MRI; SPECT; Parkinson‘s disease
Abbreviations: BR, specific/nonspecific binding ratio of 99mTc-TRODAT-1 to dopamine transporters; C, caudate nucleus; EBR, error in binding ratio; GUI, graphical user interface; L, left; MI, mutual information; MRI, magnetic resonance imaging; NMI, normalized mutual information; NMISB, proposed similarity metric based on NMI and SB; NMISB-A, NMISB with automatic MRI labeling; P, putamen; PD, Parkinson‘s disease; PET, positron emission tomography; R, right; SB, specific binding of nuclear medicine; SM, similarity metric; S.D., standard deviation; SPECT, single-photon emission computed tomography; ROI, region of interest; TRE, target registration error
Article Outline 1.Introduction 2.Materials and methods 2.1.Clinical MRI and SPECT data 2.2.Two-cluster SPECT normalization algorithm 2.3.Mutual information and NMI 2.4.Proposed similarity metric
3.Experiment results 3.1.Results of two-cluster SPECT normalization 3.2.Results of MRI/SPECT registration by NMISB
4.Discussion and conclusionReferences
1. Introduction
Medical images are typically characterized as functional or structural. Functional images reveal metabolic or neurochemical changes, and structural images show geometric and structural characteristics. Image registration allows these two classes of images to be fused to yield functional information localized over anatomical images. Accordingly, multimodality image registration[1] and[2], which involves bringing the two classes of images from different modalities into spatial correspondence, is becoming increasingly important in numerous applications[3] and[4].
The registration algorithms used for medical images are feature based or similarity based according to intrinsic features thereof. Feature-based approaches exploit anatomical knowledge to extract corresponding features or landmarks, which could appear as points[5], edges[6], contours or surfaces[7]. Feature-based approaches, relying on a relatively few feature points, are faster than similarity-based approaches. However, the performance of feature-based approaches typically depends on the success of the feature extraction strategy. Similarity-based approaches estimate spatial transformation from all intensity information in the images. They are implemented by iteratively updating the transformation by optimizing a similarity metric. Similarity-based approaches may be fully automatic and avoid difficulties associated with a feature extraction stage, but, unfortunately, computational burden is heavy.
Parkinson‘s disease (PD) is a progressive neurodegenerative disorder that is characterized by symptoms of akinesia, rigidity and tremors. PD affects approximately 0.15% of the total population but 0.5% of people older than 50 years. In the clinical application of PD[8] and[9], functional imaging techniques such as positron emission tomography (PET) or single-photon emission computed tomography (SPECT) provide useful information for detecting in vivo metabolic and neurochemical changes characteristic of PD pathology.
Recent research has demonstrated that 99mTc-TRODAT-1 is useful in diagnosing PD[10] because TRODAT-1 is a cocaine analog that can bind to such dopamine transporter sites as putamens and caudate nuclei at the presynaptic neuron membrane and can be easily labeled with 99mTc. The shapes of a healthy subject‘s putamens and caudate nuclei are quite similar to those of a patient on magnetic resonance imaging (MRI), but on a patient‘s 99mTc-TRODAT-1 brain SPECT images, the intensities of the putamens and caudate nuclei differ markedly from those of a healthy subject. Hence, in a clinical diagnosis of PD, the fusion image of MRI and SPECT is adopted to help physicians compute a quantitative index of the specific/nonspecific binding ratio (BR) of 99mTc-TRODAT-1 to dopamine transporters in the putamens and caudate nuclei. This index enables an accurate evaluation of complex presynaptic, postsynaptic and intrasynaptic dopaminergic phenomena in a living brain to better clarify the pathophysiology of PD and related syndromes. Accordingly, a suitable or even automatic algorithm for MRI/SPECT registration is useful in aiding clinical diagnoses.
Previous reviews of MRI/SPECT registrations[11],[12],[13] and[14] have all included intensity-based approaches based on the registration framework proposed by Maes et al.[15]. Grova et al.[11] generated simulated SPECT images from MRI using Monte Carlo simulations and studied the performance of numerous similarity metrics. Yokoi et al.[12] implemented 3D registration by dividing it into three 2D registrations performed sequentially in the transverse, sagittal and coronal planes. Thurfjell et al.[13] sequentially registered a single MRI image and two SPECT images obtained from a single individual. This sequential chain, called “registration circuits,” is adopted to evaluate how parameters and the sparseness of sampling influence the accuracy and robustness of the registration. Zhu and Cochoff[14] compared strategies for implementing MRI/SPECT registration using mutual information (MI) and a predefined standard implementation that uses trilinear interpolation, 64×64 bins, Powell‘s optimization and a multiresolution strategy. MI is an entropy-based similarity metric based on information theory[16] and[17], and previous studies have demonstrated that MI is useful for multimodality image registration[18]. However, in a 3D multimodality medical registration, misalignment can be so large with respect to the imaged field of view that statistical invariance is an important consideration. Therefore, Studholme et al.[19] proposed a new entropy-based similarity metric called normalized mutual information (NMI), which is insensitive to the overlapping region of the two images; NMI has also been employed in some investigations[20],[21] and[22].
Our clinical studies reveal that conventional MI-based registration algorithms, according to a reading physician, typically have a misalignment error of 3–12 mm (one to four MRI slices). Hence, an automatic MRI/SPECT registration algorithm based on anatomical features and the distribution of image intensity is proposed to increase the accuracy of MRI/SPECT registration. This work also presents a two-cluster SPECT normalization algorithm, which yields a clinically meaningful visual contrast. Validation studies that use brain MRI/SPECT data on 18 healthy subjects and 13 PD patients were performed.
2. Materials and methods
2.1. Clinical MRI and SPECT data
This study involved 31 subjects (18 healthy subjects and 13 PD patients) listed inTable 1. Each healthy subject underwent one MRI study and one 99mTc-TRODAT-1 SPECT study, while each patient underwent one MRI study and two SPECT studies to track the progress of one‘s disease.
Table 1.
Summary demographic data for each group Characteristic Healthy subjects (n=18) PD patients (n=13)
Gender [n (%)]
Male 7 (39) 9 (69)
Female 11 (61) 4 (31)
Age (years) (mean±S.D.) 45±17 63±11
Interval between two SPECT scans (m) (mean±S.D.) 21±11
 
High-resolution T2-weighted MRI was conducted using a VISION VB33D 1.5-T instrument (SIEMENS Medical Systems) with a repetition time of 4000 ms, an echo time of 90 ms and a slice thickness of 3 mm. Data series was obtained from axial view and covered approximately 36 images. Each image was a 256×256 matrix with a pixel size of 0.78×0.78 mm.
Before PD patients underwent SPECT scanning, all antiparkinsonian drugs and neuroleptic medications were discontinued for at least 12 h. A dose of 925 MBq of 99mTC-TRODAT-1 was injected intravenously. After 4 h, SPECT images were obtained using a Siemens MULTISPECT triple-head γ-camera with fan-beam collimators and 120 equally spaced projections over360°, taking 20 s per step. Individual images were reconstructed with backprojection using a ramp-Butterworth filter, with a cutoff of 0.3 cm−1 and an order of 10. SPECT data, obtained from the Chang Gung Memorial Hospital (Lin-Kou, Taiwan), were reconstructed in a 128×128×64 matrix that consisted of cubic voxels with volumes of 2.9 mm3.
The T2-weighted MR image inFig. 1A outlines the putamens and caudate nuclei of healthy subjects.Fig. 1B and C shows 99mTC-TRODAT-1 SPECT images scanned from a healthy subject and a PD patient, respectively. As presented inFig. 1, on 99mTC-TRODAT-1 SPECT images, specific binding regions (putamens and caudate nuclei) appear more intense in healthy subjects than in PD subjects. This difference can be quantified by the so-called “specific/nonspecific binding ratio (BR)”:
BR=(Csb−Cnsb)/Cnsb (1)
 
where Csb and Cnsb denote the mean count per voxel in the specific binding region (putamens or caudate nuclei) and in the nonspecific binding region (the occipital cortex), respectively. From the perspective of physicians, the BR quantifies the progression of PD, indicating the severity of the degeneration of specific binding regions.
Display Full Size version of this image (27K)
Fig. 1. Example of T2-weighted MRI and 99mTC-TRODAT-1 SPECT. (A) MRI image. (B) SPECT image from a healthy subject. (C) SPECT image from a PD patient.
 
2.2. Two-cluster SPECT normalization algorithm
Fig. 2 presents a 99mTC-TRODAT-1 SPECT image and its corresponding gray-level histogram. A gray-level histogram typically has three peaks — two relatively large peaks corresponding to the background and nonspecific binding regions, including the occipital cortex, cerebrum and cerebellum, and a smaller peak corresponding to specific binding regions, which are the putamens and caudate nuclei. The term “background” here is a set of nonintensity image voxels that do not belong to the subject under detection. The relative location of specific binding regions in the gray-level histogram depends on whether the subject scanned was healthy or was a PD patient, with specific binding regions closer to the nonspecific binding regions for a PD patient than for a healthy subject.
Display Full Size version of this image (36K)
Fig. 2. The gray-level histogram of a 99mTC-TRODAT-1 SPECT image.
 
At least three factors influence the intensity and contrast of SPECT images; these three factors are the interval between the injection of a radiotracer and scanning, the rate and degree of tissue uptake of the radiotracer and the rate and degree of the radiotracer‘s radiolabeled metabolites. Consequently, a two-cluster and two-step algorithm is proposed to adjust for these confounding variables, as follows, to normalize SPECT images among subjects:
Step 1 A Parzen window, which is also known as kernel density estimator, is utilized to smooth SPECT histogram in order to find two significant clusters, the background v0 and the nonspecific binding region v1. The kernel density function k(x) is an estimator of the probability density function of a data set and is defined by the following equation:
(2)
 
where G is a Gaussian function with a mean of 0 and a variance of 1, H is the maximum intensity of the gray-level histogram and w is the bandwidth parameter. Larger values of w yield a smoother estimator of density function, while smaller values result in a sharper estimator. Starting with a large value of w and reducing it yields a two-peak histogram. Following this step, the background and the nonspecific binding regions in a SPECT histogram can be located automatically.
Step 2 After v0 and v1 are located, each voxel with intensity i is normalized to a new intensity I according to the following equation:
(3)
 
where s is the normalization scale factor. Therefore, the gray-level value of the background becomes 0, that of the nonspecific binding regions becomes 1/s and the gray-level value of specific binding regions exceeds 1/s and is smaller than 1. Based on clinical trials of 99mTC-TRODAT-1 brain SPECT images, the gray-level value corresponding to the most intense voxel in specific binding regions is always less than four times that in nonspecific binding regions. Therefore, s is set to 4 herein.
2.3. Mutual information and NMI
The choice of similarity metric strongly influences similarity-based image registration. Holden et al.[23] and Skerl et al.[24] surveyed various similarity metrics for rigid registration. The similarity metric is based on image correlation (specified by mean square intensity differences, crosscorrelation or gradient correlation) or on image entropy, described by mutual information (MI) or NMI. Statistically, similarity metrics based on image correlations perform better with intramodality registrations, and those based on image entropy perform better with intermodality registrations[23].
MI, which is based on entropy information, has been employed in the context of image registration to measure how much information one image contains about the other, and this dependency makes MI very appropriate as a similarity metric of multimodality registration. MI is typically defined as in Eq.(4):
MI(A;B)=H(A)+H(B)−H(A;B) (4)
 
where H(A) and H(B) are the entropies of Images A and B, and H(A;B) is their joint entropy.
Joint entropy is given by:
(5)
 
(6)
 
(7)
 
where p(a;b), the so-called joint probability distribution, is the probability of co-occurrence of the gray-level value a in Image A and the gray-level value b in Image B at the same coordinates, and m1 and m2 represent the number of bins (the samples of gray level) in Images A and B, respectively. When the images are aligned, the joint probability should be less than that when they are misaligned.
Studholme et al.[19] proposed NMI as another entropy-based metric that is invariant throughout the region of overlap of the two images and is given by Eq.(8):
(8)
 
NMI has been demonstrated to be superior to MI[19] and is adopted as a part of a similarity metric in the registration algorithm herein.
2.4. Proposed similarity metric
Image registration processing is as follows. Given a fixed Image A and a movable Image B, after spatial transformation has been performed, both images have some overlapping voxels V in common where A,B,VRn , with n=2 for 2D image registration and n=3 for 3D image registration. More precisely, V={vi|i=1,…,n1, viA∩T(B)}, where n1 is the number of voxels in volume V and T denotes the transformation to transform the movable image onto the fixed image. The overlapping voxels of the fixed image and the transformed movable image are denoted as AV and BV, where AVA and BVT(B), respectively. The intensities of the corresponding voxel vi in AV and BV are fa(vi) and fb(vi), respectively. Registering A and B is equivalent to estimating the transformation T by optimizing a similarity metric SM using the following equation:
(9)
 
Fig. 3 presents the flowchart of the proposed algorithm, where MRI yields the fixed Image A and SPECT yields the movable Image B. First, regions of interest (ROI) VROIV, particular tissues corresponding to specific binding regions of SPECT, are labeled from the MRI manually or using a deformable registration algorithm. The second step involves the two-cluster SPECT normalization described in the preceding section. After B undergoes an initial transformation T to align the centers of masses of A and B, the similarity between the two images can be computed using a novel similarity metric called normalized mutual information and specific binding (NMISB). An optimization algorithm iteratively updates the transformation T until a maximum NMISB is reached.
Display Full Size version of this image (71K)
Fig. 3. The flowchart of the NMISB-based MRI/SPECT registration.
 
The proposed similarity metric NMISB considers anatomical features and the distribution of image intensity in the rigid registration framework and is given by the following equation:
NMISB=NMI+αSB (10)
 
where NMI represents the NMI of AV and BV and SB represents specific binding specified by the mean of voxel intensity within the ROI. The weighting parameter α represents a tradeoff between NMI and SB. SB is given by the following equation:
(11)
 
where n2 is the number of voxels in VROI.
All steps on the flowchart displayed inFig. 3 can be implemented automatically, except for the ROI labeling of MRI. Therefore, an automatic segmentation algorithm, which can label ROI from MRI, is required. A T2-weighted MRI atlas, which was downloaded from BrainWeb (http://www.bic.mni.mcgill.ca/brainweb) and then manually labeled, is adopted as a template to register elastically to the clinical MRI image. A B-spline deformable registration algorithm implementing the Insight Segmentation and Registration Toolkit[25], where the mean square metric[26] and LBFGS[27] are applied to the framework of the B-spline deformable registration, is employed. The mean square metric quantifies the mean squared pixelwise difference in intensity between two images; LBFGS is an optimization algorithm for solving a problem with numerous variables.
However, registering two 3D volumes takes half an hour or more on a standard PC. For reasons of practical time consumption, only a 2D deformable registration was adopted to label ROI in three continuous MR slices, in which the putamens and caudate nuclei were clearly visible: three MR slices and their corresponding slices in the MRI template are picked up, and then a 2D deformable registration is applied to each image pair. After 2D labeling results are integrated, they can be used for the proposed registration algorithm.Fig. 4 presents an example of ROI labeling. The MRI slice and a labeled MRI template are imported. The template and its labels are deformed by elastically registering them on the MRI slice. Finally, the MRI slice is labeled by laying the labels of the deformed template on it.
Display Full Size version of this image (49K)
Fig. 4. Illustration of ROI labeling by deformable registration.
 
3. Experiment results
3.1. Results of two-cluster SPECT normalization
Forty-four SPECT data sets, which comprised sets of data from 18 healthy subjects and dual sets from 13 PD patients, were used to validate the proposed two-cluster SPECT normalization algorithm. Normalized SPECT data sets were rescaled from the range [0,1] to the range [0,255] to compare them to the original SPECT data.Fig. 5 presents an example of the significant difference before and after SPECT normalization. Three SPECT images were selected from experimental SPECT data sets. After SPECT normalization, all of the images appeared similar. The average gray-level values before normalization were 23.49 in (A), 36.45 in (B) and 51.59 in (C). Those after normalization were 64.26 in (A), 57.18 in (B) and 57.21 in (C).
Display Full Size version of this image (32K)
Fig. 5. Illustration of the proposed two-cluster SPECT normalization. The average gray-level values before normalization were 23.49 in (A), 36.45 in (B) and 51.59 in (C). Those after normalization were 64.26 in (A), 57.18 in (B) and 57.21 in (C).
 
As shown inTable 2, the average gray-level value of each individual SPECT was computed initially, and then the mean and standard deviation (S.D.) were determined from the two groups (healthy subjects and patients). The mean of all SPECT data sets increased from 33.39 to 61.13, and the S.D. decreased from 8.12 to 3.87. As stated inSection 2.2, at least three factors influence the intensity and contrast of SPECT images, resulting in a large S.D., but they can be normalized using the proposed algorithm. Given a suitable color map, such as “rainbow,” for image display, the normalized SPECT image exhibits useful visual contrast for clinical readers. As shown inFig. 5, blue regions are the nonspecific binding regions, while yellow and red regions are specific binding regions.
Table 2.
The mean and S.D. of the gray-level histogram of SPECT data sets Mean S.D.
Nonnormalized Normalized Nonnormalized Normalized
Healthy subjects 34.57 60.94 8.33 3.78
Patients 32.58 61.27 8.03 3.88
Total subjects 33.39 61.13 8.12 3.87
 
3.2. Results of MRI/SPECT registration by NMISB
Numerous strategies for implementing the multimodality registration scheme proposed by Maes et al.[15], such as interpolation methods, histogram bin strategies, optimization algorithms, multiresolution or subsampling schemes and similarity metrics, have been considered[12],[13],[14],[26],[28] and[29]. Based on the results thus obtained, this study adopted trilinear interpolation, used Powell‘s optimization, exploited a subsampling scheme and assessed similarities using NMISB. Powell‘s optimization, a direction-set method in multidimensions, iteratively determines the optimum in each direction in the set[30]. However, this optimization method cannot guarantee a global optimal result. A three-level subsampling approach was added to prevent trapping at a local optimum and to accelerate convergence. The idea of a multilevel and subsampling approach is to register coarse images first and then to use previously registered result as the starting point in further fine-image registration. The coarse image is obtained by subsampling the original input image. In this work, the x:y:z sampling rates were 4:4:2 on the first level, 2:2:1 on the second level and 1:1:1 on the third level. The number of bins was 64 on the first level, 128 on the second level and 256 on the third level. MRI and SPECT data sets were rescaled to the same number of bins at each level.
The difficulty in validating registration algorithms that are applied to clinical data is that the ground truth is unknown. Thurfjell et al.[13] evaluated MRI/SPECT registration that involves the sequential registration of three scans (one MRI scan and two SPECT scans on different occasions). Grova et al.[11] and Yokoi et al.[12] utilized simulated SPECT data from segmented MRI as a gold standard to evaluate registration algorithms. Since modeling the full complexity of human data is difficult, simulated data should not be expected to behave as realistically as actual clinical data. However, Pfluger et al.[31] demonstrated that errors were small when experienced radiologists interactively and manually registered MRI/SPECT using a suitable graphical user interface (GUI). Therefore, the gold standard adopted in this study are the interaction registration results.
Fig. 6 shows the GUI developed in C++ for interactive registration. Sagittal, coronal and transverse views of a color-coded SPECT image overlaid onto a gray-level MR image are displayed simultaneously on a split screen. For registration, the three orthogonal views of a SPECT image can be freely rotated and/or translated simultaneously; the resulting transformation is denoted as TIR. However, such manual registration depends strongly on an operator‘s experiences and takes a very large computation time. To expedite the process, an initial transformation TINIT was conducted midway between NMI and NMISB transformations, denoted TNMI and TNMISB, respectively. Since the operator has no a priori information about the results of registering by NMI or NMISB, manual registration relies only on the experience of the expert, regarded as a gold standard.
Display Full Size version of this image (104K)
Fig. 6. The GUI for interactive registration.
 
The target registration error (TRE) was adopted to evaluate the spatial error of transformations between registration algorithm T and interactive registration TIR. The registration error calculates the mean of spatial distance between 5000 voxels at the corresponding points on two images. The following equation gives the TRE:
(12)
 
where T can be either TNMI or TNMISB, pi represents points selected randomly in the brain and ||·|| is Euclidean distance.
First, the effect of the tradeoff parameter α in Eq.(10) was evaluated. All data sets were registered following the flowchart presented inFig. 2 for various α. The ROI in MRI were labeled manually to prevent mislabeling by an automatic labeling algorithm.Fig. 7 plots the mean TRE in MRI/SPECT registration as a function of α. When α is zero, NMISB is equivalent to NMI. The curve declines steadily until α reaches 1, and then both curves rise slowly as α increases. Therefore, α=1 is the appropriate choice herein.
Display Full Size version of this image (19K)
Fig. 7. The TRE of MRI/SPECT registration versus the tradeoff parameter α.
 
When α is set to 1, the registration results of 18 healthy subjects and 13 PD patients are as plotted inFig. 8 andFig. 9, where NMISB-A refers to automatic labeling of the ROI by elastically registering them with a labeled MRI template. The mean and S.D. of TRE among healthy subjects and patients are then calculated and are presented inTable 3. If the mean is close to 0, then the result is considered to be accurate. If the S.D. is small, then the result is regarded as precise. The results reveal that NMISB outperforms NMI and NMISB-A. Since only an automatic 2D ROI labeling, which was less accurate than manual labeling, was considered, NMISB-A was worse than NMISB but still better than NMI.
Display Full Size version of this image (16K)
Fig. 8. The TRE of the MRI/SPECT registration of 18 healthy subjects.
 
Display Full Size version of this image (20K)
Fig. 9. The TRE of the MRI/SPECT registration of 13 PD patients. Each PD patient has two SPECT data sets scanned on different dates. The digit in front of and behind the underline stands for the number of PD patients and the corresponding SPECT studies, respectively.
 
Table 3.
The mean and S.D. of TRE by registration with NMI, NMISB and NMISB-A among healthy subjects and PD patients Healthy subjects Patients
Mean (mm) S.D. (mm) Mean (mm) S.D. (mm)
NMI 7.71 2.59 7.50 2.19
NMISB 4.13 1.98 4.85 1.61
NMISB-A 5.31 1.99 5.91 1.72
 
The specific/nonspecific binding ratio (Eq. (1)) calculating the intensity ratio of specific and nonspecific regions is an important indicator in distinguishing healthy subjects from patients. Generally, the binding ratio of a healthy subject is between 1.5 and 2.5, and that of a patient is <1.5[10].Table 4 reveals the mean and S.D. of the error in the binding ratio (EBR; Eq.(13)) of n MRI/SPECT registration couples (n=18, healthy subjects; n=26, patients), where R and L stand for right and left, and C and P stand for caudate nucleus and putamen, respectively:
EBR=|BRIR−BR| (13)
 
where BRIR denotes the binding ratio obtained by interactive registration and BR denotes the binding ratio obtained by NMI, NMIBI or NMISB-A. The result inTable 4 reveals that the proposed registration scheme reduces the EBR from 0.45±0.22 to 0.08±0.06 among healthy subjects and from 0.28±0.18 to 0.12±0.09 among PD patients.
Table 4.
The mean and S.D. of EBR among healthy subjects and PD patients Healthy subjects Patients
NMI NMISB NMISB-A NMI NMISB NMISB-A
R, C 0.49±0.19 0.09±0.06 0.16±0.08 0.36±0.20 0.16±0.11 0.21±0.13
L, C 0.43±0.24 0.11±0.09 0.19±0.13 0.44±0.29 0.16±0.11 0.19±0.14
R, P 0.44±0.24 0.05±0.05 0.11±0.14 0.16±0.11 0.09±0.07 0.12±0.08
L, P 0.42±0.23 0.07+0.06 0.12±0.12 0.16±0.12 0.08±0.07 0.13±0.07
Average 0.45±0.22 0.08±0.06 0.15±0.12 0.28±0.18 0.12±0.09 0.16±0.11
 
4. Discussion and conclusion
NMI was adopted as a similarity-based metric and is suitable for multimodality registration. The NMI is measured from the two-dimensional joint histogram of the image pair being registered. Although NMI is extensively used in intermodality registration, our clinical studies indicate that NMI-based MRI/SPECT registrations still suffer from nonnegligible errors in the judgment of reading physicians. Therefore, specific binding (SB), which is regarded as a feature-based similarity metric, is utilized with the NMI involved in the MRI/SPECT image registration to distinguish automatically healthy subjects from PD patients. In addition, a two-cluster SPECT normalization algorithm is proposed. SPECT images are normalized by relocating two significant peaks as background and nonspecific binding regions in the gray-scale histogram.
Comparing experimental results with the results of the conventional NMI-based registration algorithm reveals that the NMISB-based registration scheme reduces the TRE from >7 mm to approximately 4 mm and reduces the error in the specific/nonspecific binding ratio of 99mTc-TRODAT-1 to dopamine transporters from 0.45±0.22 to 0.08±0.06 among healthy subjects and from 0.28±0.18 to 0.12±0.09 among PD patients. NMISB has high accuracy, especially in the ROI. Comparing NMISB with NMISB-A reveals that the latter was less effective than the former because the ROI in MRI were only labeled with reference to some particular 2D slices. Therefore, a more reliable 3D brain labeling algorithm will improve the results of the use of NMISB-A.
An automatic 3D brain MRI/SPECT registration algorithm was implemented based on a novel similarity metric, NMISB. The accuracy and precision of the proposed approach were evaluated using data from 18 healthy subjects and 13 PD patients. The registration process was performed on a standard PC with Pentium 4 mobile, a 1.6-GHz central processing unit and 512 MB of random access memory (RAM) with an ATI Mobility Radeon 7500 video card with 16 MB of RAM, and took a total of approximately 12–10 min for labeling and 2 min for optimization. Moreover, a novel two-cluster normalization algorithm applied to the gray-level histogram to normalize SPECT images was proposed to enhance the background, nonspecific binding regions and specific binding regions. With a suitable color map, normalized SPECT images provide a useful visual contrast for clinical readers. Future work will consider a more precise MRI labeling algorithm and will apply this system to more clinical data.
Appendix A
List of symbols
A
Fixed image (MRI image herein)
AV
Voxel set of the fixed image at the overlapping volume V
B
Movable image (SPECT image herein)
BV
Voxel set of the movable image at the overlapping volume V
BRIR
Binding ratio obtained by iterative registration
Cnsb
Mean count per voxel in the nonspecific binding region
Csb
Mean count per voxel in the specific binding region
fa
Intensity of voxel in AV
fb
Intensity of voxel in BV
G
Gaussian function
H
Entropy
H(;)
Joint entropy
i
Voxel intensity before SPECT normalization
I
Voxel intensity after SPECT normalization
k
Kernel density function
n1
The number of voxels in V
n2
The number of voxels in VROI
p
Point selected randomly in the brain
s
Normalization scale factor
T
Spatial transformation
TIR
Transformation of interactive registration
TINIT
Initial transformation for manual registration
TNMI
Transformation obtained by NMI registration
TNMISB
Transformation obtained by NMISB registration
V
The overlapping volume of the fixed image and the movable image
VROI
ROI (putamens and caudate nuclei herein)
v
Voxel belonging to the overlapping volume V
v0
Peak stands for the background of the gray-level histogram of SPECT image
v1
Peak stands for the nonspecific binding region of the gray-level histogram of SPECT image
w
Bandwidth parameter of Parzen window
α
Tradeoff between NMI and SB
References
[1] J.B.A. Maintz and M.A. Viergever, A survey of medical image registration, Med Image Anal 2 (1998), pp. 1–36.Abstract |Abstract + References |PDF (248 K) |View Record in Scopus |Cited By in Scopus
[2] In: J.V. Hajnal, D.L.G. Hill and D.J. Hawkes, Editors, Medical image registration, CRC Press, Boca Raton (Fla) (2001).
[3] H.K. Pohjonen, S.E. Savolainen, P.H. Nikkinen, V.P. Poutanen, E.T. Korppi-Tommola and B.K. Liewendahl, Abdominal SPECT/MRI fusion applied to the study of splenic and hepatic uptake of radiolabeled thrombocytes and colloids, Ann Nucl Med 10 (1996), pp. 409–417.View Record in Scopus |Cited By in Scopus
[4] G.J. Forster, C. Laumann, O. Nickel, P. Kann, O. Rieker and P. Bartenstein, SPET/CT image co-registration in the abdomen with a simple and cost-effective tool, Eur J Nucl Med Mol Imaging 30 (2003), pp. 32–39.View Record in Scopus |Cited By in Scopus
[5] F. Bokstein, Thin-plate splines and the atlas problem for biomedical images, Inf Process Med Imaging (1991), pp. 326–342.
[6] W. Kerwin and C. Yuan, Active edge maps for medical image registration, Proceeding of the SPIE — The International Society for Optical Engineering (2001), pp. 516–526.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[7] A. Dale, B. Fischl and M. Sereno, Cortical surface-based analysis: I. Segmentation and surface reconstruction, Neuroimage 9 (2) (1999), pp. 179–194.Abstract |Abstract + References |PDF (1029 K) |View Record in Scopus |Cited By in Scopus
[8] J.B.A. Habraken, J. Booij, P. Slomka, E.B. Sokole and E.A.V. Royen, Quantification and visualization of defects of the functional dopaminergic system using an automatic algorithm, J Nucl Med 40 (1999), pp. 1091–1097.View Record in Scopus |Cited By in Scopus
[9] G.E. Fakrhi, S.C. Moore, P. Maksud, A. Aurengo and M.F. Kijewski, Absolute activity quantitation in simultaneous 123I/99mTc brain SPECT, J Nucl Med 42 (2001), pp. 300–308.
[10] Y.H. Weng, T.C. Yen, M.C. Chen, P.F. Kao, K.Y. Tzen and R.S. Chen et al., Sensitivity and specificity of 99mTc-TRODAT-1 SPECT imaging in differentiating patients with idiopathic Parkinson‘s disease from healthy subjects, J Nucl Med 45 (2004), pp. 393–401.View Record in Scopus |Cited By in Scopus
[11] C. Grova, A. Biraben, J.M. Scarabin, P. Jannin, I. Buvat and H. Benali et al., A methodology to validate MRI/SPECT registration methods using realistic simulated SPECT data, Medical image computing and computer-assisted intervention MICCAI 2001. Lecture notes in computer science 2208, Springer, Netherlands (2001), pp. 275–282.
[12] T. Yokoi, T. Soma, H. Shinohara and H. Matsuda, Accuracy and reproducibility of co-registration techniques based on mutual information and normalized mutual information for MRI and SPECT brain images, Ann Nucl Med 18 (2004), pp. 659–667.View Record in Scopus |Cited By in Scopus
[13] L. Thurfjell, Y.H. Lau, J.L.R. Andersson and B.F. Hutton, Improved efficiency for MR-SPET registration based on mutual information, Eur J Nucl Med 27 (2000), pp. 847–856.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[14] Y.M. Zhu and S.M. Cochoff, Influence of implementation parameters on registration of MR and SPECT brain images by maximization of mutual information, J Nucl Med 43 (2002), pp. 160–166.View Record in Scopus |Cited By in Scopus
[15] F. Maes, A. Collignon, D. Vandermeulen, G. Marchal and P. Suetens, Multimodality image registration using mutual information, IEEE Trans Med Imaging 16 (1997), pp. 187–198.Full Text via CrossRef
[16] A. Collignon, F. Maes, D. Delaere, D. Vandermeulen, P. Suetens and G. Marchal, Automated multi-modality image registration based on information theory. In: Y. Bizais, C. Barillot and R. Di Paola, Editors, Information processing in medical imaging, Kluwer Academic Publishers, Dordrecht (1995), pp. 263–274.
[17] P. Viola and W.M. Wells III, Alignment by maximization of mutual information, Proceedings of the 5th International Conference on Computer Vision. (1995), pp. 16–23 Cambridge (Mass).Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[18] J.P.W. Pluim, J.B.A. Maintz and M.A. Viergever, Mutual-information-based registration of medical images: a survey, IEEE Trans Med Imaging 22 (2003), pp. 986–1004.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[19] C. Studholme, D.L.G. Hill and D.J. Hawkes, An overlap invariant entropy measure of 3D medical image alignment, Pattern Recognition 32 (1999), pp. 71–86.Abstract |PDF (745 K)
[20] E.R. Denton, L.I. Sonoda, D. Rueckert, S.C. Rankin, C. Hayes and M.O. Leach et al., Comparison and evaluation of rigid, affine, and nonrigid registration of breast MR images, J Comput Assist Tomogr 23 (1999), pp. 800–805.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[21] P.E. Radau, P.J. Slomka, P. Julin, L. Svensson and L.-O. Wahlund, Evaluation of linear registration algorithms for brain SPECT and the errors due to hypoperfusion lesions, Med Phys 28 (2001), pp. 1660–1668.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[22] C. Stodholme, E. Novotny, I.G. Zubal and J.S. Duncan, Estimating tissue deformation between functional images induced by intracranial electrode implantation using anatomical MRI, Neuroimage 13 (2001), pp. 561–576.
[23] M. Holden, D.L.G. Hill, E.R.E. Denton, J.M. Jarosz, T.C.S. Cox and T. Rohlfing et al., Voxel similarity measures for 3-D serial MR brain image registration, IEEE Trans Med Imaging 19 (2) (2000), pp. 94–102.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[24] D. Skerl, B. Likar and F. Pernus, Evaluation of nine similarity measures used in rigid registration, Proceedings of the 17th International Conference on Pattern Recognition Vol. 3 (2004), pp. 794–797.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[25] Insight Segmentation and Registration Toolkit (ITK), National Library of Medicine segmentation and registration toolkithttp://www.itk.org (2004).
[26] J.V. Hajnal, N. Saeed, E.J. Soar, A. Oatridge, I.R. Young and G.M. Bydder, A registration and interpolation procedure for subvoxel matching of serially acquired MR images, J Comput Assist Tomogr 19 (2) (1995), pp. 289–296.View Record in Scopus |Cited By in Scopus
[27] D.C. Liu and J. Nocedal, On the limited memory BFGS method for large-scale optimization, Math Program 45 (1989), pp. 503–528.MathSciNet |Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[28] F. Maes, D. Vandermeulen and P. Suetens, Comparative evaluation of multiresolution optimization strategies for multimodality image registration by maximization of mutual information, Med Image Anal 3 (1999), pp. 373–386.Abstract |Abstract + References |PDF (270 K) |View Record in Scopus |Cited By in Scopus
[29] C. Studholme, D.L.G. Hill and D.J. Hawkes, Automatic three-dimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures, Med Phys 24 (1997), pp. 25–35.Full Text via CrossRef |View Record in Scopus |Cited By in Scopus
[30] W.H. Press, B.P. Flannery, S.A. Teukolsky and W.T. Vetterling, Numerical recipes in C (2nd ed.), Cambridge University Press (1997).
[31] T. Pfluger, C. Vollmar, A. Wismuller, S. Dresel, F. Berger and P. Suntheim et al., Quantitative comparison of automatic and interactive methods for MRI-SPECT image registration of the brain based on 3-dimensional calculation of error, J Nucl Med 41 (2000), pp. 1823–1829.View Record in Scopus |Cited By in Scopus
Corresponding author. Department of Electrical Engineering, Chang Gung University, Tao-Yuan 333, Taiwan, ROC. Tel.: +886 3 2118800x5316; fax: +886 3 2118026.
Nuclear Medicine and Biology
Volume 34, Issue 4, May 2007, Pages 447-457
Home Browse Search My Settings Alerts Help
About ScienceDirect  | Contact Us  | Terms & Conditions  | Privacy Policy

Copyright © 2007Elsevier B.V. All rights reserved. ScienceDirect® is a registered trademark of Elsevier B.V.