Abstract

Introduction

Many movement disorders are related first to Parkinson’s disease, with symptoms that include essential tremor, rigidity, and akinesia, and less frequently to dystonia that is a syndrome of sustained muscle contractions producing writhing movements and abnormal postures, to essential tremor or to Tourette syndrome. When pharmaceutical treatments lose effectiveness, such disorders may require functional neurosurgery. In such surgeries, deep brain electrodes are implanted to inhibit the activity of target structures such as the subthalamic nucleus (STN), medial globus pallidus (GPm), and caudal part of the thalamic ventral lateral nucleus (VLc) [1–3]. Following surgery, motor effects and neuropsychological or psychiatric secondary side effects of the stimulation may occur depending on the location and trajectory of the electrodes [3–5]. The patient’s surgical outcome is related to the accuracy of nucleus targeting during surgery, and the accuracy of electrode implantation depends on different steps such as image preprocessing, intraoperative clinical testing of patients, and intraoperative electrophysiological recordings.

Brain atlases and atlas-based segmentation techniques have been developed to facilitate the accurate targeting of deep brain structures in patients with movement disorders [6–11]. Recently, several authors have demonstrated improved targeting accuracy by using multimodal databases composed of anatomical and functional data [12–14]. In these studies, basal ganglia segmentations, the coordinates of previously implanted electrodes, and functional data (mainly intracerebral electrophysiological recordings) were aligned in a common space and overlaid on a magnetic resonance (MR) image template. Although the use of multimodal images has been shown to be an effective tool in surgical planning, very few studies have correlated clinical postoperative data with anatomical data [13].

Developing a tool to improve targeting accuracy and the aggregation of anatomical and clinical data requires several steps, including the construction of a template to define a common stereotactic space, accurate and automated segmentation of target structures on the template, registration of the patient’s multimodal image data, and alignment of the patient’s image data into a common space for comparison with electrode placement. In this paper, we focus on the segmentation step by investigating the accuracy of three methods of segmenting the basal ganglia and deep brain structures of patients with Parkinson’s disease. The goal of this work was to compare these methods and determine which method was enough accurate and clinically applicable to obtain segmentation of target structures in further patients with Parkinson’s disease waiting for deep brain stimulation.

First, we used two segmentation methods that are nonlinear registration-based approaches involving a population-specific, anatomically labeled template. This kind of approach is widely used and has demonstrated good performance in the segmentation of a large variety of structures [15]. In our study, we constructed a population-specific template [16], which has been shown to be one of the most effective methods of achieving accurate segmentation of subjects with pathology. To achieve automated segmentation, the basal ganglia and other deep brain structures useful in the planning of deep brain stimulation were first manually delineated on the template, which was then nonlinearly warped onto the patient’s MR image. The segmentation of the target structures on the MR image could thus be obtained by mapping the template labels through the estimated transformation onto the patient’s MR image. We compared two nonlinear algorithms to achieve this task: the well-known Automatic Nonlinear Image Matching and Anatomical Labelling (ANIMAL) method [15] and the more recent Symmetric Image Normalization (SyN) technique [17]. The third segmentation method evaluated here was a recently published, patch-based, label fusion technique that has demonstrated very good performance for hippocampus and ventricle segmentation [18]. We compared the results if the three segmentation methods in terms of anatomical similarity with the template segmentation and discussed interest of these methods in clinical practice for deep brain stimulation in patients with Parkinson’s disease.

Materials and methods

Construction of the Parkinson’s templates

We used the procedure detailed by Fonov et al. [16] to construct multispectral, multi-subject, unbiased nonlinear average Parkinson’s disease templates. The 57 T1w images were used to create the T1w Parkinson’s template (PD_T1_57), and the 42 T2w images from the same subjects were used to create the T2w Parkinson’s template (PD_T2_42), using the steps described below. All images were converted to the Medical Image NetCDF (MINC) format [19], the format used at our institute for image processing.

The following preprocessing steps were applied to all MR scans prior to constructing the template: (1) the program N3 was used to correct for image intensity nonuniformity [20]; (2) inter-subject linear image intensity normalization was achieved using linear histogram scaling to the International Consortium Brain Mapping (ICBM152) template [16] following the technique of Nyul et al. [21]; (3) intra-subject inter-modality volume registration of T2w and T1w MR images was achieved with a rigid-body transformation (i.e., only 3 translations and 3 rotations) and a mutual-information cost function [22] implemented in the mritoself tool from the MINC mni_autoreg software package [23]; (4) inter-subject spatial alignment to the Talairach-like ICBM152 stereotactic space was achieved with a linear (nine parameter) transformation estimated with mritotal; and (5) a patient-specific brain mask was created by applying Brain Extraction Tool (BET) from the Oxford centre of Functional MRI of the brain Functional Software Library (FSL) to the T1w data [24]. Then, the T1w and T2w data for all subjects were intensity normalized and spatially aligned within the ICBM152 stereotactic space.

After preprocessing, the average high-signal-to-noise T1w and T2w templates were created using an iterative unbiased method [16] (see Figure 1). At each iteration, all T1w datasets are nonlinearly registered to an intermediate target and then averaged together on a voxel-by-voxel basis. The inverse of the average nonlinear transformation is applied to the intermediate template to remove any potential spatial bias. This procedure is repeated multiple times in a hierarchical fashion, gradually increasing the resolution of the nonlinear transformations, with number of iterations estimated empirically to achieve good results from a similar dataset. The result is an unbiased, high-contrast, highly detailed T1w average MR image template. Nonlinear registrations were obtained with the ANIMAL procedure [15], chosen because of our extensive experience with this technique. Following the unbiased non-linear registration, the transformations estimated on the T1w data were applied to the patient’s corresponding T2w data to create the unbiased T2w average template. The templates (PD_T1_57 and PD_T2_42) enable visualization of deep brain structures such as amygdala or hippocampus and basal ganglia structures such as subthalamic nucleus or thalamus. These structures have either T1w contrast or T2w contrast, or are visible in both modalities, and can be manually segmented on each slice of the PD_T1_57 and/or PD_T2_42 templates.

Open in a separate window

Fig. 1

Parkinson’s templates obtained from 57 T1w and 42 T2w average MR images of patients with Parkinson’s disease. From left to right: Axial, coronal, coronal and sagittal slices of the T1w (top) and T2w (bottom) templates.

Manual segmentation

Manual segmentation of the template was achieved using the MNI Display software (Montreal Neurological Institute, Canada, www.bic.mni.mcgill.ca/ServicesSoftware) by the first author, Claire Haegelen. Segmentation consisted of manual delineations of 24 structures bilaterally on each slice of the PD_T1_57 and PD_T2_42 templates. The Display software allows the user to paint the structure in any chosen slice (e.g., transverse) with simultaneous update in the other two views (e.g., coronal and sagittal), enabling consistent segmentation in 3D while viewing either template. The 24 structures were identified with the help of anatomical atlases [9,10,25]. The structures were defined bilaterally and included the amygdala, hippocampus, caudate nucleus, putamen, GPm and lateral globus pallidus, thalamus, medial and lateral geniculate bodies, red nucleus, substantia nigra, and STN (see Figure 2). All the structures were visible on the PD_T2_42 template. While a rapid 36-slice T2w image, centered on the thalamus was acquired on the morning of the PD surgery, it did not permit to identify for segmentation the entire caudate nucleus, amygdala and hippocampus. For this reason, these structures were segmented on the PD_T1_57 template where they were visible. Display allows both T1w and T1w images to be visible simultaneously to ensure label consistency. It took 120 hours to manual segment the structures on the templates.

Open in a separate window

Fig. 2

Anterior (top left), lateral left (top right), posterior (bottom left) and superior (bottom right) 3D views of the manual segmentation of the Parkinson’s templates. The painted structures included bilateral amygdala (brown), hippocampus (white), putamen (dark blue), GPm (yellow), lateral globus pallidus (dark pink), caudate nucleus (light pink), thalami (light blue), substantia nigra (dark violet), STN (dark green), red nucleus (light violet), and medial geniculate body (light green).

To validate the automated template-based segmentation procedure, a gold standard labelling of patient data was required. Fourteen bilateral structures were manually painted on the T1w and T2w images of 10 patients with Parkinson’s disease. The images used were a subset of those used for the Parkinson’s templates after N3 nonuniformity correction [20]. Using the Display program, the structures labelled included the amygdala, hippocampus, putamen, thalamus, red nucleus, substantia nigra, and STN (see Figure 3). As for the template, the amygdala and hippocampus were segmented on the patients T1w image, an the remaining structures on the T2w image. We chose these nine structures for their different sizes: large (e.g., thalamus and hippocampus), intermediate (e.g., amygdala and putamen), and small (e.g., red nucleus, substantia nigra, and STN). The five remaining structures (caudate nucleus, GPm and lateral globus pallidus, medial and lateral geniculate bodies) were not identified on the patient scans.

Open in a separate window

Fig. 3

Screenshot of a patient’s manual painting using ITKsnap. Axial (top left), sagittal (top right), and coronal (bottom right) slices of MR image and 3D (bottom left) view of the following bilateral painted structures: amygdala (brown), hippocampus (white), putamen (dark blue), thalami (light blue), substantia nigra (dark violet), STN (dark green), and red nucleus (light violet).

To measure the quality of the manual segmentation used as a gold standard, two experts (Claire Haegelen and Louis Collins) segmented the putamen, subthalamic nucleus and substantia nigra, on both the left and right sides, on the MRI of one patient. These segmentations were used to estimate the intra- and inter-rater variability in structure segmentation.

Registration with ANIMAL and SyN

As described earlier, the registration-based segmentation procedure maps the labels from the PD_T1_57 template onto the MR image using the nonlinear registration transformation that aligns the template and patient MR image data. The template’s labels, customized to the anatomy of each patient, can then be compared with the gold standard labels described above. Two methods of inter-subject nonlinear registration were evaluated: ANIMAL [15] and SyN [17]. So, as not to bias the comparison between the two methods, only the T1w modality was used to drive the non-linear registration. Furthermore, the PD_T1_57 template was more appropriate for the whole-brain registration since the PD_T2_42 only covered part of the brain.

Linear and nonlinear transformations were calculated between the PD_57_T1 template and each patient’s T1w image with both ANIMAL and SyN, resulting in two spatial warping transformations. Then, the template’s labels were mapped through each of the two transformations onto the patient’s T1w image to achieve two automatic structure segmentations: one for ANIMAL and one for SyN (see Figure 4).

Open in a separate window

Fig 4

Generation of labels by ANIMAL registration, SyN registration and patch-based method. All labels are shown on axial (top row), parasagittal (second row), and coronal (third row) slices of a T1w map and on a 3D reconstruction (bottom row). Labelled structures: bilateral amygdala (brown), hippocampus (white), putamen (dark blue), thalami (light blue), substantia nigra (dark violet), STN (dark green), and red nuclei (light violet).

Results

The results of intra- and inter-rater variability of manual segmentation are presented in Table 2. The results of the center-of-gravity values between the deep brain structures obtained with 2 registration methods and the patch-based method are presented in Table 3.

The paired-t test showed a significant difference between the SyN registration and patch-based method (p = 0.01) and the ANIMAL registration and patch-based method (p = 0.002). There was no difference between ANIMAL and SyN registration methods in terms of distances (p = 0.13). The mean distances measured between the repeated segmentations of the two experts were 0.82 (standard deviation (SD) 0.49) and 0.96 (SD 0.36) for respectively Claire Haegelen and Louis Collins. These mean distances were not significantly different (p = 0.15). The kappa index values obtained with the three methods are presented in Table 4 and Figure 5.

Open in a separate window

Fig 5

Box plots of the kappa index distribution for ANIMAL registration (top row), SyN registration (second row), and patch-based method of segmentation (bottom row). Boxes represent the lower quartile, median (red line), and upper quartile of the kappa index distribution. Whiskers show the most extreme values within 1.5 times the interquartile range. Outliers (red +) are data with values beyond the ends of the whiskers. l_AG: left amygdala; r_AG: right amygdala; l_HC: left hippocampus; r_HC: right hippocampus; l_PUT: left putamen; r_PUT: right putamen; l_THAL: left thalamus; r_THAL: right thalamus; l_RN: left red nucleus; r_RN: right red nucleus; l_SN: left substantia nigra; r_SN: right substantia nigra; l_STN: left subthalamic nucleus; r_STN: right subthalamic nucleus.

Patch-based method

The median kappa values computed between the labels obtained with the patch-based method and the manual labels were greater than 0.76 for the bilateral thalami, putamen, amygdala, hippocampus, and red nuclei, and between 0.62 and 0.70 for the bilateral substantia nigra and subthalamic nuclei (Fig. 5).

The Friedman test showed a significant difference between the three methods for the left amygdala (p = 0.014), left red nucleus (p = 0.045), and left (p = 0.045) and right thalamus (p = 0.001) (Fig. 6). For the left amygdala, the patch-based segmentation method was better than with ANIMAL. For the left red nucleus and left thalamus, registration with ANIMAL was better than with SyN. For the right thalamus, registration with ANIMAL was better than the patch-based method.

Open in a separate window

Fig 6

Mean kappa index distribution for segmentation of cerebral structures in 10 patients with Parkinson’s disease using ANIMAL registration, SyN registration, and patch-based method. * represents a significant difference between two methods according to Friedman and post hoc multi-comparison test (see the text in RESULTS for details).

Discussion

The template-based segmentation approach is one method that has been developed to assist neurosurgeons in the target planning step of deep brain stimulation procedures. For example, Bardinet et al. [6] and Chakravarty et al. [7] developed different 3D histological atlases that were aligned to a MR image of the same brain obtained after autopsy, and to an average MR image, respectively. The template can then be registered to the patient’s MR image to delineate deep brain structures like the STN or GPm that will be targeted during surgery. We built a high-resolution/high-signal-to-noise ratio template by averaging the nonlinearly aligned MR images of many patients with Parkinson’s disease. This process allowed us to obtain T1w and T2w templates with a high degree of anatomical detail for deep brain structures and the cerebral cortex. In contrast to the atlases of Bardinet et al. [6] and Chakravarty et al. [7], we have a direct high-quality visualization of the basal ganglia on the PD_T1_57 template and even more so on the PD_T2_42 template for those nuclei that are only visible on the T2w images, such as the red nucleus, substantia nigra, and subthalamic nucleus (see Figure 1). The segmentation of the basal ganglia and deep brain structures is an important step in obtaining their 3D representation and importing them onto the patient’s MR image. The manual segmentation of deep brain structures by the first author, Claire Haegelen, was found to be easier on the PD_T1_57 and PD_T2_42 templates than on a 3T digitized mono-subject template [8] obtained from a normal subject, but this fact was difficult to quantify and demonstrate, and beyond the scope of this paper where the goal was to compare three segmentation procedures. One of the limits of our study was that the segmentation used during methods comparison was performed by only one expert. However, the segmentation quality was verified by a second expert, Xavier Morandi and two experts (Claire Heagelen and D. Louis Collins) repeatedly segmented three cerebral structures of a patient with Parkinson’s disease to estimate intra- and inter-rater reliability (see Table 2). The intra-rater segmentation reliability was quite high with values above 0.83 for putamen, around 0.80 for substantia nigra, and between 0.71 and 0.74 for the subthalamic nucleus. The inter-rater segmentation variability was slighty higher than the intra-rater for putamen and SN while significantly higher for STN. This suggests that the manual delineation of STN from T1w MR scan is a very challenging task due to low contrast and small structure size that limits the possible consensus between expert’s segmentations. Moreover, the manual segmentation of putamen, SN and STN in patients with Parkinson’s disease is difficult because these patients have a specific pathological loss of neurons of these three structures affecting their MRI visualization. In the literature, very few studies have done manual segmentation of basal ganglia from MRI or on histological slides [6, 7] because of the great anatomical variability and the lack of very specific MRI sequences to identify accurately the STN, SN and putamen in patients with basal ganglia pathologocial changes like those seen in Parkinson’s disease. In the literature, many segmentation methods [18,29,34,36] are first tested on lateral ventricles, hippocampus or amygdala, structures for which the boundaries and limits are better defined. In our opinion, these facts explain why in our study the STN had higher inter-rater variability compare to other reports of automatic segmentation techniques. Considering these manual segmentation difficulties, our results should be taken with the caveat that the inter-rater variability is relatively high. Nevertheless, the mean distances between centers of gravity were not different in the two sets of manual segmentations, and these distances were lower in labels obtained with the patch-based method than with the two registration methods. Overall, the mean distance of 1.5mm for the automatic techniques is only higher than the 0.82mm of inter-rater variability. Furthermore, the mean Dice Kappa values for inter-rater segmentation are at approximately the same level as the mean Dice Kappa values comparing the manual labels and those generated by the three automatic techniques. In the future, we will investigate the effect of consensus segmentations from multiple experts.

To our knowledge, this is the first study to compare different registration approaches and the patch-based method to identify deep brain structures in the context of Parkinson’s disease. We applied two nonlinear registration-based segmentation tools to MR images, as nonlinear registration has been shown to perform better than linear registration for the automated identification of subcortical targets [30–32]. The accuracy of the automated segmentation depends on the accuracy of registration. In Klein et al. [33], the two nonlinear tools, ANIMAL and SyN, were assessed based on overlap measures of manually labelled anatomical regions, as was done in our study. SyN was one of the methods that reached top rank for all label sets [33]. In our case, the two nonlinear registrations were quite similar for all the structures, even for the smallest ones, except for the left thalamus and red nucleus, for which ANIMAL performed better than SyN. The advantage of this type of segmentation approach is that the manual segmentation is performed on a high-contrast image (i.e., the template), which facilitates the delineation procedure. However, the difficulty is then in registering the template with an individual MR image because of inter-subject variability. A linear (or affine) transformation consists in transforming a template into the image space of the patient’s MR data. Although it can account for variations in size, orientation, and position, a linear transformation cannot radically alter the shape of the template as is needed to account for anatomical differences between the patient’s image and the template. Instead, nonlinear transformations are needed to address this issue. By using nonlinear transformations, we obtained a more accurate segmentation result; however, the nonlinear registration of the template to the patient MR image is time consuming, taking in our case from thirty minutes to one hour per patient for ANIMAL and from one to three hours per patient for SyN.

Another way to obtain better segmentation accuracy than by registering images to a single labelled template is to fuse several labelled images [34]. Fusing the labelled images in the training library can be done by nonlinear registration [34–35] or patch comparison [18]. The advantage of label fusion approaches is that they can handle inter-subject variability correctly by aggregating labels from several images. In this manner, different anatomical patterns can be optimally fused to fit the patient’s anatomy. The patch-based label fusion approach evaluated here is also less time consuming than nonlinear registration, since it required only one minute of processing.

The difficulty with the patch-based label fusion approach is that it requires manual segmentation of structures on the MR images of several subjects to form the library. Not only is this task time consuming, but it is also difficult due to low image contrast, especially in the context of Parkinson’s disease where motion artefacts greatly impact the image quality. While the overall segmentation quality was similar for the patch-based and registration-based methods, there was more variability in the segmentation quality of the patch-based approach (see Fig. 5 and Table 4). As shown in Coupe et al. [18], the quality of the segmentation of patch-based label fusion depends on the number of training templates in the library. During our leave-one-out experiment, only nine templates were used to segment the subject under study. This number of training templates is small compared with what is suggested in Coupe et al. [18] (between 20 and 30). By increasing the number of training templates, the segmentation quality is expected to increase and the robustness of the segmentation to improve.

The kappa metric is often used to compare the agreement between structures in the literature [18,32,34–36]. The kappa metric depends on the surface-to-volume ratio of the structure, with decreasing kappa values in small structures for which the ratio increases. We found lower kappa values for the hippocampus than what is reported in Coupe et al. [18] (κ = 0.884) and Collins et al. [34] (κ = 0.887), and also for the amygdala, for which Collins et al. [34] found (κ = 0.826). As mentioned, this lower segmentation quality may be due to the smaller number of training subjects used in our segmentation procedure. It might also be due to the low-contrast images of patients with Parkinson’s disease or to the use of a different database [18,34]. Patenaude et al. [36] used a Bayesian appearance model approach to provide accurate segmentation of 15 subcortical structures on 336 manually labelled T1w MR images. They found that the best structures in terms of kappa values were the thalamus and putamen with a median kappa as high as 0.85, and the hippocampus with a median kappa between 0.80 and 0.85, as in our study, though we found lowest kappa values for the same structures. Patenaude et al. [36] found that the worst structures in terms of kappa values were the amygdala and the nucleus accumbens because, in those cases, the kappa penalized small structures or structures with a higher surface-to-volume ratio. We also found lower median kappa values for the smallest structures, the STN and substantia nigra, but not for the red nucleus. One hypothesis may be that, in patients with Parkinson’s disease, compared with the red nucleus, the STN and substantia nigra are more affected by the loss of dopaminergic neurons and therefore more difficult to identify with automated segmentation tools.

The three methods were relatively equivalent in terms of mean and median kappa values, suggesting they perform similarly in producing the automated segmentation of deep brain structures. Differences between the three methods were observed only for some of the structures, which mean there was, in fact, very little dissimilarity. However, when applied to a new patient, the patch-based method is notably more efficient than the other two methods, as it avoids the lengthy time-consuming non-linear registration step to achieve segmentation. Moreover, considering the resulting high inter-rater variability in our study, the automatic segmentation was as variable as another expert rater but was not subject to intra-rater variability. The STN deep brain stimulation proved a great efficacy in reducing symptoms of Parkinson’s disease [37,38]. The patch-based segmentation method may improve the surgical workflow by quickly and automatically identifying the STN on the preoperative MRI data of new patients, thus reducing part of the preoperative planning time in STN targeting. In clinical practice, the patch-based method is not yet applicable because it requires an engineer. From a point of view of a neurosurgeon, the three methods of segmentation were accurate enough to visualize the STN or the thalamus and to be applicable to new patients. The limits were the long time consuming step of the ANIMAL and SyN registrations method in clinical practice. More work is needed to have a software solution that is usable by the clinicians without supervision of the engineers.

Conclusions

We have developed the first step toward a multimodal database on deep brain stimulation, specifically, the construction of an MR image template specific to Parkinson’s disease, and also evaluated a method of achieving the accurate segmentation of the basal ganglia and deep brain structures on patient data.

We compared three methods to determine the best strategy for segmenting the basal ganglia and deep brain structures on patients’ MR images. We assessed that the intra-rater variability between the manual segmentations was lower than the inter-rater variability. Although template-based approaches are the most frequently used techniques to warp basal ganglia segmentation onto MR images, we found that the patch-based method provides similar results and is much less time consuming.

Acknowledgments

References