Fully automated 3D segmentation of dopamine transporter SPECT images using an estimation-based approach
Ziping Liu, Hae Sol Moon, Richard Laforest, Joel S. Perlmutter, Scott A. Norris, Abhinav K. Jha
FFully automated 3D segmentation of dopamine transporter SPECT images usingan estimation-based approach
Ziping Liu a , Hae Sol Moon a , Richard Laforest b , Joel S. Perlmutter b,c , Scott A. Norris b,c , Abhinav K. Jha a,b, ∗ a Department of Biomedical Engineering, Washington University in St. Louis, St. Louis, MO, USA b Department of Radiology, Washington University School of Medicine, St. Louis, MO, USA c Department of Neurology, Washington University School of Medicine, St. Louis, MO, USA
Abstract
Quantitative measures of uptake in the caudate, putamen, and globus pallidus in dopamine transporter (DaT) brainSPECT have potential as biomarkers for the severity of Parkinson disease. This potential is critical for clinicalcare and research. Reliable quantification of uptake requires accurate segmentation of these regions. However,segmentation is challenging in DaT SPECT due to partial-volume effects, system noise, physiological variability,and the small size of these regions. To address these challenges, we propose an estimation-based approach tosegmentation. This approach estimates the posterior mean of the fractional volume occupied by caudate, putamen,and globus pallidus within each voxel of a 3D SPECT image. The estimate is obtained by minimizing a cost functionbased on the binary cross-entropy loss between the true and estimated fractional volumes over a population ofSPECT images, where the distribution of the true fractional volumes is obtained from magnetic resonance (MR)images from clinical populations. The proposed method accounts for both the sources of partial-volume effectsin SPECT, namely the limited system resolution and tissue-fraction effects. The method was implemented usingan encoder-decoder network and evaluated using realistic clinically guided SPECT simulation studies, where theground-truth fractional volumes of caudate, putamen, and globus pallidus were known. The method significantlyoutperformed all other considered segmentation methods, including a U-net-based method, and yielded accuratesegmentation with dice similarity coefficients of ∼ .
80 for all regions. The method was relatively insensitive to thechanges in voxel size. Further, the method was relatively robust up to ± ◦ of patient head tilt along transaxial,sagittal, and coronal planes. Overall, the results demonstrate the efficacy of the proposed method to yield accuratefully automated segmentation of caudate, putamen, and globus pallidus in 3D DaT-SPECT images. Keywords:
Parkinsonian syndromes, single-photon emission computed tomography, estimation, segmentation, ∗ Corresponding author
Email address: [email protected] (Abhinav K. Jha) a r X i v : . [ phy s i c s . m e d - ph ] J a n artial volume effects, tissue fraction effects
1. Introduction
Parkinson disease (PD) is the second-most common neurodegenerative disease. The disease is relentlessly pro-gressive and will affect 12 million people worldwide by 2040 (Fahn and Sulzer, 2004; Dorsey et al., 2018; Dorseyand Bloem, 2018). Dopamine transporter (DaT) single-photon emission computed tomography (SPECT) measurespre-synaptic dopaminergic neurons that are known to degenerate in patients with PD. Of the various DaT-basedligands (Emamzadeh and Surguchov, 2018), DaTscan (Ioflupane I-123 injection) is FDA approved to assist withthe evaluation of adult patients with suspected parkinsonian syndromes (I13, 2011). For this clinical application,DaT-SPECT images typically are analyzed visually by trained readers. However, this is error-prone and suffersfrom intra- and inter-reader variability (Augimeri et al., 2016). To address this issue, several studies are investi-gating whether quantitative uptake in striatal regions, i.e. the caudate and putamen, provide more useful clinicalinformation (Joling et al., 2017). Further, shape (Staff et al., 2009; Prashanth et al., 2016) and texture (Martinez-Murcia et al., 2014; Augimeri et al., 2016) analyses of striatal regions are also being explored to enhance the clinicalapplications of DaT SPECT. Conducting these quantification, shape- and texture-analysis studies requires reliablesegmentation of caudate and putamen from DaT-SPECT images.Further, the greatest unmet therapeutic need in PD is for disease-modifying therapies (Lang and Espay, 2018).While no such therapies have been identified, several are being investigated (Lang and Espay, 2018; AlDakheel et al.,2014; Elkouzi et al., 2019). Biomarkers that accurately reflect disease severity are critical for these studies and theneed is pressing (McGhee et al., 2013). Most studies on developing such biomarkers have focused on the striataluptake, but such measures may correlate with severity only early in the disease (Perlmutter and Norris, 2014; Saariet al., 2017; Honkanen et al., 2019; Karimi et al., 2013). Thus, there remains an important need for improvedbiomarkers that can measure severity throughout the range of the disease. Measurements in globus pallidus (GP)may play a key role. Several post-mortem studies demonstrated altered DaT uptake in GP (Ciliax et al., 1999;Porritt et al., 2005). Clinical studies have observed correlations between DaT reduction in GP and resting tremorseverity (Helmich et al., 2011). In addition, small-sized studies suggest that degeneration of nigrostriatal dopamineneurons in early PD is compensated in part by increased DaT uptake in GP that is lost in more advanced stagesof disease (Whone et al., 2003). The observations have led to the hypothesis that the plasticity of the nigropallidalpathway helps maintain a more normal pattern of pallidal output to ventral thalamus and motor cortex in earlyPD. Further, the loss of this pathway in advanced disease may be a pivotal step in disease progression. Similarfindings were reported in parkinsonism secondary to occupational manganese toxicity (Emamzadeh and Surguchov,2018). Thus, investigating DaT uptake in GP may help provide a measure of PD severity. DaT-SPECT studiesprovide a mechanism to study these effects in vivo . However, this requires the availability of tools to segment theGP in SPECT images, and such tools are currently not available. Thus, there is an important need to develop suchtools.Segmentation in SPECT is challenging due to limited system resolution, system noise, and physiological vari-ability in patient populations (Badiavas et al., 2011; Dewaraja et al., 2001). In addition, the caudate, putamen,and GP are in proximity to each other. The limited system resolution in SPECT blurs the boundaries of theseregions, thus making the segmentation task challenging. In fact, GP is visually almost impossible to demarcatefrom SPECT images, and to the best of our knowledge, there are no validated automated or semi-automated toolsto segment the GP. Another challenge is the small size of these regions. GP is ∼ ∼
10 cc. Segmentation is typically performed manually by trained readers, but thatis expensive, tedious, time-consuming, and suffers from intra- and inter-reader variability, especially due to thepoor resolution in SPECT (Jha et al., 2017a). To address this issue, several semi-automated SPECT segmentationmethods have been proposed, such as those based on thresholding (Long et al., 1992; Tsujimoto et al., 2018), edgedetection (Long et al., 1992), region growing (Slomka et al., 1995), and clustering (Mignotte et al., 2001). However,our studies showed that these methods typically segmented the caudate, putamen, and GP as highly overlappedregions in DaT-SPECT images. Further, a recent study showed that semi-automated segmentation methods yieldedlarge variabilities in DaT-SPECT-derived metrics, and thus may have limited clinical utility (Matesan et al., 2018).Thus, there remains an important need for improved methods to segment DaT-SPECT images, and in particular,methods that can segment the GP.To address this need, we recognize that a major barrier to segmenting the caudate, putamen, and GP is thelack of clear boundaries between these regions in SPECT images. This barrier arises due to partial-volume effects(PVEs) in the reconstructed images. PVEs have two sources, namely the limited system resolution and tissue-fraction effects (TFEs). The resolution of clinical brain SPECT systems is typically ∼ ∼
12 mm (Stam et al., 2018; Council et al., 1996). Thus, structures such as caudate,putamen, and GP that are close to each other, are heavily affected by this limited resolution. Further, due to thereconstruction of image over finite-sized voxel grids, voxels on the boundary of two regions likely contain both theregions. While this is generally an issue in medical imaging, the effect is more prominent in SPECT due to thelarger voxel sizes.Conventional segmentation methods are typically designed to perform segmentation through voxel-wise classifi-cation, i.e. each image voxel is assigned as exclusively belonging to one region. Thus, these methods are inherently3imited in accounting for TFEs. Probabilistic techniques such as fuzzy segmentation methods (Chen et al., 2019)and probabilistic atlas-based segmentation methods (Lee and Lee, 2005), can provide a probabilistic estimate ofeach image voxel belonging to a certain region. However, this probabilistic estimate is only a measure of uncer-tainty in classification and thus has no relation to the TFEs. Conventional deep-learning (DL)-based segmentationmethods (Leung et al., 2020; Lin et al., 2020) are designed and trained on the task of classifying each image voxelas belonging to one region. While these methods can also output probabilistic estimates for each image voxel, theestimates again only infer classification uncertainty and do not relate to the TFEs.To account for both the sources of PVEs, Liu et al. (2021) proposed an estimation-based segmentation technique,in the context of segmenting tumors in oncological positron emission tomography (PET) images. This techniqueis designed to estimate the fractional volume that a tumor occupies within each voxel of a PET image. Throughthis strategy, TFEs can be explicitly modeled. In this manuscript, we propose to advance this estimation-basedtechnique in the context of segmenting the caudate, putamen, and GP in 3-D DaT-SPECT images. Our proposedmethod is a Bayesian approach, and as shown later, requires a distribution of the fractional volumes that caudate,putamen, and GP occupy within each image voxel. This distribution is obtained from clinical T1-weighted magneticresonance (MR) images, where these regions are visible separately and can be delineated. The method then usesthis distribution to estimate the posterior mean of the fractional volumes of caudate, putamen, and GP in DaT-SPECT images. As our results show, the method yielded accurate segmentation of these regions and significantlyoutperformed all other considered segmentation methods.
2. Methods
Consider a patient injected with a DaT-based tracer such as 123-Ioflupane. Let f ( r ) describe the DaT-tracerdistribution within the brain, where r = ( x, y, z ) is a 3-D vector denoting spatial coordinates. Consider a SPECTsystem that images this patient and yields sinogram data. This data is then input to a reconstruction algorithm,yielding a reconstructed image ˆ f consisting of M voxels. Denote the process of obtaining the reconstructed imagefrom the original tracer distribution by the operator Θ : L ( R ) → E M . As outlined in Sec. 1, our objective isto design an approach that estimates the fractional volume occupied by the k th region of left and right caudate,putamen, and GP in each voxel of the reconstructed image ˆ f . Using this approach, we are able to explicitly modelTFEs while performing segmentation. 4et the support of the k th region be given by φ k ( r ), i.e. φ k ( r ) = , if region k occupies location r .0 , otherwise. (1)Next, define the voxel function ψ m ( r ) as ψ m ( r ) = , if r lies within the m th voxel in the DaT-SPECT image.0 , otherwise. (2)Let V denote the total volume of each voxel. The fractional volume occupied by the k th region in the m th voxel inthe SPECT image, denoted by v ideal ,k,m , is given by v ideal ,k,m = 1 V (cid:90) d r φ k ( r ) ψ m ( r ) . (3)Estimating the fractional volumes within each voxel thus requires estimating the values of φ k ( r ) from ˆ f . This isan ill-posed problem due to the null functions of the Θ operator. One way to alleviate this ill-posedness is toincorporate prior distribution of φ k ( r ). To obtain this distribution, we benefit from the availability of existing MR-image populations, which can provide the distribution of φ k ( r ) at relatively higher resolution. More specifically, theboundaries of caudate, putamen, and GP are distinguishable on MR images due to the higher resolution of MRI.Thus, the acquired MR images can be segmented to obtain a distribution of the support of these regions.Denote the MR image by an N -dimensional vector f MR , where N > M . The voxel function for this image isdenoted as ψ MRn ( r ). Consider that this image has been segmented into K distinct regions. For each region, wedefine an N -dimensional vector φ MRk , which denotes the support of that region. The elements of this vector aredefined as follows: φ MRk,i = , if voxel i in the MR image is assigned to region k .0 , otherwise. (4)Consider a patient population where the MR and SPECT images are co-registered. Thus, from the MR imagesin this population, we can obtain a discrete version of the support of the k th region as defined in Eq. (1), at ahigher resolution than the SPECT images. The ground-truth fractional volume of k th region in the m th voxel ofthe SPECT image can thus be calculated as follows: v k,m = 1 V N (cid:88) n =1 φ MRk,n (cid:90) d r ψ MRn ( r ) ψ m ( r ) , (5)5here the integral computes the volume that the n th voxel in the MR image occupies within the m th voxel ofthe SPECT image. We frame our problem as the task of estimating these true fractional volumes { v k,m , k =1 , , . . . K, m = 1 , , . . . M } from the SPECT images. Let the estimate of v k,m be defined as ˆ v k,m . Denote the vector { v k,m , m = 1 , , . . . , M } by v k , and the estimateof v k by ˆv k . To compute ˆv k from the reconstructed SPECT image ˆ f , we first need to define a cost function. In thisproblem, values of v k,m and ˆ v k,m are constrained to lie between 0 and 1. Using binary cross-entropy (BCE) lossfunction allows us to automatically incorporate this constraint on the range of these values (Creswell et al., 2017).Thus, the BCE loss, denoted by l BCE ( v k,m , ˆ v k,m ) and given by the equation below, is chosen as the basis of our costfunction: l BCE ( v k,m , ˆ v k,m ) = − v k,m log(ˆ v k,m ) − (1 − v k,m ) log(1 − ˆ v k,m ) . (6)The cost function, denoted by C ( v k , ˆv k ), is defined to minimize the negative of aggregate BCE loss between v k,m and ˆ v k,m over all M voxels, averaged over the ensemble of true values v k,m and the noise ˆf in the imaging process.The cost function is then given by: C ( v k , ˆv k ) = − M (cid:88) m =1 (cid:90) d M ˆ f (cid:90) dv k,m pr(ˆ f , v k,m ) l BCE ( v k,m , ˆ v k,m ) . (7)Using conditional probability to expand pr(ˆ f , v k,m ) and replacing the expression from Eq. (6) in Eq. (7), we obtain C ( v k , ˆv k ) = M (cid:88) m =1 (cid:90) d M ˆ f pr(ˆ f ) (cid:90) dv k,m pr( v k,m | ˆ f ) (cid:104) v k,m log(ˆ v k,m ) + (1 − v k,m ) log(1 − ˆ v k,m ) (cid:105) = M (cid:88) m =1 (cid:90) d M ˆ f pr(ˆ f ) (cid:90) dv k,m pr( v k,m | ˆ f ) (cid:104) log(1 − ˆ v k,m ) + v k,m { log(ˆ v k,m ) − log(1 − ˆ v k,m ) } (cid:105) . (8)Since (cid:82) dv k,m pr( v k,m | ˆ f ) = 1, Eq. (8) becomes C ( v k , ˆv k ) = M (cid:88) m =1 (cid:90) d M ˆ f pr(ˆ f ) (cid:20) log(1 − ˆ v k,m ) + { log(ˆ v k,m ) − log(1 − ˆ v k,m ) } (cid:90) dv k,m pr( v k,m | ˆ f ) v k,m (cid:21) . (9)Our objective is to find the value of ˆ v k,m that minimizes this cost function. This is the point at which the derivativeof the cost function with respect to ˆ v k,m is 0. Since the term pr(ˆ f ) is always non-negative, the cost function isminimized by setting ∂∂ ˆ v k,m (cid:20) log(1 − ˆ v k,m ) + { log(ˆ v k,m ) − log(1 − ˆ v k,m ) } (cid:90) dv k,m pr( v k,m | ˆ f ) v k,m (cid:21) = 0 . (10)6he solution is given by ˆ v ∗ k,m = (cid:90) dv k,m pr( v k,m | ˆ f ) v k,m . (11)This is simply the posterior mean estimate of v k,m . Thus, by developing an optimization procedure that minimizesthe cost function defined in Eq. (9), we obtain an estimator that yields the posterior mean estimate of the ground-truth fractional volume occupied by the left and right caudate, putamen, and GP in each voxel of the SPECTimage. Note that the same estimator is obtained when the ensemble mean squared error is defined as the costfunction. Thus, this estimator also yields the lowest mean squared error among all estimators. We next describeour optimization procedure. To minimize the cost function given by Eq. (9) requires sampling from the posterior distribution pr( v k,m | ˆf ). Thisis challenging since this distribution is high dimensional and an analytical form for the distribution is unavailable.To address this challenge, the proposed method was implemented by constructing an encoder-decoder network(Fig. 1). In the training phase, the network was input a population of 3-D SPECT images and the correspondingground-truth fractional volumes occupied by caudate, putamen, and GP, i.e. v k for each image, as described inSec. 2.1.1. The network was trained by minimizing the cost function defined in Eq. (9) to yield a posterior meanestimate of the fractional volumes occupied by these regions within each image voxel.The network design for the proposed method is similar to other networks that are designed for estimation-basedtasks, such as those for image denoising (Creswell et al., 2017) and reconstruction (Nath et al., 2020). Briefly, thenetwork consists of an encoder and decoder, each containing cascades of convolutional layers. The encoder extractslocal spatial features from the input SPECT image and the decoder maps the extracted features to the fractionalvolumes. In the final layer, the network outputs the fractional volumes occupied by left and right caudate, putamen,and GP in each voxel of the input SPECT image. To improve training performance, several features were adopted inthe network design. Following each convolutional layer, an activation function of leaky rectified linear unit (ReLU)was applied to accelerate the convergence speed (Xu et al., 2015). Dropout was used as a regularization strategy toprevent overfitting (Srivastava et al., 2014). In addition, skip connections with element-wise addition were appliedbetween the output of layers in the encoder and decoder to stabilize network training (Mao et al., 2016). Further,instead of conducting slice-by-slice segmentation or patch-based training (Kamnitsas et al., 2017), the network wasdesigned in fully 3-D to provide the maximal amount of global contextual information per patient image. The detailof the network architecture is given in the appendix (Table A1).7 nput Output caudateputamenglobuspallidus
Fractional volume occupied by globus pallidus within each image voxel
Figure 1: Illustration of the implementation of the proposed method. The network, when input a 3-D DaT-SPECT image, outputs theestimated fractional volumes occupied by left and right caudate, putamen and GP within each voxel of the image.
The network architecture and training were implemented in Python 3.6.9 and Keras 2.2.4. For effective trainingperformance on 3-D images, experiments were performed on a Red Hat Enterprise Linux 7.7 operating system withtwo 24 GB NVIDIA Titan RTX graphics processing unit cards.
3. Evaluation
Evaluation of the proposed method using clinical data requires access to DaT-SPECT images, where the ground-truth fractional volumes of caudate, putamen, and GP are known. One approach to conducting such a study is tohave expert readers delineate these regions. However, manual delineations are not guaranteed to be accurate dueto PVEs in SPECT. In fact, GP is almost impossible to delineate manually on the DaT-SPECT images. Further,the boundaries between caudate and putamen are unclear. Even more importantly, accounting for TFEs whileperforming manual delineations is challenging. Finally, manual delineations are well known to suffer from intra-and inter-reader variability (Augimeri et al., 2016). All these factors make manual delineations a limited measureof ground truth. This issue of lack of ground truth when evaluating methods for quantitative imaging is oftenaddressed by conducting realistic simulation studies (Du et al., 2005; Ouyang et al., 2006; Du and Frey, 2009; Jinet al., 2013). Realistic simulations provide a mechanism to evaluate imaging methods with known ground truth,model imaging physics, and account for patient-population variability. Thus, we developed a clinically guided and8ealistic simulation-based approach to evaluate the proposed method. In this approach, clinical MR images wereused to generate clinically realistic SPECT images by accurately modeling the tracer distributions and SPECTimaging physics. Further, this approach allowed us to quantitatively evaluate the effect of factors such as patienthead tilt on the performance of the proposed method.Note that while this simulation-based approach requires simulation of SPECT system, the ground-truth distri-bution of the fractional volumes of caudate, putamen, and GP is not simulated and is obtained directly from clinicalstudies. Below, we describe the process to rigorously evaluate the proposed method using this simulation-basedapproach.
A total of 580 T1-weighted MR images of cognitively healthy patients from the Open Access Series of ImagingStudies-3 (LaMontagne et al., 2019) were used. Caudate, putamen, and GP for both left and right hemisphereswere segmented from these images using the Freesurfer software (Fischl, 2012). The segmentations were performedin Montreal Neurological Institute (MNI) space and then transformed back into each patient’s native space, yielding580 anatomical brain templates. These templates were of 256 × ×
256 voxels, with the voxel size of 1 mm × × Table 1: Mean and standard deviation of the specific binding ratio of caudate, putamen, and globus pallidus based on Son et al. (2016).The occipital region was chosen as the reference region.
These ground-truth tracer distributions were then used to generate realistic DaTscan SPECT images. A SPECTsystem with acquisition parameters similar to the GE Discovery 670 scanner (GE Healthcare, Haifa, Israel) was9imulated. The acquisition modeled a DaTscan SPECT protocol with 123-Ioflupane tracer and a low-energy high-resolution (LEHR) collimator. Projection data were generated using the SIMIND Monte Carlo simulation software(Ljungberg et al., 2012). This simulation modeled major image-degrading processes in SPECT, including photonattenuation and scatter, the finite extent of the collimator with thickness of 3 . .
8% and the intrinsic spatial resolution of 0 .
39 cm at 159 keV of the detector. Following the Society of NuclearMedicine Practice guideline for DaTscan SPECT (Djang et al., 2012), 120 projection views for full 360 ◦ coverageof the head were acquired. The number of total detected counts were scaled to clinically realistic value of 2 millioncounts, to which Poisson noise was added. The noisy projection data were then reconstructed using an iterative 3-Dordered subset expectation maximization (OSEM)-based algorithm with 4 iterations and 8 subsets that compensatedfor attenuation, scatter and detector response. The overall procedure to generate the SPECT images is summarizedin Fig. 2. caudateputamenglobus pallidus Extract structural boundaries from T1-weighted MR images using FreesurferDefine clinically realistic dopamine transporter uptake ratio in caudate, putamen and globus pallidus
Ground-truth dopamine transporter uptake map
Highly realistic 3-D SPECT simulation using SIMIND 3-D OSEM reconstruction that compensates for attenuation, scatter and detector response
Forward projection Reconstructed SPECT images
Figure 2: Strategy to generate highly realistic simulated SPECT images. A digital phantom population of 580 patients were generatedwith anatomical templates obtained from existing MR images and tracer distributions guided by clinical data. The digital phantomswere input to SIMIND Monte Carlo simulation software to obtain projection data. The projection data were then fed into a 3-DOSEM-based reconstruction algorithm to generate reconstructed SPECT images. .2. Training strategy Following the simulation procedure described above, a total of 580 reconstructed 3-D DaTscan SPECT imageswere generated. Of the 580 SPECT images, 480 images were used for network training and validation using 5-foldcross validation. The remaining 100 completely independent images were reserved for evaluating the performanceof the proposed method.Our training strategy was designed specifically for the goal of explicitly modeling TFEs, as described in Sec. 2.1.1.Thus, the ground truth was defined as the fractional volumes occupied by caudate, putamen, and GP within eachvoxel of a SPECT image. The inputs to the network were the SPECT image and the corresponding true fractionalvolumes within each voxel of this image. The true fractional volumes were obtained using the Freesurfer-definedsegmentations on corresponding MR images. These segmentations provided the high-resolution support of eachregion, i.e. φ MRk as defined in Sec. 2.1.1. From this support, using Eq. (5), we obtained the ground-truth fractionalvolumes. The network was then trained to yield a posterior mean estimate of the true fractional volumes byminimizing the cost function C ( v k , ˆv k ) given by Eq. (9). The proposed method was quantitatively evaluated on the 100 test images, using metrics that evaluated spatialoverlap and shape similarity. In the category of spatial-overlap metrics, we used the dice similarity coefficient (DSC)and jaccard similarity coefficient (JSC). For each voxel in a SPECT image, the method yields continuous-valuedfractional volumes. Thus, the DSC and JSC were defined as in Taha and Hanbury (2015) to evaluate segmentationmethods that yield such non-binary outputs. Higher values of DSC and JSC infer more accurate segmentationperformance. In addition, the shape similarity between the true and estimated segmentations was quantified usingthe hausdorff distance (HD) (Huttenlocher et al., 1993). For this purpose, a topographic map was constructed foreach of the left and right caudate, putamen, and GP, using the corresponding estimated fractional volumes. Theground-truth topographic map of each region was also constructed using the true fractional volumes. From thetopographic map of each region, the set of points with the value of fractional volume equal to 0 . t -test with p -value < .
01 was performed to assess statistical significance.11 .3.2. Evaluating accuracy and comparing to other segmentation methods
The proposed method was quantitatively compared to several commonly used segmentation methods, on thebasis of DSC, JSC, and HD. Common methods for segmenting nuclear-medicine images can be classified as thosebased on thresholding, boundary detection, stochastic modeling, and learning (Foster et al., 2014). In our study,40% SUV-max thresholding (King et al., 1991), Snakes (Kass et al., 1988), Markov random fields-Gaussian mixturemodel (MRF-GMM) and a U-net-based method (Leung et al., 2020) were chosen in each of these categories. Notethat the U-net-based method, similar to conventional DL-based methods, was designed and trained on the taskof classifying each voxel as exclusively belonging to one region. The 40% SUV-max, Snakes and MRF-GMMmethods required manual inputs in the form of a seed voxel and/or an initial boundary estimate. The proposedand U-net-based method instead did not require any manual input and were fully automated.
To conduct this evaluation, we generated two sets of 580 simulated SPECT images, with the voxel size of2 mm × × × × The caudate, putamen, and GP are small-sized structures. Thus, slight differences in patient head tilt can causeartifacts in the acquired SPECT images and provide a false impression of the activity loss in putamen (Covingtonet al., 2013). We thus assessed the accuracy of the proposed method for different extents of head tilt in the 100 testpatients. For this purpose, the T1-weighted MR images of these patients were first registered to the MNI space.From the registered MR images, the ground-truth activity and attenuation maps were obtained. Patient head tiltwas then simulated by rotating the activity and attenuation maps in steps of 2 . ◦ for up to a maximal rotation of ± ◦ , based on experimental findings in Patterson et al. (1998). The rotation was performed along the transaxial,sagittal and coronal plane, separately, to study the sensitivity of our method to head tilt along each plane. Thesimulated SPECT images were generated following the procedure in Sec. 3.1. For this evaluation, reconstructedimages with 4 mm voxel size were considered. The performance of the proposed method was quantified using DSC.12 . Results Quantitatively, for all the left and right caudate, putamen, and GP, the proposed method significantly outper-formed (p < .
01) the 40% SUV-max, Snakes, and MRF-GMM, on the basis of DSC, JSC, and HD (Fig. 3, TableA2 and A3 in the appendix). Further, the proposed method yielded DSC of ∼ .
80 for all regions, indicating anaccurate segmentation performance (Zijdenbos et al., 1994).
LC RC LP RP LGP RGP00.20.40.60.81
LC RC LP RP LGP RGP00.20.40.60.8
LC RC LP RP LGP RGP024681012
LC RC LP RP LGP RGP00.20.40.60.81
LC RC LP RP LGP RGP00.20.40.60.8
LC RC LP RP LGP RGP02468 D S C J S C HDD S C J S C HD (a) Voxel size = 2 mm(b) Voxel size = 4 mm
LC RC LP RP LGP RGP00.20.40.60.81 proposed U-net-based 40% SUV-max Snakes MRF-GMM
Figure 3: Quantitative comparison of the segmentation performance between the proposed method and other considered methods, onthe basis of dice similarity coefficient, jaccard similarity coefficient and hausdorff distance, for SPECT images with voxel size of (a) 2mm and (b) 4 mm, respectively. (LC/RC: left/right caudate; LP/RP: left/right putamen; LGP/RGP: left/right globus pallidus)
Qualitatively, Fig. 4 shows the comparison between the ground-truth boundaries of caudate, putamen, and GPand the boundaries estimated using the proposed method. The boundaries were obtained following the proceduredescribed in Sec. 3.3.1. We observe that the proposed method yielded reliable boundaries of all regions and provideda close match to the ground-truth boundaries. 13dditionally, we observe in Fig. 5 that the 40% SUV-max, Snakes, and MRF-GMM all segmented the caudate,putamen, and GP as highly overlapped regions. In contrast, the proposed method was able to segment these regionsseparately and closely matched the ground truth. (a)(b)(c)(d) Representative slice Zoomed Zoomed adjacent slices
Figure 4: Qualitative evaluation of the proposed method for four representative simulated patients. The true boundaries of caudate,putamen and globus pallidus and the boundaries estimated using the proposed method were compared. The shown reconstructedSPECT images were with voxel size of 2 mm. audatePutamenGlobus pallidus Ground truth Proposed40% SUV-max Snakes MRF-GMM
Figure 5: Qualitative comparison between the true and estimated boundaries of left and right caudate, putamen and globus pallidususing conventional semi-automated segmentation methods. Boundaries yielded by the proposed method are also shown for comparison.The shown reconstructed SPECT images were with voxel size of 2 mm.
Fig. 3 shows that the proposed method yielded accurate segmentation performance for both 2 mm and 4 mmvoxel sizes, and significantly outperformed the U-net-based method on the basis of DSC and JSC. Note that thedifference in segmentation performance between the proposed and U-net-based method decreased from the voxelsize of 4 mm to 2 mm. This was because TFEs were less dominant in the reconstructed images with smaller voxelsize.For the proposed method, DSC and JSC remained relatively unchanged on varying the voxel size from 2 mmto 4 mm, indicating that the method was relatively insensitive to changes in voxel size. This insensitivity was dueto the capability of the method to explicitly model the TFEs. In contrast, the U-net-based method yielded limited15erformance for the 4 mm voxel size, demonstrating that not modeling TFEs has a tangibly adverse impact onsegmentation performance.
Fig. 6 shows the performance of the proposed method for different degrees of tilt along transaxial, sagittal, andcoronal planes, respectively. The method consistently yielded reliable segmentation performance between ± ◦ ofrotation along all three planes and yielded DSC higher than 0 .
70 for all regions. More specifically, the methodyielded high DSC of ∼ -10° -5° 0° +5° +10°0.650.70.750.80.85 -10° -5° 0° +5° +10°0.650.70.750.80.85 -10° -5° 0° +5° +10°0.650.70.750.80.85 -10° -5° 0° +5° +10°0.650.70.750.80.85 -10° -5° 0° +5° +10°0.650.70.750.80.85 -10° -5° 0° +5° +10°0.650.70.750.80.85 Transaxial plane Sagittal plane Coronal plane (a)(b) D S CD S C D S CD S C D S CD S C Degrees of tiltDegrees of tilt Degrees of tiltDegrees of tilt Degrees of tiltDegrees of tilt -10° -5° 0° +5° +10°0.650.70.750.80.85 left/right caudateleft/right putamenleft/right GP -10° -5° 0° +5° +10°0.650.70.750.80.85 left/right caudateleft/right putamenleft/right GP -10° -5° 0° +5° +10°0.650.70.750.80.85 left/right caudateleft/right putamenleft/right GP left/right caudate left/right putamen left/right GP
Figure 6: Sensitivity of the proposed method to patient head tilt along transaxial, sagittal and coronal planes. The performance ofsegmenting (a) left and (b) right caudate, putamen, and globus pallidus was quantified on the basis of dice similarity coefficient. . Discussion We proposed an estimation-based approach to fully automated 3-D DaT-SPECT segmentation. The proposedmethod accounts for PVEs, and in particular, TFEs, while performing segmentation. Our results demonstrate theefficacy of the method to accurately segment caudate, putamen, and GP in DaT-SPECT images. In particular,ability to segment the small-sized GP is distinctive as this region is visually almost impossible to demarcate andcurrently, to the best of our knowledge, no validated tools are available to segment this region in DaT-SPECTimages. This is clinically important as delineating GP presents an opportunity to rigorously evaluate pallidaluptake as a biomarker for measuring the severity of PD. This also opens up a new and important research frontieron analyzing the functional characteristics of the GP in patients with PD. In addition, the method accuratelysegmented the caudate and putamen in DaT-SPECT images. This provides a tool to evaluate quantitative features,including the shape and texture analysis of these regions, to determine whether differences exist in different formsof parkinsonism. These studies may lead to new biomarkers for measuring the severity of PD and possibly fordifferential diagnosis.We observe from Fig. 3 that the proposed method yielded accurate segmentation for all regions and significantlyoutperformed the other considered segmentation methods, including a U-net-based method (Leung et al., 2020).While the U-net-based and proposed method both learn anatomical variability and PVEs due to the limited systemresolution, the proposed method additionally accounts for PVEs due to TFEs that arise from the larger voxel sizein reconstructed SPECT images. Thus, the proposed method yielded significantly improved performance comparedto segmentation methods that do not account for increased TFEs. This explains the larger difference in the valuesof DSC and JSC between the proposed and U-net-based method when the reconstructed images have the larger4 mm voxel size. The higher accuracy of the proposed method for 4 mm voxel size is especially important as manyclinical DaT-SPECT scans are conducted with 4 mm voxel size (Djang et al., 2012). Additionally, we observe inFig. 3 that a change in the voxel size from 2 mm to 4 mm did not impact the performance of the proposed method.However, the DSC of the U-net-based method decreased by more than 10%. This provides evidence that theproposed method is relatively robust to changes in voxel size and provides distinct and quantitatively measurableadvantages to existing methods.Further, we observe in Fig. 6 that the proposed method consistently yielded reliable segmentation performancefor patient head tilt ranging from − ◦ to +10 ◦ along transaxial, sagittal, and coronal planes, respectively. Thisresult has great importance since slight differences in head tilt can create artifacts in DaTscan SPECT imagesthat can lead to the false impression of decreased or absent DaT activity in putamen and cause a misread clinicalscan (Covington et al., 2013). However, our method accurately segmented the putamen with high DSC of ∼ ± ◦ of head tilt along all anatomical planes, demonstrating the relative robustness of the method torelevant degrees of head tilt.Application of the proposed method in clinical and research settings requires inputs of only patient SPECTimages, and no images from any other modalities such as MR, are needed. This is an important advantage ofour method. Another class of methods to segment striatal regions in DaT-SPECT images uses MR images of thesame patient acquired from a different scanner (Rahmim et al., 2016). These methods have not been evaluated forsegmenting GP. Further, these methods require SPECT and MR images of the patient. However, both these imagingmodalities may not be available for the patient. While there has been recent progress in designing simultaneousSPECT/MR systems (Hutton et al., 2018), no clinical simultaneous SPECT/MR systems are currently available.The proposed method, by virtue of having no such requirement, can segment the DaT-SPECT image of a patienteven when the corresponding MR image is not available.There are some limitations in our study. First, the proposed method was evaluated using simulation studies.The simulations used clinical MR images as anatomical templates so that the distribution of the true fractionalvolumes was obtained directly from clinical imaging data. Further, the simulations accurately modeled the tracerdistributions and SPECT imaging physics. However, the simulations may not have modeled all aspects of patientphysiology and system instrumentation. Evaluation using physical-phantom and patient studies can help addressthese limitations by modeling all aspects of system instrumentation and modeling patient physiology accurately,respectively. Validation with patient studies requires a gold standard and that is not available for these types ofimages. To address this challenge, no-gold-standard (NGS) evaluation techniques have been developed (Kupinskiet al., 2002; Jha et al., 2012, 2016, 2017b; Liu et al., 2020). These techniques would provide an approach to evaluatethe proposed method with patient studies.Another limitation is that we focused on segmenting the caudate, putamen, and GP in DaT-SPECT images andconsidered the rest of brain as background. However, post-mortem and PET imaging studies identify DaT uptake inother brain regions such as nucleus accumbens and substantia nigra (Sun et al., 2012; Brown et al., 2013). Further,the GP can itself be separated into two parts, namely the internal and external GP. It remains to be determinedhow these two parts may contribute independently to PD progression. Results from this study motivate extendingthe proposed method to segment these regions. Finally, in our simulation studies, we assume that the uptake withinthe various regions is uniform. However, the tracer uptake within the putamen may be heterogeneous in patientswith PD. Modeling this heterogeneity and evaluating the performance of the proposed method in the presence ofthis heterogeneity are important research areas. 18 . Conclusion In this manuscript, we proposed an estimation-based fully automated method to segment the caudate, putamen,and globus pallidus in 3-D DaT-SPECT images. The method accounts for both the sources of partial-volume effectsin SPECT, namely system-generated blur and tissue-fraction effects. Essentially, the method estimates the posteriormean of the fractional volume that the different regions occupy within each voxel of a SPECT image, where the priordistribution of the fractional volumes were obtained from existing MR-image populations. Evaluation using realisticclinically guided simulation studies shows that the method accurately segmented the caudate, putamen, and globuspallidus in 3D DaT-SPECT images. The method significantly outperformed all other considered segmentationmethods, including a U-net-based method, by yielding high DSC of ∼ .
80 for all regions. The method wasrelatively insensitive to changes in voxel size. Further, the method was relatively robust to patient head tilt up to ± ◦ along transaxial, sagittal and coronal planes. Overall, the results demonstrate the efficacy of our method in3-D DaT-SPECT segmentation and motivate further evaluation with physical-phantom and patient studies. Acknowledgements
This work was supported by the National Institute of Biomedical Imaging and Bioengineering Trailblazer R21Award (R21-EB024647) and the NVIDIA GPU grant, the American Parkinson Disease Association (APDA), theGreater St. Louis Chapter of the APDA, the Barnes-Jewish Hospital Foundation (Elliot Stein Family Fund), andthe Paula & Rodger Riney Fund. The authors also thank the Washington University Center for High PerformanceComputing for providing computational resources. The center is partially funded by NIH grants 1S10RR022984-01A1 and 1S10OD018091-01. 19 ppendix
Detailed architecture of the encoder-decoder network designed for the proposed method is given in Table A1.Quantification results of the proposed method and other considered segmentation methods are provided inTable A2 and Table A3.
Table A1: Detailed architecture parameters. (Conv.: convolutional)
Layer Type × × × × × × × × × × × × × × × × ×
32 64 × × × × × × × × × ×
32 64 × × × × × × × × × ×
64 32 × × × × × × × × × ×
64 32 × × × × × × × × × ×
128 16 × × × × × × × × × ×
128 16 × × × × × × × × × ×
256 32 × × × × × ×
128 32 × × × × × × × × × ×
128 32 × × × × × × × × × ×
128 64 × × × × × ×
64 64 × × × × × × × × × ×
64 64 × × × × × × × × × ×
64 128 × × × × × ×
32 128 × × × × × × × × × ×
32 128 × × × × × × × × × ×
32 128 × × × × × × × × × able A2: Quantitative comparison between the proposed method and other segmentation methods on the basis of dice similaritycoefficient, jaccard similarity coefficient and hausdorff distance, for reconstructed images with voxel size of 2 mm. (LC/RC: left/rightcaudate; LP/RP: left/right putamen; LGP/RGP: left/right globus pallidus Proposed U-net-based 40% SUV-max Snakes MRF-GMMDSC (LC) 0.8(0.79, 0.81) 0.77(0.76, 0.78) 0.38(0.35, 0.4) 0.34(0.32, 0.35) 0.34(0.32, 0.35)DSC (RC) 0.8(0.79, 0.81) 0.77(0.76, 0.78) 0.38(0.36, 0.4) 0.33(0.32, 0.35) 0.31(0.3, 0.32)DSC (LP) 0.83(0.83, 0.84) 0.82(0.81, 0.82) 0.56(0.55, 0.58) 0.54(0.53, 0.55) 0.36(0.35, 0.37)DSC (RP) 0.85(0.84, 0.85) 0.82(0.82, 0.83) 0.57(0.56, 0.58) 0.53(0.52, 0.54) 0.41(0.4, 0.42)DSC (LGP) 0.75(0.73, 0.77) 0.71(0.69, 0.74) 0.3(0.29, 0.31) 0.28(0.27, 0.29) 0.4(0.39, 0.42)DSC (RGP) 0.79(0.78, 0.81) 0.77(0.76, 0.78) 0.3(0.3, 0.31) 0.29(0.28, 0.3) 0.37(0.35, 0.38)JSC (LC) 0.67(0.66, 0.68) 0.63(0.62, 0.64) 0.24(0.22, 0.25) 0.2(0.19, 0.21) 0.2(0.19, 0.21)JSC (RC) 0.67(0.66, 0.68) 0.63(0.61, 0.64) 0.24(0.22, 0.26) 0.2(0.19, 0.21) 0.18(0.18, 0.19)JSC (LP) 0.72(0.71, 0.72) 0.69(0.68, 0.7) 0.4(0.38, 0.41) 0.37(0.36, 0.38) 0.22(0.21, 0.23)JSC (RP) 0.73(0.73, 0.74) 0.7(0.69, 0.71) 0.4(0.39, 0.41) 0.36(0.35, 0.37) 0.26(0.25, 0.27)JSC (LGP) 0.61(0.59, 0.63) 0.57(0.54, 0.59) 0.18(0.17, 0.18) 0.17(0.16, 0.17) 0.26(0.24, 0.27)JSC (RGP) 0.66(0.65, 0.68) 0.63(0.62, 0.65) 0.18(0.17, 0.19) 0.17(0.16, 0.18) 0.23(0.22, 0.24)HD (LC) 1.98(1.8, 2.17) 2(1.9, 2.1) 9.99(9.66, 10.3) 8.21(8.06, 8.36) 7.19(7, 7.39)HD (RC) 2.15(2.02, 2.29) 2.22(2.05, 2.39) 9.91(9.62, 10.2) 8.68(8.51, 8.85) 7.9(7.72, 8.07)HD (LP) 1.84(1.76, 1.93) 2.03(1.75, 2.31) 5.25(5.04, 5.46) 6.14(5.81, 6.47) 8.4(8.21, 8.58)HD (RP) 1.86(1.77, 1.94) 1.93(1.84, 2.02) 5.03(4.84, 5.22) 5.94(5.66, 6.23) 7.89(7.69, 8.08)HD (LGP) 1.61(1.52, 1.7) 1.7(1.61, 1.8) 6.08(5.94, 6.22) 6.2(6.01, 6.39) 5.49(5.1, 5.89)HD (RGP) 1.46(1.39, 1.54) 1.6(1.53, 1.67) 6.24(6.1, 6.37) 5.96(5.85, 6.07) 6.4(6.01, 6.79) able A3: Quantitative comparison between the proposed method and other segmentation methods on the basis of dice similaritycoefficient, jaccard similarity coefficient and hausdorff distance, for reconstructed images with voxel size of 4 mm. (LC/RC: left/rightcaudate; LP/RP: left/right putamen; LGP/RGP: left/right globus pallidus Proposed U-net-based 40% SUV-max Snakes MRF-GMMDSC (LC) 0.81(0.8, 0.82) 0.64(0.63, 0.66) 0.25(0.23, 0.26) 0.2(0.19, 0.21) 0.23(0.22, 0.24)DSC (RC) 0.81(0.8, 0.82) 0.64(0.63, 0.65) 0.25(0.24, 0.27) 0.2(0.19, 0.21) 0.21(0.21, 0.22)DSC (LP) 0.85(0.84, 0.85) 0.73(0.72, 0.74) 0.43(0.42, 0.45) 0.41(0.4, 0.42) 0.28(0.26, 0.29)DSC (RP) 0.85(0.84, 0.85) 0.73(0.72, 0.74) 0.44(0.42, 0.45) 0.41(0.4, 0.41) 0.42(0.4, 0.43)DSC (LGP) 0.77(0.75, 0.78) 0.62(0.6, 0.64) 0.21(0.2, 0.22) 0.24(0.23, 0.25) 0.31(0.3, 0.32)DSC (RGP) 0.8(0.79, 0.82) 0.64(0.63, 0.66) 0.23(0.22, 0.23) 0.26(0.25, 0.26) 0.3(0.29, 0.32)JSC (LC) 0.68(0.67, 0.69) 0.48(0.46, 0.49) 0.14(0.13, 0.15) 0.11(0.11, 0.12) 0.13(0.12, 0.14)JSC (RC) 0.68(0.67, 0.7) 0.47(0.46, 0.49) 0.15(0.14, 0.15) 0.11(0.11, 0.12) 0.12(0.11, 0.13)JSC (LP) 0.73(0.72, 0.74) 0.57(0.56, 0.58) 0.28(0.27, 0.29) 0.26(0.25, 0.26) 0.16(0.15, 0.17)JSC (RP) 0.74(0.73, 0.75) 0.58(0.57, 0.59) 0.28(0.27, 0.29) 0.26(0.25, 0.26) 0.27(0.26, 0.28)JSC (LGP) 0.63(0.61, 0.65) 0.45(0.43, 0.47) 0.12(0.11, 0.12) 0.14(0.13, 0.14) 0.19(0.18, 0.19)JSC (RGP) 0.68(0.66, 0.7) 0.48(0.46, 0.5) 0.13(0.12, 0.13) 0.15(0.14, 0.15) 0.18(0.17, 0.19)HD (LC) 1.62(1.45, 1.78) 1.92(1.73, 2.1) 5.05(4.91, 5.18) 6.59(6.36, 6.81) 5.39(5.21, 5.58)HD (RC) 1.58(1.41, 1.75) 1.9(1.69, 2.12) 5.36(5.24, 5.48) 6.73(6.49, 6.96) 5.75(5.58, 5.91)HD (LP) 0.95(0.907, 0.993) 1.16(1.11, 1.21) 3.5(3.35, 3.65) 4(3.84, 4.17) 5.17(5.01, 5.33)HD (RP) 0.944(0.908, 0.981) 1.14(1.09, 1.18) 3.31(3.19, 3.42) 3.83(3.67, 3.99) 3.68(3.44, 3.92)HD (LGP) 0.958(0.902, 1.01) 1.21(1.06, 1.36) 3.97(3.88, 4.06) 3.73(3.66, 3.8) 3.16(3.02, 3.29)HD (RGP) 0.847(0.794, 0.901) 1.05(0.997, 1.11) 3.92(3.85, 3.99) 3.73(3.67, 3.79) 3.38(3.2, 3.56) eferences DaTscan (Ioflupane I 123 Injection) for Intravenous Use, CII Initial U.S. Approval. 2011. URL: .AlDakheel A, Kalia LV, Lang AE. Pathogenesis-targeted, disease-modifying therapies in parkinson disease. Neu-rotherapeutics 2014;11(1):6–23.Augimeri A, Cherubini A, Cascini GL, Galea D, Caligiuri ME, Barbagallo G, Arabia G, Quattrone A.CADA—computer-aided DaTSCAN analysis. EJNMMI Phys 2016;3(1):4.Badiavas K, Molyvda E, Iakovou I, Tsolaki M, Psarrakos K, Karatzas N. SPECT imaging evaluation in movementdisorders: far beyond visual assessment. EJNMMI Phys 2011;38(4):764–73.Brown CA, Karimi MK, Tian L, Flores H, Su Y, Tabbal SD, Loftin SK, Moerlein SM, Perlmutter JS. Valida-tion of midbrain positron emission tomography measures for nigrostriatal neurons in macaques. Ann Neurol2013;74(4):602–10.Chen J, Jha AK, Frey EC. Incorporating CT prior information in the robust fuzzy C-means algorithm for QSPECTimage segmentation. In: SPIE Medical Imaging. International Society for Optics and Photonics; volume 10949;2019. p. 109491W.Ciliax BJ, Drash GW, Staley JK, Haber S, Mobley CJ, Miller GW, Mufson EJ, Mash DC, Levey AI. Immunocy-tochemical localization of the dopamine transporter in human brain. J Comp Neurol 1999;409(1):38–56.Council NR, et al. Single Photon Emission Computed Tomography. In: Mathematics and Physics of EmergingBiomedical Imaging. National Academies Press (US); 1996. .Covington MF, McMillan NA, Avery RJ, Kuo PH. The semicolon sign: dopamine transporter imaging artifactfrom head tilt. Journal of Nuclear Medicine Technology 2013;41(2):105–7.Creswell A, Arulkumaran K, Bharath AA. On denoising autoencoders trained to minimise binary cross-entropy.arXiv preprint arXiv:170808487 2017;.Dewaraja YK, Ljungberg M, Koral KF. Monte Carlo evaluation of object shape effects in iodine-131 SPECT tumoractivity quantification. Eur J Nucl Med 2001;28(7):900–6.23jang DS, Janssen MJ, Bohnen N, Booij J, Henderson TA, Herholz K, Minoshima S, Rowe CC, Sabri O, SeibylJ, et al. SNM practice guideline for dopamine transporter imaging with 123I-Ioflupane SPECT 1.0. J Nucl Med2012;53(1):154–63.Dorsey E, Sherer T, Okun MS, Bloem BR. The emerging evidence of the Parkinson pandemic. J Parkinsons Dis2018;8(s1):S3–8.Dorsey ER, Bloem BR. The Parkinson pandemic—a call to action. JAMA Neurol 2018;75(1):9–10.Du Y, Frey EC. Quantitative evaluation of simultaneous reconstruction with model-based crosstalk compensationfor dual-isotope simultaneous acquisition brain SPECT. Med Phys 2009;36(6Part1):2021–33.Du Y, Tsui BM, Frey EC. Partial volume effect compensation for quantitative brain SPECT imaging. IEEE TransMed Imag 2005;24(8):969–76.Elkouzi A, Vedam-Mai V, Eisinger RS, Okun MS. Emerging therapies in Parkinson disease—repurposed drugs andnew approaches. Nat Rev Neurol 2019;15(4):204–23.Emamzadeh FN, Surguchov A. Parkinson’s disease: biomarkers, treatment, and risk factors. Front Neurosci2018;12:612.Fahn S, Sulzer D. Neurodegeneration and neuroprotection in Parkinson disease. NeuroRx 2004;1(1):139–54.Fischl B. Freesurfer. Neuroimage 2012;62(2):774–81.Foster B, Bagci U, Mansoor A, Xu Z, Mollura DJ. A review on segmentation of positron emission tomographyimages. Comp Biol Med 2014;50:76–96.Helmich RC, Janssen MJ, Oyen WJ, Bloem BR, Toni I. Pallidal dysfunction drives a cerebellothalamic circuit intoParkinson tremor. Ann Neurol 2011;69(2):269–81.Honkanen EA, Saari L, Orte K, Gardberg M, Noponen T, Joutsa J, Kaasinen V. No link between striatal dopamin-ergic axons and dopamine transporter imaging in parkinson’s disease. Mov Disord 2019;34(10):1562–6.Huttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing images using the Hausdorff distance. IEEE TransPattern Anal Mach Intell 1993;15(9):850–63.Hutton BF, Occhipinti M, Kuehne A, M´ath´e D, Kov´acs N, Waiczies H, Erlandsson K, Salvado D, Carminati M,Montagnani GL, et al. Development of clinical simultaneous SPECT/MRI. Br J Radiol 2018;91(1081):20160690.24ha A, Rowe S, Du Y, Frey E. An unsupervised semi-automated segmentation technique to delineate tumors andbones in SPECT images of patients with bone metastasis. J Nucl Med 2017a;58(supplement 1):613–.Jha AK, Caffo B, Frey EC. A no-gold-standard technique for objective assessment of quantitative nuclear-medicineimaging methods. Phys Med Biol 2016;61(7):2780.Jha AK, Kupinski MA, Rodriguez JJ, Stephen RM, Stopeck AT. Task-based evaluation of segmentation algorithmsfor diffusion-weighted MRI without using a gold standard. Phys Med Biol 2012;57(13):4425.Jha AK, Mena E, Caffo BS, Ashrafinia S, Rahmim A, Frey EC, Subramaniam RM. Practical no-gold-standardevaluation framework for quantitative imaging methods: application to lesion segmentation in positron emissiontomography. J Med Imag 2017b;4(1):011011.Jin X, Mulnix T, Gallezot JD, Carson RE. Evaluation of motion correction methods in human brain PET imag-ing—A simulation study based on human motion data. Medical physics 2013;40(10):102503.Joling M, Vriend C, Van den Heuvel OA, Raijmakers PG, Jones PA, Berendse HW, Booij J. Analysis of ex-trastriatal 123I-FP-CIT binding contributes to the differential diagnosis of Parkinsonian diseases. J Nucl Med2017;58(7):1117–23.Kamnitsas K, Ledig C, Newcombe VF, Simpson JP, Kane AD, Menon DK, Rueckert D, Glocker B. Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Med Image Anal 2017;36:61–78.Karimi M, Tian L, Brown CA, Flores HP, Loftin SK, Videen TO, Moerlein SM, Perlmutter JS. Validation ofnigrostriatal positron emission tomography measures: critical limits. Ann Neurol 2013;73(3):390–6.Kass M, Witkin A, Terzopoulos D. Snakes: Active contour models. Int J Comp Vision 1988;1(4):321–31.King MA, Long DT, Brill AB. SPECT volume quantitation: influence of spatial resolution, source size and shape,and voxel size. Med Phys 1991;18(5):1016–24.Kupinski MA, Hoppin JW, Clarkson E, Barrett HH, Kastis GA. Estimation in medical imaging without a goldstandard. Acad Radiol 2002;9(3):290–7.LaMontagne PJ, Benzinger TL, Morris JC, Keefe S, Hornbeck R, Xiong C, Grant E, Hassenstab J, Moulder K,Vlassenko A, et al. Oasis-3: longitudinal neuroimaging, clinical, and cognitive dataset for normal aging andalzheimer disease. medRxiv 2019;. 25ang AE, Espay AJ. Disease modification in parkinson’s disease: current approaches, challenges, and futureconsiderations. Mov Disord 2018;33(5):660–77.Lee JS, Lee DS. Analysis of functional brain images using population-based probabilistic atlas. Current MedicalImaging 2005;1(1):81–7.Leung KH, Marashdeh W, Wray R, Ashrafinia S, Pomper MG, Rahmim A, Jha AK. A physics-guided modulardeep-learning based automated framework for tumor segmentation in PET. Phys Med Biol 2020;.Lin Q, Luo M, Gao R, Li T, Man Z, Cao Y, Wang H. Deep learning based automatic segmentation of metastasishotspots in thorax bone SPECT images. PLOS One 2020;15(12):e0243253.Liu J, Liu Z, Moon HS, Mhlanga J, Jha A. A no-gold-standard technique for objective evaluation of quantitativenuclear-medicine imaging methods in the presence of correlated noise. J Nucl Med 2020;61(supplement 1):523–.Liu Z, Mhlanga JC, Laforest R, Derenoncourt PR, Siegel BA, Jha AK. An estimation-based approach to tumorsegmentation in oncological PET. 2021. arXiv:2003.00317 .Ljungberg M, Strand S, King M. The simind monte carlo program. Monte Carlo calculation in nuclear medicine:Applications in diagnostic imaging 2012;:145–63.Long DT, King MA, Sheehan J. Comparative evaluation of image segmentation methods for volume quantitationin SPECT. Med Phys 1992;19(2):483–9.Mao XJ, Shen C, Yang YB. Image restoration using very deep convolutional encoder-decoder networks withsymmetric skip connections. arXiv preprint arXiv:160309056 2016;.Martinez-Murcia F, G´orriz J, Ram´ırez J, Moreno-Caballero M, G´omez-R´ıo M, Initiative PPM. Parametriza-tion of textural patterns in 123I-ioflupane imaging for the automatic detection of Parkinsonism. Med Phys2014;41(1):012502.Matesan M, Gaddikeri S, Longfellow K, Miyaoka R, Elojeimy S, Elman S, Hu SC, Minoshima S, Lewis D. I-123 DaTscan SPECT Brain Imaging in Parkinsonian Syndromes: Utility of the Putamen-to-Caudate Ratio. JNeuroimaging 2018;28(6):629–34.McGhee DJ, Royle PL, Thompson PA, Wright DE, Zajicek JP, Counsell CE. A systematic review of biomarkersfor disease progression in parkinson’s disease. BMC Neurol 2013;13(1):35.26ignotte M, Meunier J, Soucy J, Janicki C. Segmentation and classification of brain SPECT images using 3DMarkov random field and density mixture estimations. In: Proc. 5th World Multi-Conf. Systemics, Cyberneticsand Informatics, Concepts and Applications of Systemics and Informatics. Citeseer; volume 10; 2001. p. 239–44.Nath V, Pathak SK, Schilling KG, Schneider W, Landman BA. Deep learning estimation of multi-tissue constrainedspherical deconvolution with limited single shell DW-MRI. In: SPIE Medical Imaging. International Society forOptics and Photonics; volume 11313; 2020. p. 113130S.Ouyang J, El Fakhri G, Moore SC, Kijewski MF. Fast Monte Carlo Simulation Based Joint Iterative Reconstructionfor Simultaneous 99mTc/123I Brain SPECT Imaging. In: 2006 IEEE Nuclear Science Symposium ConferenceRecord. IEEE; volume 4; 2006. p. 2251–6.Patterson H, Clarke GH, Guy R, McKay WJ. Head movement in normal subjects during simulated brain imaging.Journal of nuclear medicine technology 1998;26(4):257–61.Perlmutter JS, Norris SA. Neuroimaging biomarkers for Parkinson’s disease: Facts and fantasy. Ann Neurol2014;76(6):769–83.Porritt M, Stanic D, Finkelstein D, Batchelor P, Lockhart S, Hughes A, Kalnins R, Howells D. Dopaminergicinnervation of the human striatum in Parkinson’s disease. Mov Disord 2005;20(7):810–8.Prashanth R, Roy SD, Mandal PK, Ghosh S. High-accuracy classification of Parkinson’s disease through shapeanalysis and surface fitting in 123I-Ioflupane SPECT imaging. IEEE J Biomed Health Inform 2016;21(3):794–802.Rahmim A, Salimpour Y, Jain S, Blinder SA, Klyuzhin IS, Smith GS, Mari Z, Sossi V. Application of textureanalysis to DAT SPECT imaging: relationship to clinical assessments. NeuroImage: Clinical 2016;12:e1–9.Saari L, Kivinen K, Gardberg M, Joutsa J, Noponen T, Kaasinen V. Dopamine transporter imaging does notpredict the number of nigral neurons in parkinson disease. Neurology 2017;88(15):1461–7.Slomka PJ, Hurwitz GA, Clement GS, Stephenson J. Three-dimensional demarcation of perfusion zones corre-sponding to specific coronary arteries: application for automated interpretation of myocardial SPECT. J NuclMed 1995;36(11):2120–6.Son SJ, Kim M, Park H. Imaging analysis of parkinson’s disease patients using spect and tractography. Scientificreports 2016;6:38070. 27rivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a simple way to prevent neuralnetworks from overfitting. J Mach Learn Res 2014;15(1):1929–58.Staff RT, Ahearn TS, Wilson K, Counsell CE, Taylor K, Caslake R, Davidson JE, Gemmell HG, Murray AD.Shape analysis of 123I-N- ω -fluoropropyl-2- β -carbomethoxy-3 ββ