Note: This copy is for your personal non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at www.rsna.org/rsnarights.
Federica Agosta, MD, PhD Pilar M. Ferraro, MS Elisa Canu, PhD Massimiliano Copetti, PhD Sebastiano Galantucci, MD Giuseppe Magnani, MD Alessandra Marcone, MD Paola Valsasina, MS Alessandro Sodero, MD Giancarlo Comi, MD Andrea Falini, MD, PhD Massimo Filippi, MD
Purpose:
To test a multimodal magnetic resonance (MR) imaging– based approach composed of cortical thickness and white matter (WM) damage metrics to discriminate between variants of primary progressive aphasia (PPA) that are nonfluent and/or agrammatic (NFVPPA) and semantic (SVPPA).
Materials and Methods:
This study was approved by the local ethics committees on human studies, and written informed consent from all patients was obtained before their enrollment. T1-weighted and diffusion-tensor (DT) MR images were obtained from 13 NFVPPA patients, 13 SVPPA patients, and 23 healthy control participants. Cortical thickness and DT MR imaging indices from the long-associative and interhemispheric WM tracts were obtained. A random forest (RF) analysis was used to identify the image features associated with each clinical syndrome. Individual patient classification was performed by using receiver operator characteristic curve analysis with cortical thickness, DT MR imaging, and a combination of the two modalities.
Results:
RF analysis showed that the best markers to differentiate the two PPA variants at an individual patient level among cortical thickness and DT MR imaging metrics were diffusivity abnormalities of the left inferior longitudinal and uncinate fasciculi and cortical thickness measures of the left temporal pole and inferior frontal gyrus. A combination of cortical thickness and DT MR imaging measures (the so-called gray-matter-andWM model) was able to distinguish patients with NFVPPA and SVPPA with the following classification pattern: area under the curve, 0.91; accuracy, 0.89; sensitivity, 0.92; specificity, 0.85. Leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy was 0.86, 0.73, and 0.68 for the gray matter and WM model, the gray matter–only model, and the WM-only model, respectively).
Conclusion:
A combination of structural and DT MR imaging metrics may provide a quantitative procedure to distinguish NFVPPA and SVPPA patients at an individual patient level. The discrimination accuracies obtained suggest that the gray matter and WM model is potentially relevant for the differential diagnosis of the PPA variants in clinical practice.
1
From the Neuroimaging Research Unit (F.A., P.M.F., E.C., S.G., P.V., A.S., M.F.), Department of Neurology, Institute of Experimental Neurology (G.M., G.C., M.F.), Department of Clinical Neurosciences (A.M.), and Department of Neuroradiology and CERMAC, Division of Neuroscience (A.F.), San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Via Olgettina 60, 20132 Milan, Italy; and Biostatistics Unit, IRCCS-Ospedale Casa Sollievo della Sofferenza, San Giovanni Rotondo, Foggia, Italy (M.C.). Received August 6, 2014; revision requested October 7; revision received November 19; accepted December 5; final version accepted December 18. Address correspondence to M.F. (e-mail: [email protected]). RSNA, 2015
q
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
RSNA, 2015
q
Online supplemental material is available for this article. 1
Original Research n Neuroradiology
Differentiation between Subtypes of Primary Progressive Aphasia by Using Cortical Thickness and Diffusion-Tensor MR Imaging Measures1
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
P
rimary progressive aphasia (PPA) is a relatively rare neurodegenerative syndrome in which language impairment is the presenting and most salient feature (1). In 2011, a group of experts (2) developed consensus criteria that recognized three discrete syndrome-related variants based on specific speech and language features characteristic of each subtype: nonfluent and/or agrammatic PPA (NFVPPA), semantic PPA (SVPPA), and logopenic PPA. Clinical-pathologic studies indicate that NFVPPA is more frequently associated with a frontotemporal lobar degeneration (FTLD)-t,
Advances in Knowledge nn Among cortical thickness and diffusion-tensor (DT) MR imaging indices, the best markers to differentiate variants of primary progressive aphasia (PPA) that are nonfluent and/or agrammatic and at an individual patient level, as selected by the random forest analysis, were diffusivity abnormalities of the left inferior longitudinal (area under the curve [AUC], 0.83) and uncinate fasciculi (AUC, 0.81) and cortical thickness measures of the left temporal pole (AUC, 0.81) and inferior frontal gyrus (AUC, 0.77). nn A combination of cortical thickness and DT MR imaging measures (the so-called gray matter and white matter [WM] model) was able to distinguish patients with nonfluent and/or agrammatic PPA and semantic PPA with the following classification pattern: AUC, 0.91; accuracy, 0.89; sensitivity, 0.92; specificity, 0.85. nn Leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy: 0.86, 0.73, and 0.68 for the gray matter and WM model, gray matter–only model, and WMonly model, respectively). 2
Agosta et al
FTLD–transactive response DNA binding protein 43 kDa is the typical pathologic substrate of SVPPA, and logopenic PPA is more likely to be associated with the pathologic substrate in Alzheimer disease (3). Therefore, a correct clinical classification of PPA subtypes is of paramount importance because of the emergence of a new generation of etiologic-specific diseasemodifying treatments. Clinical presentations of PPA closely depend on the initial site of anatomic damage. Clinical classification can take advantage of imaging findings based on the patterns of atrophy and/or hypometabolism that are found (2). NFVPPA is associated with damage to the left posterior frontoinsular regions, SVPPA is associated with involvement of the left ventral and lateral portions of the anterior temporal lobes, and logopenic PPA is associated with atrophy or hypometabolism of the left temporoparietal junction (2). Despite this, challenges emerged regarding an accurate clinical diagnosis of these conditions. Recent reports (4) highlighted the relative insensitivity of some of the administered language tests that classify PPA patients into the three variants, particularly in milder cases. Although atrophy patterns associated with each variant were documented at a group level, individual cases may present with inconclusive pattern of brain atrophy. Furthermore, the initial distinctive location of brain atrophy may be lost as degeneration progresses (5,6). Therefore, the determination of further specific and reliable in vivo markers able to aid in the differential diagnosis of these three syndromes, in addition to those based on atrophy measurements, is needed. Several investigations used diffusiontensor (DT) magnetic resonance (MR)
Implication for Patient Care nn A multimodal MR approach that includes cortical thickness and DT MR imaging measures is potentially relevant for the differential diagnosis of the PPA variants in clinical practice.
imaging to interrogate the patterns of white matter (WM) abnormalities in PPA patients, which suggests a relative specificity of WM damage to language networks in these conditions (7–13). NFVPPA is predominantly associated with left-sided abnormalities of anterior tracts, including frontoparietal, frontotemporal, and subcortical projections, and SVPPA is associated with the involvement of the left ventral frontotemporal and occipital WM tracts. The role of WM damage that contributes to individual PPA patient classification, relative to and in combination with cortical atrophy measures, remains to be assessed. The purpose of this study was to test the ability of a multiparametric MR-based approach, composed of cortical thickness and diffusivity WM metrics, to determine in vivo the two major variants of PPA.
Materials and Methods This study was approved by the local ethical committee on human studies and written informed consent from all participants was obtained before they
Published online before print 10.1148/radiol.15141869 Content codes: Radiology 2015; 000:1–9 Abbreviations: AUC = area under the curve DT = diffusion tensor FTLD = frontotemporal lobar degeneration IFG = inferior frontal gyrus NFVPPA = nonfluent and/or agrammatic PPA PPA = primary progressive aphasia RF = random forest SVPPA = semantic PPA WM = white matter Author contributions: Guarantor of integrity of entire study, M.F.; study concepts/ study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, F.A., G.M., P.V., A.S.; clinical studies, G.M., A.M., A.S., A.F.; experimental studies, F.A., P.M.F., S.G., G.M., A.S.; statistical analysis, E.C., M.C., G.M., P.V., A.S., A.F.; and manuscript editing, F.A., E.C., M.C., G.M., A.M., P.V., A.S., G.C., A.F., M.F. Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
were enrolled. Participants were studied from October 2009 to April 2013.
Participants Inclusion criteria were a clinical diagnosis of sporadic NFVPPA or SVPPA according to current international criteria (2), right-handedness, and the patient was native to Italy and fluent in Italian. Patients received a comprehensive evaluation that included a structured history and a neurologic examination, neuropsychologic testing (see Appendix E1 [online]), and neuroimaging. Clinical assessment was performed by experienced neurologists (G.M. and A.M., each with 30 years of experience in clinical neurology) who were blinded to MR imaging results. Patients were excluded if they had any of the following: a family history of dementia or FTLD-related disorders; significant medical illnesses or substance abuse that could interfere with cognitive functioning; any other systemic, psychiatric, or neurologic illnesses; or other causes of focal or diffuse brain damage, including cerebrovascular disorders at routine MR imaging. A total of 31 PPA patients were screened. Three patients were excluded because of extensive cerebrovascular disorder and two patients were not able to perform a complete MR examination. We included 26 PPA patients (13 NFVPPA and 13 SVPPA patients) in the study (Table 1). Right-handed, native Italian-speaking, age- and sex-matched healthy control participants were recruited among spouses of patients and by word of mouth. Healthy control participants underwent a multidimensional assessment, including neurologic and neuropsychological evaluation, and were included only if results were in the normal range (three patients were excluded for extensive cerebrovascular disorders at MR imaging). We studied 23 control participants (Table 1). MR Imaging Study By using a 3-T MR imager, three-dimensional T1-weighted fast-field-echo and DT MR imaging sequences were performed in all study participants. The Appendix E1 (online) shows the
complete MR imaging protocol and details on cortical thickness and tractographic analysis. Briefly, WM hyperintensity load was measured on T2-weighted images by using a semiautomatic threshold-based approach as implemented in a software package (Jim5; Xinapse Systems, Northants, England; http://www.xinapse.com). Analysis of MR images was performed by two experienced observers who were blinded to the identity of the patients (P.M.F., with 3 years of experience in neuroimaging, performed cortical thickness analysis; S.G., with 6 years of experience in neuroimaging, performed the DT MR imaging analysis). Cortical reconstruction and estimation of cortical thickness were performed on the three-dimensional T1-weighted fast-field-echo images by using an image analysis suite (FreeSurfer version 5.0; Harvard Medical School, Boston, Mass; http://surfer.nmr.mgh.harvard.edu). DT MR imaging analysis was performed with software library tools (FMRIB; Oxford, England; http://www.fmrib.ox.ac. uk/fsl/fdt/index.html) and software (Jim5; Xinapse Systems). Tractography was performed in the corpus callosum and superior longitudinal, inferior longitudinal, and uncinate fasciculi. For each tract, the average mean diffusivity, fractional anisotropy, axial diffusivity, and radial diffusivity were obtained.
Statistical Analysis Demographic, clinical, and cognitive data.—Normal distribution assumption was checked by means of Q-Q plot and Shapiro-Wilk and Kolmogorov-Smirnov tests. Group comparisons were performed by using analysis of variance models followed by posthoc pairwise comparisons with Bonferroni correction. All analyses were performed by using statistical software (SAS version 9.3; SAS Institute, Cary, NC). Statistical analysis was performed by M.C., with 6 years of experience in biostatistics. MR imaging data.—A vertex-byvertex analysis was used to assess differences of cortical thickness between control and PPA patients, and between patient subgroups, by using a general linear model (FreeSurfer; Harvard
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
Agosta et al
Medical School). Maps that show between-group differences were generated by using the t statistic as a threshold, a P value of less than .05 indicative of statistical significance, and a falsediscovery rate corrected for multiple comparisons when patients and control participants were compared, and an uncorrected (cluster extent, 100 mm2) P value of .01 when patient groups were compared. The mean cortical thickness of 34 regions of interest per hemisphere (14) and mean DT MR imaging measures from WM tracts were compared between groups by using analysis-of-variance models and falsediscovery rate adjusted for multiple comparisons. Between-patient group comparisons were also performed by adjusting for disease duration. Discrimination analysis.—For each pairwise comparison, a random forest (RF) approach (15) was used to select the MR imaging measures that best predicted the two PPA phenotypes. RF is a powerful machine-learning statistical algorithm based on an ensemble of classification trees. For the RF method, 100 000 classification trees were built. The training set used to grow each tree was a 0.632+ bootstrap resample of the observations. The best split at each node was selected from a random subset of covariates. The left-out observations (ie, the so-called out-of-bag observations) were then used to obtain the classification error of each tree considered. The goodness of fit of the RF was assessed by averaging the individual tree classification errors. Furthermore, the RF framework estimated the importance of a predictor by examining how much the classification error increases when out-of-bag data for that variable were permuted, while all others were left unchanged. The importance of variables was ranked by assigning to each metric a score based on its ability to classify correctly the patients according to the increase of classification error when values of that covariate in a node were permuted randomly. For easier interpretation, variable importance was normalized with respect to the best predictor. The bootstrap resampling and the permutation strategy simulated, 3
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Table 1 Demographic, Clinical, and Neuropsychological Features of Healthy Control Participants and Patients with PPA P Value Parameter No. of patients Age (y) Sex No. of women No. of men Education (y) Disease duration (y) WMH load (mL) General cognition CDR sum of boxes MMSE Language Object knowledge Single word comprehension Confrontation naming Written language Syntactic comprehension Repetition Memory Digit span Spatial span Rey list immediate recall Rey list delay recall Rey figure recall Visuospatial abilities Rey’s figure copy Attention and executive functions Semantic fluency Phonemic fluency Raven colored progressive matrices Attentive matrices
Cut-off Value
Healthy Control Participants
NFVPPA
SVPPA
NFVPPA vs HC
SVPPA vs HC
NFVPPA vs SVPPA
… …
23 64.2 6 5.9
13 66.9 6 6.2
13 65.2 6 8.8
...
...
...
… … … … …
15 8 12.5 6 5.0 ... 0.7 6 0.7
9.3 6 6.4 2.4 6 1.2 1.5 6 1.3
6 7 10.2 6 5.1 3.4 6 1.5 1.4 6 1.2
.63 ..999 … … .24 ... .33
.90 .93 … … .52 ... .42
.90 .93 … … .60 .10 .78
… 24
... 29.3 6 0.9
2.6 6 0.9 22.2 6 6.7
3.2 6 2.7 22.1 6 7.0
... ,.001*
... ,.001*
..999 .96
… … … … … …
... ... ... ... ... ...
20.7 6 2.3 21.7 6 2.9 25.2 6 7.5 24.9 6 8.1 212.2 6 13.8 220.9 6 35.3
22.1 6 1.8 224.2 6 34.8 221.0 6 14.6 22.4 6 3.3 26.2 6 5.2 24.8 6 6.7
... ... ... ... ... ...
... ... ... ... ... ...
.20 .08 .02* ..999 .46 .21
3.75 3.75 28.53 4.69 9.47
6.0 6 1.2 5.1 6 1.2 42.3 6 9.9 8.5 6 3.3 17.6 6 6.5
4.1 6 1.2 3.4 6 0.9 30.4 6 12.3 7.3 6 4.4 9.6 6 7.5
5.1 6 1.2 4.2 6 0.9 20.4 6 12.5 3.8 6 3.54 11.6 6 7.0
28.88
33.7 6 2.4
23.4 6 8.2
25 17 18
42.7 6 8.9 38.3 6 9.5 30.3 6 3.7
31
50.7 6 7.3
9 4
.003* .06 .06 .56 .03*
.08 .06 ,.001* .003* .04*
.08 .16 .17 .14 .61
26.5 6 7.9
,.001*
.002*
.30
20.4 6 12.4 10.0 6 8.5 22.1 6 7.9
11.8 6 9.1 12.2 6 11.0 24.6 6 9.2
,.001* ,.001* .01*
,.001* ,.001* .26
.14 .72 .36
35.1 6 12.7
40.7 6 10.9
.01*
.02*
.25
Note.—Unless otherwise indicated, values are mean 6 standard deviation. Values for the language domains are expressed as z scores. Values for the other neuropsychologic tests are the scores obtained at each test, corrected for age, sex, and education, with respect to normative values. P values refer to Fisher exact test or analysis of variance models, followed by posthoc pairwise comparisons with Bonferroni correction. CDR = Clinical Dementia Rating, HC = healthy control participants, MMSE = Mini-mental State Examination, WMH = WM hyperintensity. * P value is significant.
de facto, the natural variability of measures and provided a direct validation of the results: by using a different bootstrap sample of the data and a different subset of predictors, randomly chosen to build each tree of the forest, RF overcomes the concern of false-positive discoveries. Discrimination (ie, the ability to distinguish patients who belong to one of the two PPA phenotypes) of the MR imaging variables selected by RF was assessed by receiver operating 4
characteristic curve analysis and by computing the area under the curve (AUC). Pairwise AUC comparisons were performed by using the DeLong test for paired receiver operating characteristic curves. A leave-one-out procedure was performed to achieve an internal validation of the results. We calculated accuracy, sensitivity, and specificity. The discrimination between the two PPA variants was tested and adjusted for disease duration. Analyses were performed by
using statistical software (SAS version 9.3, SAS Institute; and randomForest, R Project for Statistical Computing, Vienna, Austria).
Results Cortical Thickness Peaks of cortical thinning in NFVPPA patients compared with healthy control participants included the bilateral
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Figure 1
Figure 1: Three-dimensional reconstructed MR imaging maps show distribution of the cortical thinning on the pial surface in the left (L) and right (R ) sides of the cortex in, A, patients with NFVPPA variant compared with healthy control participants (false-discovery rate, P , .05), B, SVPPA compared with healthy control participants (false-discovery rate, P , .05), and, C, in NFVPPA compared with SVPPA patients (uncorrected P , .01). A–C, Regions of cortical thinning in NFVPPA relative to SVPPA are shown in the gray-to-yellow color keys and, C, regions of cortical thinning in SVPPA relative to NFVPPA are shown in the gray-to-cyan color key. The color key represents t values.
inferior frontal gyrus (IFG) and bilateral supplementary motor areas, and left temporoparietal junction and middle cingulate cortex (falsediscovery rate, P , .05; Fig 1, A). Smaller areas of cortical thinning were found in bilateral superior frontal gyrus and right middle cingulate cortex and superior temporal gyrus (false-discovery rate, P , .05; Fig 1, A). Peaks of cortical thinning in SVPPA patients relative to control participants were located in the left anterior temporal lobe (false-discovery rate, P , .05; Fig 1, B). A much smaller area of cortical thinning was found in the right anterior inferior temporal gyrus (falsediscovery rate, P , .05; Fig 1, B). Patterns of cortical thinning differed between the two variants: NFVPPA patients had more cortical thinning in the bilateral IFG and bilateral superior frontal gyri, left temporoparietal
junction and inferior parietal lobule and right middle frontal gyrus and supplementary motor areas, while SVPPA patients had more cortical thinning in the bilateral anterior temporal lobes (uncorrected P , .01; Fig 1, C). Adjusted for disease duration, the results were similar although the regions of temporal thinning in SVPPA relative to NFVPPA patients were relatively small (uncorrected P , .01; Fig E1 [online]). Region-of-interest analysis (Table E1 [online]) confirmed that NFVPPA relative to SVPPA patients had a greater cortical thinning of the left IFG, and SVPPA patients showed a more severe cortical thinning of the left temporal pole compared with the other group (false-discovery rate, P , .05).
DT MR Imaging Compared with healthy control participants, NFVPPA patients had increased
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
mean diffusivity, axial diffusivity, and radial diffusivity and decreased fractional anisotropy of the body and genu of the corpus callosum and left superior longitudinal fasciculus, increased mean diffusivity and radial diffusivity and decreased fractional anisotropy of the right superior longitudinal fasciculus, increased mean diffusivity and radial diffusivity of the left inferior longitudinal fasciculus, and increased mean diffusivity and axial diffusivity of the splenium of the corpus callosum (falsediscovery rate, P , .05; Tables E2, E3 [online]). Compared with control participants, SVPPA patients had the following: increased mean diffusivity, axial diffusivity, radial diffusivity and decreased fractional anisotropy of the left superior longitudinal fasciculus; increased mean diffusivity, radial diffusivity and decreased fractional 5
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Table 2 MR Imaging Variables Ranked for Their Diagnostic Importance by the RF Analysis in Patients with PPA Variants Relative to Healthy Control Participants and Each Other
Parameter NFVPPA versus control participants CC body fractional anisotropy CC body radial diffusivity Left SLF fractional anisotropy Left caudal MFG cortical thickness SVPPA versus control participants Left ILF axial diffusivity Left ILF mean diffusivity Left ILF radial diffusivity Left entorhinal cortical thickness NFVPPA versus SVPPA Left temporal pole cortical thickness Left uncinate axial diffusivity Left IFG-pars opercularis cortical thickness Left ILF axial diffusivity
Normalized Variable Importance
AUC
Accuracy
Sensitivity
Specificity
100 71.2 57.1 55.8
0.94 0.93 0.92 0.86
0.89 0.80 0.86 0.83
1.00 0.70 0.91 0.96
0.69 1.00 0.77 0.61
100 79.0 77.3 60.2
0.97 0.93 0.90 0.90
0.94 0.94 0.94 0.92
1.00 1.00 1.00 0.96
0.85 0.85 0.85 0.85
100 86.6 42.4
0.81 0.81 0.77
0.85 0.85 0.81
1.00 1.00 0.69
0.69 0.69 0.92
38.3
0.83
0.81
0.85
0.77
Note.— CC = corpus callosum, ILF = inferior longitudinal fasciculus, MFG = middle frontal gyrus, SLF = superior longitudinal fasciculus.
anisotropy of the body of the corpus callosum; increased mean diffusivity, axial diffusivity, and radial diffusivity of the bilateral uncinate fasciculus and left inferior longitudinal fasciculus; increased mean diffusivity and radial diffusivity of the genu of the corpus callosum; and increased mean diffusivity and axial diffusivity of the right inferior longitudinal fasciculus and splenium of the corpus callosum (false-discovery rate, P , .05; Tables E2, E3 [online]). Compared with NFVPPA, SVPPA cases showed increased mean diffusivity, axial diffusivity, and radial diffusivity of the left inferior longitudinal fasciculus (false-discovery rate, P , .05; Tables E2, E3 [online]). However, NFVPPA patients showed higher radial diffusivity of the body of the corpus callosum and higher radial diffusivity and lower fractional anisotropy of right superior longitudinal fasciculus compared with SVPPA patients (false-discovery rate, P , .05; Tables E2, E3 [online]). Adjusted for disease duration, the results were similar but the significance was lower (Tables E2, E3 [online]). 6
Discrimination Analysis Table 2 shows the results of the RF analysis. For each comparison, the first four MR imaging variables were provided to correctly predict, according to their importance, the diagnostic category in which each individual belonged, as well as the AUC, accuracy, sensitivity, and specificity. The highest discriminatory ability to distinguish NFVPPA patients from control participants was achieved by DT MR imaging metrics of the body of the corpus callosum and left superior longitudinal fasciculus, followed by thickness of the left caudal middle frontal gyrus. The DT MR imaging variables helped to correctly classify NFVPPA patients versus control participants with AUC greater than 0.90 (ie, more than 90% of patients were correctly classified by using an individual variable). The highest discriminatory ability to distinguish SVPPA patients from control participants was achieved by mean diffusivity and axial diffusivity values of the left inferior longitudinal fasciculus (AUCs, 0.93 and 0.97,
respectively), followed by radial diffusivity values of the same tract and cortical thickness of the left entorhinal cortex (AUC, 0.90). When the two PPA patient groups were compared with each other by adjusting for disease duration, the highest discrimination ability was achieved by axial diffusivity values of the left inferior longitudinal fasciculus (AUC, 0.83) and uncinate fasciculus (AUC, 0.81), and cortical thickness of the left temporal pole (AUC, 0.81) and IFG-pars opercularis (AUC, 0.77). Receiver operating characteristic curve analysis demonstrated that a model that combined left temporal pole and IFG–pars opercularis thickness (the so-called gray matter–only model) discriminated the two patient groups with an AUC of 0.90 (accuracy, 0.89; sensitivity, 0.85; and specificity, 0.92), while a combination of the two best DT MR imaging metrics selected by the RF analysis (the so-called WM-only model) separated NFVPPA and SVPPA patients with an AUC of 0.85 (accuracy, 0.81; sensitivity, 0.77; and specificity, 0.85) (Fig 2a, 2b). Finally, the model that combined the four selected gray matter and WM measures (the so-called gray matter and WM model) separated NFVPPA from SVPPA patients with an AUC of 0.91 (accuracy, 0.89; sensitivity, 0.92; and specificity, 0.85) (Fig 2c). Although the gray matter and WM model had a better classification pattern relative to both gray matter–only and WM-only models, AUC values were not significantly different (P = .48 compared with gray matter only and P = .24 compared with WM only by using the DeLong test). However, the leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy: 0.86, 0.73, 0.68 for the gray matter and WM model, gray matter–only model, and the WM-only model, respectively).
Discussion This study shows that a combination of structural and DT MR imaging metrics may provide a quantitative procedure to
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Figure 2
Figure 2: Graphs of receiver operator characteristic curves for discrimination between patients with nonfluent and semantic variants of PPA based on the optimal measures from each variable selected by the RF analysis. (a) Receiver operator characteristic curve for the gray matter MR imaging measures and gray matter–only model (optimal cut-off values: 1.5 for the left temporal pole thickness [th ], 1.8 for the IFG-pars opercularis thickness, and 0.7 for the combined model). (b) Receiver operator characteristic curve for the WM MR imaging measures and WM-only model (optimal cut-off values: 1.3 for the axial diffusivity of the left inferior longitudinal fasciculus [L ILF axD] and uncinate fasciculi [L UNC axD ], and 0.6 for the combined model). (c) Receiver operator characteristic curve for the model that combines the four gray matter and WM measures selected by the RF analysis and shows thickness of the left temporal pole and IFG-pars opercularis and axial diffusivity values of the left inferior longitudinal fasciculus and uncinate fasciculus (gray matter and WM model; optimal cut-off value: 0.6).
distinguish NFVPPA and SVPPA in vivo at an individual patient level. The gray matter and WM model is potentially relevant for the differential diagnosis of the PPA variants in the clinical practice. These findings also emphasize the role of DT MR imaging, in addition to cortical atrophy, for in vivo discrimination of individuals affected by the PPA variants of the FTLD spectrum. The major strengths of our report compared with previous MR imaging studies of PPA are the combination of different MR imaging modalities to explore the GM and WM damage in these patients, the application of a sophisticated statistical approach (ie, the RF analysis) (15) to select the most probable cortical thickness, and in vivo DT MR imaging predictors of each PPA variant at an individual patient level. The most intriguing finding of this study is that the combination of structural and DT MR imaging modalities resulted in a more accurate classification of individual PPA patients than each of the MR techniques in isolation. Our multimodal approach was indeed both sensitive and specific. In addition, the leave-one-out validation
analysis suggested that our results are likely to work properly in new cases. Thus, such a method might provide in vivo markers of NFVPPA and SVPPA to be implemented in clinical practice through automated, computer-based tools (16). Previous studies have combined volumetric and DT MR imaging modalities to distinguish FTLD from Alzheimer disease (17–20). Our results are consistent with their claims that a multimodal approach yields the highest discrimination accuracy. It is well established that NFVPPA and SVPPA have distinct patterns of cortical damage, according to their language deficits. Our finding of predominantly asymmetric, left greater than right, anterior temporal lobe atrophy in SVPPA and predominantly left-sided inferior and superior frontal, premotor, insular, and temporoparietal atrophy in NFVPPA is consistent with previous reports, including those from FTLD cases that were confirmed with pathologic analysis (5,21). Cortical thickness measurements were obtained in previous studies of PPA (5,22–24), although their diagnostic use was not thoroughly assessed. In our study, RF
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
analysis allowed us to identify—compared with control subjects—the cortical thicknesses of the caudal middle frontal gyrus as the best gray matter predictor of NFVPPA diagnosis and the entorhinal cortex as the best gray matter predictor of SVPPA diagnosis. For the discrimination between NFVPPA and SVPPA, our model based on cortical thickness measures of the left IFG and temporal pole resulted in a correct classification of 90% of individual patients. Frontal and anterior temporal cortical regions are related to specific language abnormalities in PPA. Left frontal lobe damage was associated with poor fluency and grammatical processing in NFVPPA patients (22). All left anterior temporal lobe structures subserving semantic memory are known to be affected in SVPPA, and entorhinal cortex, amygdala, middle and inferior temporal, and fusiform gyri were the most severely damaged regions (25). The findings of the present study extend the value of frontal and temporal cortical thickness measures as potential diagnostic markers in PPA. Similarly, an automated MR imaging–based approach with T1 images provided 7
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
an accuracy of 89% in distinguishing NFVPPA and SVPPA variants (16). A recent study used an a priori region-ofinterest–based approach and demonstrated that a combination of left IFG, temporopolar, and superior temporal sulcal thickness measures completely separated the three PPA variants (24). It is noteworthy that, for the discrimination of each PPA variant from healthy control participants, we found that DT MR imaging variables had higher classification accuracies relative to cortical thickness measures. A growing number of studies investigated WM disease in PPA by highlighting distinct disruptions of large-scale networks that contribute to language impairment. Diffusivity abnormalities of the dorsal pathways that we found in NFVPPA patients agree with results of previous DT MR imaging studies (8–13), which are related to syntactic deficits (26). Superior longitudinal fasciculus and body callosal fibers also contain connections between frontal motor and premotor regions, which may contribute to speech fluency (10). Previous reports showed that left ventral temporal WM areas are selectively vulnerable in patients with SVPPA (7–9,11–13), which is consistent with their semantic processing deficits. Pathologic studies of FTLD cases showed that although the regional extent of WM alterations roughly parallels that of cortical degeneration, there are WM abnormalities that extend beyond the expected boundaries based on cortical involvement (27). Primary WM damage in FTLD may involve the accumulation of pathologic proteins in WM regions and microglia activation (28,29). A key pathophysiologic role of WM alterations in FTLD syndromes is also suggested by using MR imaging evidence in asymptomatic mutation carriers (30). An important caveat to consider regarding interpretation of our findings is the lack of pathologic confirmation of diagnosis. However, prediction of clinical diagnosis and prediction of pathologic substrates are two separate issues, and here we only focused on the first aspect. Converging evidence suggests a greater WM burden associated with FTLD-t relative to FTLD–transactive 8
Agosta et al
response DNA binding protein 43 kDa cases (31,32), and DT MR imaging helps to provide optimal sensitivity and specificity for an in vivo FTLD-t and FTLD-transactive response DNA binding protein 43 kDa diagnosis in individual patients (32). Future studies with pathologic confirmation should explore whether our approach is useful to distinguish FTLD subtypes in PPA patients. Another main limitation of our study is the relatively small number of cases included; however, this is similar to previous studies (5,7–13,22–24,33) and reflects the rarity of these conditions. Congruence with a priori hypothesis and consistency with previous findings should mitigate this concern. In addition, disease duration was slightly longer in SVPPA compared with NFVPPA patients, which suggests that the former patients might be at a more advanced stage of the disease than the latter; nonetheless, patients were well matched on general measures of cognitive status and dementia severity, and results were adjusted for disease duration. Furthermore, we did not report quantitative findings of apraxia of speech and agrammatism of the recorded spontaneous speech samples; the quantitative analysis of extended speech production is central for the distinction between NFVPPA and logopenic PPA, because differences can be subtle. However, a qualitative approach is sufficient to allow a clinical distinction between NFVPPA and SVPPA. In this study, we did not include patients with logopenic PPA or mixed PPA. Because NFVPPA and logopenic PPA are difficult to distinguish based on linguistic measures alone and mixed PPA is difficult to classify, the next research step would be to validate our approach in a more challenging scenario. Healthy control participants did not undergo language-specific neurocognitive testing. Finally, the use of a leave-one-out method in a small dataset with a large number of covariates can be associated with a high risk of data overfitting; this was not the case in our study because the most important variables were selected by using the RF analysis, and leave-one-out method was used only as
an internal validation of AUC estimates. Certainly, the accuracy of our models should be tested by independent and larger datasets, and this highlights the urgent need for multicenter neuroimaging studies of PPA. To conclude, development of reliable and objective measures to aid in the classification of PPA subtypes is increasingly important because FTLDspecific modification treatments are likely to become available in the near future. Individuals with NFVPPA and SVPPA have important brain volumetric and diffusivity alterations, which can offer novel markers useful for differential diagnostic purposes. Disclosures of Conflicts of Interest: F.A. Activities related to the present article: author received a grant from the Italian Ministry of Health. Activities not related to the present article: author received speaker’s honoraria from Biogen Idec and Serono Symposia International Foundation; author received a grant from Arisla. Other relationships: disclosed no relevant relationships. P.M.F. disclosed no relevant relationships. E.C. disclosed no relevant relationships. M.C. disclosed no relevant relationships. S.G. disclosed no relevant relationships. G.M. disclosed no relevant relationships. A.M. disclosed no relevant relationships. P.V. disclosed no relevant relationships. A.S. disclosed no relevant relationships. G.C. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: author receives personal fees from Novartis, Teva, Sanofi, Genzyme, Merck Serono, Biogen, Bayer, Serono Symposia, Excemed, Almirall, Chugai, and Receptos. Other relationships: disclosed no relevant relationships. A.F. disclosed no relevant relationships. M.F. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: author receives fees for consulting services and/or speaking activities from Bayer Schering Pharma, Biogen Idec, Merck Serono, Teva Pharmaceutical Industries, Bayer Schering Pharma, and Biogen Idec; author’s institution receives money from grants from Bayer Schering Pharma, Biogen Idec, Merck Serono, and Teva Pharmaceutical Industries. Other relationships: disclosed no relevant relationships.
References 1. Mesulam MM. Slowly progressive aphasia without generalized dementia. Ann Neurol 1982;11(6):592–598. 2. Gorno-Tempini ML, Hillis AE, Weintraub S, et al. Classification of primary progressive aphasia and its variants. Neurology 2011;76(11): 1006–1014.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
3. Rohrer JD, Lashley T, Schott JM, et al. Clinical and neuroanatomical signatures of tissue pathology in frontotemporal lobar degeneration. Brain 2011;134(Pt 9):2565– 2581. 4. Mesulam MM, Weintraub S. Is it time to revisit the classification guidelines for primary progressive aphasia? Neurology 2014;82(13): 1108–1109. 5. Rohrer JD, Warren JD, Modat M, et al. Patterns of cortical thinning in the language variants of frontotemporal lobar degeneration. Neurology 2009;72(18):1562–1569. 6. Rogalski E, Cobia D, Harrison TM, Wieneke C, Weintraub S, Mesulam MM. Progression of language decline and cortical atrophy in subtypes of primary progressive aphasia. Neurology 2011;76(21):1804–1810. 7. Agosta F, Henry RG, Migliaccio R, et al. Language networks in semantic dementia. Brain 2010;133(Pt 1):286–299. 8. Agosta F, Scola E, Canu E, et al. White matter damage in frontotemporal lobar degeneration spectrum. Cereb Cortex 2012;22(12): 2705–2714. 9. Galantucci S, Tartaglia MC, Wilson SM, et al. White matter damage in primary progressive aphasias: a diffusion tensor tractography study. Brain 2011;134(Pt 10):3011–3029. 10. Grossman M, Powers J, Ash S, et al. Disruption of large-scale neural networks in nonfluent/agrammatic variant primary progressive aphasia associated with frontotemporal degeneration pathology. Brain Lang 2013; 127(2):106–120.
dromic variants of frontotemporal dementia. Neurology 2010;74(16):1279–1287. 14. Desikan RS, Ségonne F, Fischl B, et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 2006;31(3):968–980. 15. Breiman L. Random forests. Mach Learn 2001;45(1):5–32. 16. Wilson SM, Ogar JM, Laluz V, et al. Automated MRI-based classification of primary progressive aphasia variants. Neuroimage 2009;47(4):1558–1567. 17. Avants BB, Cook PA, Ungar L, Gee JC, Grossman M. Dementia induces correlated reductions in white matter integrity and cortical thickness: a multivariate neuroimaging study with sparse canonical correlation analysis. Neuroimage 2010;50(3):1004– 1016.
24. Sapolsky D, Bakkour A, Negreira A, et al. Cortical neuroanatomic correlates of symptom severity in primary progressive aphasia. Neurology 2010;75(4):358–366. 25. Chan D, Fox NC, Scahill RI, et al. Pat terns of temporal lobe atrophy in semantic dementia and Alzheimer’s disease. Ann Neurol 2001;49(4):433–442. 26. Wilson SM, Galantucci S, Tartaglia MC, et al. Syntactic processing depends on dorsal language tracts. Neuron 2011;72(2):397–403. 27. Englund E, Brun A. Frontal lobe degeneration of non-Alzheimer type. IV. White matter changes. Arch Gerontol Geriatr 1987;6(3): 235–243. 28. Neumann M, Kwong LK, Truax AC, et al. TDP-43-positive white matter pathology in frontotemporal lobar degeneration with ubiquitin-positive inclusions. J Neuropathol Exp Neurol 2007;66(3):177–183.
18. Zhang Y, Schuff N, Ching C, et al. Joint assessment of structural, perfusion, and diffusion MRI in Alzheimer’s disease and frontotemporal dementia. Int J Alzheimers Dis 2011;2011:546871.
29. Schofield E, Kersaitis C, Shepherd CE, Kril JJ, Halliday GM. Severity of gliosis in Pick’s disease and frontotemporal lobar degeneration: tau-positive glia differentiate these disorders. Brain 2003;126(Pt 4):827–840.
19. McMillan CT, Brun C, Siddiqui S, et al. White matter imaging contributes to the multimodal diagnosis of frontotemporal lobar degeneration. Neurology 2012;78(22):1761–1768.
30. Dopper EG, Rombouts SA, Jiskoot LC, et al. Structural and functional brain connectivity in presymptomatic familial frontotemporal dementia. Neurology 2013;80(9):814–823.
20. McMillan CT, Avants BB, Cook P, Ungar L, Trojanowski JQ, Grossman M. The power of neuroimaging biomarkers for screening frontotemporal dementia. Hum Brain Mapp 2014;35(9):4827–4840.
11. Mahoney CJ, Malone IB, Ridgway GR, et al. White matter tract signatures of the progressive aphasias. Neurobiol Aging 2013;34(6): 1687–1699.
21. Gorno-Tempini ML, Dronkers NF, Rankin KP, et al. Cognition and anatomy in three variants of primary progressive aphasia. Ann Neurol 2004;55(3):335–346.
12. Sajjadi SA, Acosta-Cabronero J, Patterson K, Diaz-de-Grenu LZ, Williams GB, Nestor PJ. Diffusion tensor magnetic resonance imaging for single subject diagnosis in neurodegenerative diseases. Brain 2013;136(Pt 7): 2253–2261.
22. Gunawardena D, Ash S, McMillan C, Avants B, Gee J, Grossman M. Why are patients with progressive nonfluent aphasia nonfluent? Neurology 2010;75(7):588–594.
13. Whitwell JL, Avula R, Senjem ML, et al. Gray and white matter water diffusion in the syn-
Agosta et al
31. Caso F, Mandelli ML, Henry M, et al. In vivo signatures of nonfluent/agrammatic primary progressive aphasia caused by FTLD pathology. Neurology 2014;82(3):239–247. 32. McMillan CT, Irwin DJ, Avants BB, et al. White matter imaging helps dissociate tau from TDP-43 in frontotemporal lobar degeneration. J Neurol Neurosurg Psychiatry 2013;84(9):949–955. 33. Shim H, Hurley RS, Rogalski E, Mesulam MM. Anatomic, clinical, and neuropsychological correlates of spelling errors in primary progressive aphasia. Neuropsychologia 2012;50(8):1929–1935.
23. Gefen T, Wieneke C, Martersteck A, et al. Naming vs knowing faces in primary progressive aphasia: a tale of 2 hemispheres. Neurology 2013;81(7):658–664.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
9
Federica Agosta, MD, PhD Pilar M. Ferraro, MS Elisa Canu, PhD Massimiliano Copetti, PhD Sebastiano Galantucci, MD Giuseppe Magnani, MD Alessandra Marcone, MD Paola Valsasina, MS Alessandro Sodero, MD Giancarlo Comi, MD Andrea Falini, MD, PhD Massimo Filippi, MD
Purpose:
To test a multimodal magnetic resonance (MR) imaging– based approach composed of cortical thickness and white matter (WM) damage metrics to discriminate between variants of primary progressive aphasia (PPA) that are nonfluent and/or agrammatic (NFVPPA) and semantic (SVPPA).
Materials and Methods:
This study was approved by the local ethics committees on human studies, and written informed consent from all patients was obtained before their enrollment. T1-weighted and diffusion-tensor (DT) MR images were obtained from 13 NFVPPA patients, 13 SVPPA patients, and 23 healthy control participants. Cortical thickness and DT MR imaging indices from the long-associative and interhemispheric WM tracts were obtained. A random forest (RF) analysis was used to identify the image features associated with each clinical syndrome. Individual patient classification was performed by using receiver operator characteristic curve analysis with cortical thickness, DT MR imaging, and a combination of the two modalities.
Results:
RF analysis showed that the best markers to differentiate the two PPA variants at an individual patient level among cortical thickness and DT MR imaging metrics were diffusivity abnormalities of the left inferior longitudinal and uncinate fasciculi and cortical thickness measures of the left temporal pole and inferior frontal gyrus. A combination of cortical thickness and DT MR imaging measures (the so-called gray-matter-andWM model) was able to distinguish patients with NFVPPA and SVPPA with the following classification pattern: area under the curve, 0.91; accuracy, 0.89; sensitivity, 0.92; specificity, 0.85. Leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy was 0.86, 0.73, and 0.68 for the gray matter and WM model, the gray matter–only model, and the WM-only model, respectively).
Conclusion:
A combination of structural and DT MR imaging metrics may provide a quantitative procedure to distinguish NFVPPA and SVPPA patients at an individual patient level. The discrimination accuracies obtained suggest that the gray matter and WM model is potentially relevant for the differential diagnosis of the PPA variants in clinical practice.
1
From the Neuroimaging Research Unit (F.A., P.M.F., E.C., S.G., P.V., A.S., M.F.), Department of Neurology, Institute of Experimental Neurology (G.M., G.C., M.F.), Department of Clinical Neurosciences (A.M.), and Department of Neuroradiology and CERMAC, Division of Neuroscience (A.F.), San Raffaele Scientific Institute, Vita-Salute San Raffaele University, Via Olgettina 60, 20132 Milan, Italy; and Biostatistics Unit, IRCCS-Ospedale Casa Sollievo della Sofferenza, San Giovanni Rotondo, Foggia, Italy (M.C.). Received August 6, 2014; revision requested October 7; revision received November 19; accepted December 5; final version accepted December 18. Address correspondence to M.F. (e-mail: [email protected]). RSNA, 2015
q
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
RSNA, 2015
q
Online supplemental material is available for this article. 1
Original Research n Neuroradiology
Differentiation between Subtypes of Primary Progressive Aphasia by Using Cortical Thickness and Diffusion-Tensor MR Imaging Measures1
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
P
rimary progressive aphasia (PPA) is a relatively rare neurodegenerative syndrome in which language impairment is the presenting and most salient feature (1). In 2011, a group of experts (2) developed consensus criteria that recognized three discrete syndrome-related variants based on specific speech and language features characteristic of each subtype: nonfluent and/or agrammatic PPA (NFVPPA), semantic PPA (SVPPA), and logopenic PPA. Clinical-pathologic studies indicate that NFVPPA is more frequently associated with a frontotemporal lobar degeneration (FTLD)-t,
Advances in Knowledge nn Among cortical thickness and diffusion-tensor (DT) MR imaging indices, the best markers to differentiate variants of primary progressive aphasia (PPA) that are nonfluent and/or agrammatic and at an individual patient level, as selected by the random forest analysis, were diffusivity abnormalities of the left inferior longitudinal (area under the curve [AUC], 0.83) and uncinate fasciculi (AUC, 0.81) and cortical thickness measures of the left temporal pole (AUC, 0.81) and inferior frontal gyrus (AUC, 0.77). nn A combination of cortical thickness and DT MR imaging measures (the so-called gray matter and white matter [WM] model) was able to distinguish patients with nonfluent and/or agrammatic PPA and semantic PPA with the following classification pattern: AUC, 0.91; accuracy, 0.89; sensitivity, 0.92; specificity, 0.85. nn Leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy: 0.86, 0.73, and 0.68 for the gray matter and WM model, gray matter–only model, and WMonly model, respectively). 2
Agosta et al
FTLD–transactive response DNA binding protein 43 kDa is the typical pathologic substrate of SVPPA, and logopenic PPA is more likely to be associated with the pathologic substrate in Alzheimer disease (3). Therefore, a correct clinical classification of PPA subtypes is of paramount importance because of the emergence of a new generation of etiologic-specific diseasemodifying treatments. Clinical presentations of PPA closely depend on the initial site of anatomic damage. Clinical classification can take advantage of imaging findings based on the patterns of atrophy and/or hypometabolism that are found (2). NFVPPA is associated with damage to the left posterior frontoinsular regions, SVPPA is associated with involvement of the left ventral and lateral portions of the anterior temporal lobes, and logopenic PPA is associated with atrophy or hypometabolism of the left temporoparietal junction (2). Despite this, challenges emerged regarding an accurate clinical diagnosis of these conditions. Recent reports (4) highlighted the relative insensitivity of some of the administered language tests that classify PPA patients into the three variants, particularly in milder cases. Although atrophy patterns associated with each variant were documented at a group level, individual cases may present with inconclusive pattern of brain atrophy. Furthermore, the initial distinctive location of brain atrophy may be lost as degeneration progresses (5,6). Therefore, the determination of further specific and reliable in vivo markers able to aid in the differential diagnosis of these three syndromes, in addition to those based on atrophy measurements, is needed. Several investigations used diffusiontensor (DT) magnetic resonance (MR)
Implication for Patient Care nn A multimodal MR approach that includes cortical thickness and DT MR imaging measures is potentially relevant for the differential diagnosis of the PPA variants in clinical practice.
imaging to interrogate the patterns of white matter (WM) abnormalities in PPA patients, which suggests a relative specificity of WM damage to language networks in these conditions (7–13). NFVPPA is predominantly associated with left-sided abnormalities of anterior tracts, including frontoparietal, frontotemporal, and subcortical projections, and SVPPA is associated with the involvement of the left ventral frontotemporal and occipital WM tracts. The role of WM damage that contributes to individual PPA patient classification, relative to and in combination with cortical atrophy measures, remains to be assessed. The purpose of this study was to test the ability of a multiparametric MR-based approach, composed of cortical thickness and diffusivity WM metrics, to determine in vivo the two major variants of PPA.
Materials and Methods This study was approved by the local ethical committee on human studies and written informed consent from all participants was obtained before they
Published online before print 10.1148/radiol.15141869 Content codes: Radiology 2015; 000:1–9 Abbreviations: AUC = area under the curve DT = diffusion tensor FTLD = frontotemporal lobar degeneration IFG = inferior frontal gyrus NFVPPA = nonfluent and/or agrammatic PPA PPA = primary progressive aphasia RF = random forest SVPPA = semantic PPA WM = white matter Author contributions: Guarantor of integrity of entire study, M.F.; study concepts/ study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, F.A., G.M., P.V., A.S.; clinical studies, G.M., A.M., A.S., A.F.; experimental studies, F.A., P.M.F., S.G., G.M., A.S.; statistical analysis, E.C., M.C., G.M., P.V., A.S., A.F.; and manuscript editing, F.A., E.C., M.C., G.M., A.M., P.V., A.S., G.C., A.F., M.F. Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
were enrolled. Participants were studied from October 2009 to April 2013.
Participants Inclusion criteria were a clinical diagnosis of sporadic NFVPPA or SVPPA according to current international criteria (2), right-handedness, and the patient was native to Italy and fluent in Italian. Patients received a comprehensive evaluation that included a structured history and a neurologic examination, neuropsychologic testing (see Appendix E1 [online]), and neuroimaging. Clinical assessment was performed by experienced neurologists (G.M. and A.M., each with 30 years of experience in clinical neurology) who were blinded to MR imaging results. Patients were excluded if they had any of the following: a family history of dementia or FTLD-related disorders; significant medical illnesses or substance abuse that could interfere with cognitive functioning; any other systemic, psychiatric, or neurologic illnesses; or other causes of focal or diffuse brain damage, including cerebrovascular disorders at routine MR imaging. A total of 31 PPA patients were screened. Three patients were excluded because of extensive cerebrovascular disorder and two patients were not able to perform a complete MR examination. We included 26 PPA patients (13 NFVPPA and 13 SVPPA patients) in the study (Table 1). Right-handed, native Italian-speaking, age- and sex-matched healthy control participants were recruited among spouses of patients and by word of mouth. Healthy control participants underwent a multidimensional assessment, including neurologic and neuropsychological evaluation, and were included only if results were in the normal range (three patients were excluded for extensive cerebrovascular disorders at MR imaging). We studied 23 control participants (Table 1). MR Imaging Study By using a 3-T MR imager, three-dimensional T1-weighted fast-field-echo and DT MR imaging sequences were performed in all study participants. The Appendix E1 (online) shows the
complete MR imaging protocol and details on cortical thickness and tractographic analysis. Briefly, WM hyperintensity load was measured on T2-weighted images by using a semiautomatic threshold-based approach as implemented in a software package (Jim5; Xinapse Systems, Northants, England; http://www.xinapse.com). Analysis of MR images was performed by two experienced observers who were blinded to the identity of the patients (P.M.F., with 3 years of experience in neuroimaging, performed cortical thickness analysis; S.G., with 6 years of experience in neuroimaging, performed the DT MR imaging analysis). Cortical reconstruction and estimation of cortical thickness were performed on the three-dimensional T1-weighted fast-field-echo images by using an image analysis suite (FreeSurfer version 5.0; Harvard Medical School, Boston, Mass; http://surfer.nmr.mgh.harvard.edu). DT MR imaging analysis was performed with software library tools (FMRIB; Oxford, England; http://www.fmrib.ox.ac. uk/fsl/fdt/index.html) and software (Jim5; Xinapse Systems). Tractography was performed in the corpus callosum and superior longitudinal, inferior longitudinal, and uncinate fasciculi. For each tract, the average mean diffusivity, fractional anisotropy, axial diffusivity, and radial diffusivity were obtained.
Statistical Analysis Demographic, clinical, and cognitive data.—Normal distribution assumption was checked by means of Q-Q plot and Shapiro-Wilk and Kolmogorov-Smirnov tests. Group comparisons were performed by using analysis of variance models followed by posthoc pairwise comparisons with Bonferroni correction. All analyses were performed by using statistical software (SAS version 9.3; SAS Institute, Cary, NC). Statistical analysis was performed by M.C., with 6 years of experience in biostatistics. MR imaging data.—A vertex-byvertex analysis was used to assess differences of cortical thickness between control and PPA patients, and between patient subgroups, by using a general linear model (FreeSurfer; Harvard
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
Agosta et al
Medical School). Maps that show between-group differences were generated by using the t statistic as a threshold, a P value of less than .05 indicative of statistical significance, and a falsediscovery rate corrected for multiple comparisons when patients and control participants were compared, and an uncorrected (cluster extent, 100 mm2) P value of .01 when patient groups were compared. The mean cortical thickness of 34 regions of interest per hemisphere (14) and mean DT MR imaging measures from WM tracts were compared between groups by using analysis-of-variance models and falsediscovery rate adjusted for multiple comparisons. Between-patient group comparisons were also performed by adjusting for disease duration. Discrimination analysis.—For each pairwise comparison, a random forest (RF) approach (15) was used to select the MR imaging measures that best predicted the two PPA phenotypes. RF is a powerful machine-learning statistical algorithm based on an ensemble of classification trees. For the RF method, 100 000 classification trees were built. The training set used to grow each tree was a 0.632+ bootstrap resample of the observations. The best split at each node was selected from a random subset of covariates. The left-out observations (ie, the so-called out-of-bag observations) were then used to obtain the classification error of each tree considered. The goodness of fit of the RF was assessed by averaging the individual tree classification errors. Furthermore, the RF framework estimated the importance of a predictor by examining how much the classification error increases when out-of-bag data for that variable were permuted, while all others were left unchanged. The importance of variables was ranked by assigning to each metric a score based on its ability to classify correctly the patients according to the increase of classification error when values of that covariate in a node were permuted randomly. For easier interpretation, variable importance was normalized with respect to the best predictor. The bootstrap resampling and the permutation strategy simulated, 3
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Table 1 Demographic, Clinical, and Neuropsychological Features of Healthy Control Participants and Patients with PPA P Value Parameter No. of patients Age (y) Sex No. of women No. of men Education (y) Disease duration (y) WMH load (mL) General cognition CDR sum of boxes MMSE Language Object knowledge Single word comprehension Confrontation naming Written language Syntactic comprehension Repetition Memory Digit span Spatial span Rey list immediate recall Rey list delay recall Rey figure recall Visuospatial abilities Rey’s figure copy Attention and executive functions Semantic fluency Phonemic fluency Raven colored progressive matrices Attentive matrices
Cut-off Value
Healthy Control Participants
NFVPPA
SVPPA
NFVPPA vs HC
SVPPA vs HC
NFVPPA vs SVPPA
… …
23 64.2 6 5.9
13 66.9 6 6.2
13 65.2 6 8.8
...
...
...
… … … … …
15 8 12.5 6 5.0 ... 0.7 6 0.7
9.3 6 6.4 2.4 6 1.2 1.5 6 1.3
6 7 10.2 6 5.1 3.4 6 1.5 1.4 6 1.2
.63 ..999 … … .24 ... .33
.90 .93 … … .52 ... .42
.90 .93 … … .60 .10 .78
… 24
... 29.3 6 0.9
2.6 6 0.9 22.2 6 6.7
3.2 6 2.7 22.1 6 7.0
... ,.001*
... ,.001*
..999 .96
… … … … … …
... ... ... ... ... ...
20.7 6 2.3 21.7 6 2.9 25.2 6 7.5 24.9 6 8.1 212.2 6 13.8 220.9 6 35.3
22.1 6 1.8 224.2 6 34.8 221.0 6 14.6 22.4 6 3.3 26.2 6 5.2 24.8 6 6.7
... ... ... ... ... ...
... ... ... ... ... ...
.20 .08 .02* ..999 .46 .21
3.75 3.75 28.53 4.69 9.47
6.0 6 1.2 5.1 6 1.2 42.3 6 9.9 8.5 6 3.3 17.6 6 6.5
4.1 6 1.2 3.4 6 0.9 30.4 6 12.3 7.3 6 4.4 9.6 6 7.5
5.1 6 1.2 4.2 6 0.9 20.4 6 12.5 3.8 6 3.54 11.6 6 7.0
28.88
33.7 6 2.4
23.4 6 8.2
25 17 18
42.7 6 8.9 38.3 6 9.5 30.3 6 3.7
31
50.7 6 7.3
9 4
.003* .06 .06 .56 .03*
.08 .06 ,.001* .003* .04*
.08 .16 .17 .14 .61
26.5 6 7.9
,.001*
.002*
.30
20.4 6 12.4 10.0 6 8.5 22.1 6 7.9
11.8 6 9.1 12.2 6 11.0 24.6 6 9.2
,.001* ,.001* .01*
,.001* ,.001* .26
.14 .72 .36
35.1 6 12.7
40.7 6 10.9
.01*
.02*
.25
Note.—Unless otherwise indicated, values are mean 6 standard deviation. Values for the language domains are expressed as z scores. Values for the other neuropsychologic tests are the scores obtained at each test, corrected for age, sex, and education, with respect to normative values. P values refer to Fisher exact test or analysis of variance models, followed by posthoc pairwise comparisons with Bonferroni correction. CDR = Clinical Dementia Rating, HC = healthy control participants, MMSE = Mini-mental State Examination, WMH = WM hyperintensity. * P value is significant.
de facto, the natural variability of measures and provided a direct validation of the results: by using a different bootstrap sample of the data and a different subset of predictors, randomly chosen to build each tree of the forest, RF overcomes the concern of false-positive discoveries. Discrimination (ie, the ability to distinguish patients who belong to one of the two PPA phenotypes) of the MR imaging variables selected by RF was assessed by receiver operating 4
characteristic curve analysis and by computing the area under the curve (AUC). Pairwise AUC comparisons were performed by using the DeLong test for paired receiver operating characteristic curves. A leave-one-out procedure was performed to achieve an internal validation of the results. We calculated accuracy, sensitivity, and specificity. The discrimination between the two PPA variants was tested and adjusted for disease duration. Analyses were performed by
using statistical software (SAS version 9.3, SAS Institute; and randomForest, R Project for Statistical Computing, Vienna, Austria).
Results Cortical Thickness Peaks of cortical thinning in NFVPPA patients compared with healthy control participants included the bilateral
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Figure 1
Figure 1: Three-dimensional reconstructed MR imaging maps show distribution of the cortical thinning on the pial surface in the left (L) and right (R ) sides of the cortex in, A, patients with NFVPPA variant compared with healthy control participants (false-discovery rate, P , .05), B, SVPPA compared with healthy control participants (false-discovery rate, P , .05), and, C, in NFVPPA compared with SVPPA patients (uncorrected P , .01). A–C, Regions of cortical thinning in NFVPPA relative to SVPPA are shown in the gray-to-yellow color keys and, C, regions of cortical thinning in SVPPA relative to NFVPPA are shown in the gray-to-cyan color key. The color key represents t values.
inferior frontal gyrus (IFG) and bilateral supplementary motor areas, and left temporoparietal junction and middle cingulate cortex (falsediscovery rate, P , .05; Fig 1, A). Smaller areas of cortical thinning were found in bilateral superior frontal gyrus and right middle cingulate cortex and superior temporal gyrus (false-discovery rate, P , .05; Fig 1, A). Peaks of cortical thinning in SVPPA patients relative to control participants were located in the left anterior temporal lobe (false-discovery rate, P , .05; Fig 1, B). A much smaller area of cortical thinning was found in the right anterior inferior temporal gyrus (falsediscovery rate, P , .05; Fig 1, B). Patterns of cortical thinning differed between the two variants: NFVPPA patients had more cortical thinning in the bilateral IFG and bilateral superior frontal gyri, left temporoparietal
junction and inferior parietal lobule and right middle frontal gyrus and supplementary motor areas, while SVPPA patients had more cortical thinning in the bilateral anterior temporal lobes (uncorrected P , .01; Fig 1, C). Adjusted for disease duration, the results were similar although the regions of temporal thinning in SVPPA relative to NFVPPA patients were relatively small (uncorrected P , .01; Fig E1 [online]). Region-of-interest analysis (Table E1 [online]) confirmed that NFVPPA relative to SVPPA patients had a greater cortical thinning of the left IFG, and SVPPA patients showed a more severe cortical thinning of the left temporal pole compared with the other group (false-discovery rate, P , .05).
DT MR Imaging Compared with healthy control participants, NFVPPA patients had increased
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
mean diffusivity, axial diffusivity, and radial diffusivity and decreased fractional anisotropy of the body and genu of the corpus callosum and left superior longitudinal fasciculus, increased mean diffusivity and radial diffusivity and decreased fractional anisotropy of the right superior longitudinal fasciculus, increased mean diffusivity and radial diffusivity of the left inferior longitudinal fasciculus, and increased mean diffusivity and axial diffusivity of the splenium of the corpus callosum (falsediscovery rate, P , .05; Tables E2, E3 [online]). Compared with control participants, SVPPA patients had the following: increased mean diffusivity, axial diffusivity, radial diffusivity and decreased fractional anisotropy of the left superior longitudinal fasciculus; increased mean diffusivity, radial diffusivity and decreased fractional 5
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Table 2 MR Imaging Variables Ranked for Their Diagnostic Importance by the RF Analysis in Patients with PPA Variants Relative to Healthy Control Participants and Each Other
Parameter NFVPPA versus control participants CC body fractional anisotropy CC body radial diffusivity Left SLF fractional anisotropy Left caudal MFG cortical thickness SVPPA versus control participants Left ILF axial diffusivity Left ILF mean diffusivity Left ILF radial diffusivity Left entorhinal cortical thickness NFVPPA versus SVPPA Left temporal pole cortical thickness Left uncinate axial diffusivity Left IFG-pars opercularis cortical thickness Left ILF axial diffusivity
Normalized Variable Importance
AUC
Accuracy
Sensitivity
Specificity
100 71.2 57.1 55.8
0.94 0.93 0.92 0.86
0.89 0.80 0.86 0.83
1.00 0.70 0.91 0.96
0.69 1.00 0.77 0.61
100 79.0 77.3 60.2
0.97 0.93 0.90 0.90
0.94 0.94 0.94 0.92
1.00 1.00 1.00 0.96
0.85 0.85 0.85 0.85
100 86.6 42.4
0.81 0.81 0.77
0.85 0.85 0.81
1.00 1.00 0.69
0.69 0.69 0.92
38.3
0.83
0.81
0.85
0.77
Note.— CC = corpus callosum, ILF = inferior longitudinal fasciculus, MFG = middle frontal gyrus, SLF = superior longitudinal fasciculus.
anisotropy of the body of the corpus callosum; increased mean diffusivity, axial diffusivity, and radial diffusivity of the bilateral uncinate fasciculus and left inferior longitudinal fasciculus; increased mean diffusivity and radial diffusivity of the genu of the corpus callosum; and increased mean diffusivity and axial diffusivity of the right inferior longitudinal fasciculus and splenium of the corpus callosum (false-discovery rate, P , .05; Tables E2, E3 [online]). Compared with NFVPPA, SVPPA cases showed increased mean diffusivity, axial diffusivity, and radial diffusivity of the left inferior longitudinal fasciculus (false-discovery rate, P , .05; Tables E2, E3 [online]). However, NFVPPA patients showed higher radial diffusivity of the body of the corpus callosum and higher radial diffusivity and lower fractional anisotropy of right superior longitudinal fasciculus compared with SVPPA patients (false-discovery rate, P , .05; Tables E2, E3 [online]). Adjusted for disease duration, the results were similar but the significance was lower (Tables E2, E3 [online]). 6
Discrimination Analysis Table 2 shows the results of the RF analysis. For each comparison, the first four MR imaging variables were provided to correctly predict, according to their importance, the diagnostic category in which each individual belonged, as well as the AUC, accuracy, sensitivity, and specificity. The highest discriminatory ability to distinguish NFVPPA patients from control participants was achieved by DT MR imaging metrics of the body of the corpus callosum and left superior longitudinal fasciculus, followed by thickness of the left caudal middle frontal gyrus. The DT MR imaging variables helped to correctly classify NFVPPA patients versus control participants with AUC greater than 0.90 (ie, more than 90% of patients were correctly classified by using an individual variable). The highest discriminatory ability to distinguish SVPPA patients from control participants was achieved by mean diffusivity and axial diffusivity values of the left inferior longitudinal fasciculus (AUCs, 0.93 and 0.97,
respectively), followed by radial diffusivity values of the same tract and cortical thickness of the left entorhinal cortex (AUC, 0.90). When the two PPA patient groups were compared with each other by adjusting for disease duration, the highest discrimination ability was achieved by axial diffusivity values of the left inferior longitudinal fasciculus (AUC, 0.83) and uncinate fasciculus (AUC, 0.81), and cortical thickness of the left temporal pole (AUC, 0.81) and IFG-pars opercularis (AUC, 0.77). Receiver operating characteristic curve analysis demonstrated that a model that combined left temporal pole and IFG–pars opercularis thickness (the so-called gray matter–only model) discriminated the two patient groups with an AUC of 0.90 (accuracy, 0.89; sensitivity, 0.85; and specificity, 0.92), while a combination of the two best DT MR imaging metrics selected by the RF analysis (the so-called WM-only model) separated NFVPPA and SVPPA patients with an AUC of 0.85 (accuracy, 0.81; sensitivity, 0.77; and specificity, 0.85) (Fig 2a, 2b). Finally, the model that combined the four selected gray matter and WM measures (the so-called gray matter and WM model) separated NFVPPA from SVPPA patients with an AUC of 0.91 (accuracy, 0.89; sensitivity, 0.92; and specificity, 0.85) (Fig 2c). Although the gray matter and WM model had a better classification pattern relative to both gray matter–only and WM-only models, AUC values were not significantly different (P = .48 compared with gray matter only and P = .24 compared with WM only by using the DeLong test). However, the leave-one-out analysis demonstrated that the gray matter and WM model is more robust than the single MR modality models to distinguish PPA variants (accuracy: 0.86, 0.73, 0.68 for the gray matter and WM model, gray matter–only model, and the WM-only model, respectively).
Discussion This study shows that a combination of structural and DT MR imaging metrics may provide a quantitative procedure to
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
Agosta et al
Figure 2
Figure 2: Graphs of receiver operator characteristic curves for discrimination between patients with nonfluent and semantic variants of PPA based on the optimal measures from each variable selected by the RF analysis. (a) Receiver operator characteristic curve for the gray matter MR imaging measures and gray matter–only model (optimal cut-off values: 1.5 for the left temporal pole thickness [th ], 1.8 for the IFG-pars opercularis thickness, and 0.7 for the combined model). (b) Receiver operator characteristic curve for the WM MR imaging measures and WM-only model (optimal cut-off values: 1.3 for the axial diffusivity of the left inferior longitudinal fasciculus [L ILF axD] and uncinate fasciculi [L UNC axD ], and 0.6 for the combined model). (c) Receiver operator characteristic curve for the model that combines the four gray matter and WM measures selected by the RF analysis and shows thickness of the left temporal pole and IFG-pars opercularis and axial diffusivity values of the left inferior longitudinal fasciculus and uncinate fasciculus (gray matter and WM model; optimal cut-off value: 0.6).
distinguish NFVPPA and SVPPA in vivo at an individual patient level. The gray matter and WM model is potentially relevant for the differential diagnosis of the PPA variants in the clinical practice. These findings also emphasize the role of DT MR imaging, in addition to cortical atrophy, for in vivo discrimination of individuals affected by the PPA variants of the FTLD spectrum. The major strengths of our report compared with previous MR imaging studies of PPA are the combination of different MR imaging modalities to explore the GM and WM damage in these patients, the application of a sophisticated statistical approach (ie, the RF analysis) (15) to select the most probable cortical thickness, and in vivo DT MR imaging predictors of each PPA variant at an individual patient level. The most intriguing finding of this study is that the combination of structural and DT MR imaging modalities resulted in a more accurate classification of individual PPA patients than each of the MR techniques in isolation. Our multimodal approach was indeed both sensitive and specific. In addition, the leave-one-out validation
analysis suggested that our results are likely to work properly in new cases. Thus, such a method might provide in vivo markers of NFVPPA and SVPPA to be implemented in clinical practice through automated, computer-based tools (16). Previous studies have combined volumetric and DT MR imaging modalities to distinguish FTLD from Alzheimer disease (17–20). Our results are consistent with their claims that a multimodal approach yields the highest discrimination accuracy. It is well established that NFVPPA and SVPPA have distinct patterns of cortical damage, according to their language deficits. Our finding of predominantly asymmetric, left greater than right, anterior temporal lobe atrophy in SVPPA and predominantly left-sided inferior and superior frontal, premotor, insular, and temporoparietal atrophy in NFVPPA is consistent with previous reports, including those from FTLD cases that were confirmed with pathologic analysis (5,21). Cortical thickness measurements were obtained in previous studies of PPA (5,22–24), although their diagnostic use was not thoroughly assessed. In our study, RF
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
analysis allowed us to identify—compared with control subjects—the cortical thicknesses of the caudal middle frontal gyrus as the best gray matter predictor of NFVPPA diagnosis and the entorhinal cortex as the best gray matter predictor of SVPPA diagnosis. For the discrimination between NFVPPA and SVPPA, our model based on cortical thickness measures of the left IFG and temporal pole resulted in a correct classification of 90% of individual patients. Frontal and anterior temporal cortical regions are related to specific language abnormalities in PPA. Left frontal lobe damage was associated with poor fluency and grammatical processing in NFVPPA patients (22). All left anterior temporal lobe structures subserving semantic memory are known to be affected in SVPPA, and entorhinal cortex, amygdala, middle and inferior temporal, and fusiform gyri were the most severely damaged regions (25). The findings of the present study extend the value of frontal and temporal cortical thickness measures as potential diagnostic markers in PPA. Similarly, an automated MR imaging–based approach with T1 images provided 7
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
an accuracy of 89% in distinguishing NFVPPA and SVPPA variants (16). A recent study used an a priori region-ofinterest–based approach and demonstrated that a combination of left IFG, temporopolar, and superior temporal sulcal thickness measures completely separated the three PPA variants (24). It is noteworthy that, for the discrimination of each PPA variant from healthy control participants, we found that DT MR imaging variables had higher classification accuracies relative to cortical thickness measures. A growing number of studies investigated WM disease in PPA by highlighting distinct disruptions of large-scale networks that contribute to language impairment. Diffusivity abnormalities of the dorsal pathways that we found in NFVPPA patients agree with results of previous DT MR imaging studies (8–13), which are related to syntactic deficits (26). Superior longitudinal fasciculus and body callosal fibers also contain connections between frontal motor and premotor regions, which may contribute to speech fluency (10). Previous reports showed that left ventral temporal WM areas are selectively vulnerable in patients with SVPPA (7–9,11–13), which is consistent with their semantic processing deficits. Pathologic studies of FTLD cases showed that although the regional extent of WM alterations roughly parallels that of cortical degeneration, there are WM abnormalities that extend beyond the expected boundaries based on cortical involvement (27). Primary WM damage in FTLD may involve the accumulation of pathologic proteins in WM regions and microglia activation (28,29). A key pathophysiologic role of WM alterations in FTLD syndromes is also suggested by using MR imaging evidence in asymptomatic mutation carriers (30). An important caveat to consider regarding interpretation of our findings is the lack of pathologic confirmation of diagnosis. However, prediction of clinical diagnosis and prediction of pathologic substrates are two separate issues, and here we only focused on the first aspect. Converging evidence suggests a greater WM burden associated with FTLD-t relative to FTLD–transactive 8
Agosta et al
response DNA binding protein 43 kDa cases (31,32), and DT MR imaging helps to provide optimal sensitivity and specificity for an in vivo FTLD-t and FTLD-transactive response DNA binding protein 43 kDa diagnosis in individual patients (32). Future studies with pathologic confirmation should explore whether our approach is useful to distinguish FTLD subtypes in PPA patients. Another main limitation of our study is the relatively small number of cases included; however, this is similar to previous studies (5,7–13,22–24,33) and reflects the rarity of these conditions. Congruence with a priori hypothesis and consistency with previous findings should mitigate this concern. In addition, disease duration was slightly longer in SVPPA compared with NFVPPA patients, which suggests that the former patients might be at a more advanced stage of the disease than the latter; nonetheless, patients were well matched on general measures of cognitive status and dementia severity, and results were adjusted for disease duration. Furthermore, we did not report quantitative findings of apraxia of speech and agrammatism of the recorded spontaneous speech samples; the quantitative analysis of extended speech production is central for the distinction between NFVPPA and logopenic PPA, because differences can be subtle. However, a qualitative approach is sufficient to allow a clinical distinction between NFVPPA and SVPPA. In this study, we did not include patients with logopenic PPA or mixed PPA. Because NFVPPA and logopenic PPA are difficult to distinguish based on linguistic measures alone and mixed PPA is difficult to classify, the next research step would be to validate our approach in a more challenging scenario. Healthy control participants did not undergo language-specific neurocognitive testing. Finally, the use of a leave-one-out method in a small dataset with a large number of covariates can be associated with a high risk of data overfitting; this was not the case in our study because the most important variables were selected by using the RF analysis, and leave-one-out method was used only as
an internal validation of AUC estimates. Certainly, the accuracy of our models should be tested by independent and larger datasets, and this highlights the urgent need for multicenter neuroimaging studies of PPA. To conclude, development of reliable and objective measures to aid in the classification of PPA subtypes is increasingly important because FTLDspecific modification treatments are likely to become available in the near future. Individuals with NFVPPA and SVPPA have important brain volumetric and diffusivity alterations, which can offer novel markers useful for differential diagnostic purposes. Disclosures of Conflicts of Interest: F.A. Activities related to the present article: author received a grant from the Italian Ministry of Health. Activities not related to the present article: author received speaker’s honoraria from Biogen Idec and Serono Symposia International Foundation; author received a grant from Arisla. Other relationships: disclosed no relevant relationships. P.M.F. disclosed no relevant relationships. E.C. disclosed no relevant relationships. M.C. disclosed no relevant relationships. S.G. disclosed no relevant relationships. G.M. disclosed no relevant relationships. A.M. disclosed no relevant relationships. P.V. disclosed no relevant relationships. A.S. disclosed no relevant relationships. G.C. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: author receives personal fees from Novartis, Teva, Sanofi, Genzyme, Merck Serono, Biogen, Bayer, Serono Symposia, Excemed, Almirall, Chugai, and Receptos. Other relationships: disclosed no relevant relationships. A.F. disclosed no relevant relationships. M.F. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: author receives fees for consulting services and/or speaking activities from Bayer Schering Pharma, Biogen Idec, Merck Serono, Teva Pharmaceutical Industries, Bayer Schering Pharma, and Biogen Idec; author’s institution receives money from grants from Bayer Schering Pharma, Biogen Idec, Merck Serono, and Teva Pharmaceutical Industries. Other relationships: disclosed no relevant relationships.
References 1. Mesulam MM. Slowly progressive aphasia without generalized dementia. Ann Neurol 1982;11(6):592–598. 2. Gorno-Tempini ML, Hillis AE, Weintraub S, et al. Classification of primary progressive aphasia and its variants. Neurology 2011;76(11): 1006–1014.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
NEURORADIOLOGY: Differentiation between Subtypes of Primary Progressive Aphasia
3. Rohrer JD, Lashley T, Schott JM, et al. Clinical and neuroanatomical signatures of tissue pathology in frontotemporal lobar degeneration. Brain 2011;134(Pt 9):2565– 2581. 4. Mesulam MM, Weintraub S. Is it time to revisit the classification guidelines for primary progressive aphasia? Neurology 2014;82(13): 1108–1109. 5. Rohrer JD, Warren JD, Modat M, et al. Patterns of cortical thinning in the language variants of frontotemporal lobar degeneration. Neurology 2009;72(18):1562–1569. 6. Rogalski E, Cobia D, Harrison TM, Wieneke C, Weintraub S, Mesulam MM. Progression of language decline and cortical atrophy in subtypes of primary progressive aphasia. Neurology 2011;76(21):1804–1810. 7. Agosta F, Henry RG, Migliaccio R, et al. Language networks in semantic dementia. Brain 2010;133(Pt 1):286–299. 8. Agosta F, Scola E, Canu E, et al. White matter damage in frontotemporal lobar degeneration spectrum. Cereb Cortex 2012;22(12): 2705–2714. 9. Galantucci S, Tartaglia MC, Wilson SM, et al. White matter damage in primary progressive aphasias: a diffusion tensor tractography study. Brain 2011;134(Pt 10):3011–3029. 10. Grossman M, Powers J, Ash S, et al. Disruption of large-scale neural networks in nonfluent/agrammatic variant primary progressive aphasia associated with frontotemporal degeneration pathology. Brain Lang 2013; 127(2):106–120.
dromic variants of frontotemporal dementia. Neurology 2010;74(16):1279–1287. 14. Desikan RS, Ségonne F, Fischl B, et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 2006;31(3):968–980. 15. Breiman L. Random forests. Mach Learn 2001;45(1):5–32. 16. Wilson SM, Ogar JM, Laluz V, et al. Automated MRI-based classification of primary progressive aphasia variants. Neuroimage 2009;47(4):1558–1567. 17. Avants BB, Cook PA, Ungar L, Gee JC, Grossman M. Dementia induces correlated reductions in white matter integrity and cortical thickness: a multivariate neuroimaging study with sparse canonical correlation analysis. Neuroimage 2010;50(3):1004– 1016.
24. Sapolsky D, Bakkour A, Negreira A, et al. Cortical neuroanatomic correlates of symptom severity in primary progressive aphasia. Neurology 2010;75(4):358–366. 25. Chan D, Fox NC, Scahill RI, et al. Pat terns of temporal lobe atrophy in semantic dementia and Alzheimer’s disease. Ann Neurol 2001;49(4):433–442. 26. Wilson SM, Galantucci S, Tartaglia MC, et al. Syntactic processing depends on dorsal language tracts. Neuron 2011;72(2):397–403. 27. Englund E, Brun A. Frontal lobe degeneration of non-Alzheimer type. IV. White matter changes. Arch Gerontol Geriatr 1987;6(3): 235–243. 28. Neumann M, Kwong LK, Truax AC, et al. TDP-43-positive white matter pathology in frontotemporal lobar degeneration with ubiquitin-positive inclusions. J Neuropathol Exp Neurol 2007;66(3):177–183.
18. Zhang Y, Schuff N, Ching C, et al. Joint assessment of structural, perfusion, and diffusion MRI in Alzheimer’s disease and frontotemporal dementia. Int J Alzheimers Dis 2011;2011:546871.
29. Schofield E, Kersaitis C, Shepherd CE, Kril JJ, Halliday GM. Severity of gliosis in Pick’s disease and frontotemporal lobar degeneration: tau-positive glia differentiate these disorders. Brain 2003;126(Pt 4):827–840.
19. McMillan CT, Brun C, Siddiqui S, et al. White matter imaging contributes to the multimodal diagnosis of frontotemporal lobar degeneration. Neurology 2012;78(22):1761–1768.
30. Dopper EG, Rombouts SA, Jiskoot LC, et al. Structural and functional brain connectivity in presymptomatic familial frontotemporal dementia. Neurology 2013;80(9):814–823.
20. McMillan CT, Avants BB, Cook P, Ungar L, Trojanowski JQ, Grossman M. The power of neuroimaging biomarkers for screening frontotemporal dementia. Hum Brain Mapp 2014;35(9):4827–4840.
11. Mahoney CJ, Malone IB, Ridgway GR, et al. White matter tract signatures of the progressive aphasias. Neurobiol Aging 2013;34(6): 1687–1699.
21. Gorno-Tempini ML, Dronkers NF, Rankin KP, et al. Cognition and anatomy in three variants of primary progressive aphasia. Ann Neurol 2004;55(3):335–346.
12. Sajjadi SA, Acosta-Cabronero J, Patterson K, Diaz-de-Grenu LZ, Williams GB, Nestor PJ. Diffusion tensor magnetic resonance imaging for single subject diagnosis in neurodegenerative diseases. Brain 2013;136(Pt 7): 2253–2261.
22. Gunawardena D, Ash S, McMillan C, Avants B, Gee J, Grossman M. Why are patients with progressive nonfluent aphasia nonfluent? Neurology 2010;75(7):588–594.
13. Whitwell JL, Avula R, Senjem ML, et al. Gray and white matter water diffusion in the syn-
Agosta et al
31. Caso F, Mandelli ML, Henry M, et al. In vivo signatures of nonfluent/agrammatic primary progressive aphasia caused by FTLD pathology. Neurology 2014;82(3):239–247. 32. McMillan CT, Irwin DJ, Avants BB, et al. White matter imaging helps dissociate tau from TDP-43 in frontotemporal lobar degeneration. J Neurol Neurosurg Psychiatry 2013;84(9):949–955. 33. Shim H, Hurley RS, Rogalski E, Mesulam MM. Anatomic, clinical, and neuropsychological correlates of spelling errors in primary progressive aphasia. Neuropsychologia 2012;50(8):1929–1935.
23. Gefen T, Wieneke C, Martersteck A, et al. Naming vs knowing faces in primary progressive aphasia: a tale of 2 hemispheres. Neurology 2013;81(7):658–664.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
9
