MRI Texture Analysis Reveals Deep Gray Nuclei Damage in Amyotrophic Lateral Sclerosis ´ Milena de Albuquerque, Lara G.V. Anjos, Helen Maia Tavares de Andrade, Marcia S. de Oliveira, Gabriela Castellano, Thiago Junqueira Ribeiro de Rezende, Anamarli Nucci, Marcondes Cavalcante Franc¸a Junior From the Departments of Neurology and Neuroimaging Laboratory, School of Medical Sciences, University of Campinas – UNICAMP, Brazil (MdA, LGVA, HMTdA, TJRdR, AN, MCFJ); Neurophysics Group, Institute of Physics Gleb Wataghin, University of Campinas -UNICAMP, Brazil (MSdO, GC, TJRdR); and Brazilian Institute of Neuroscience and Neurotechnology (BRAINN) (MSdO, GC).
ABSTRACT BACKGROUND AND PURPOSE: Amyotrophic Lateral Sclerosis (ALS) is characterized by extensive corticospinal damage, but extrapyramidal involvement is suggested in pathological studies. Texture analysis (TA) is an image processing technique that evaluates the distribution of gray levels between pixels in a given region of interest (ROI). It provides quantitative data and has been employed in several neurodegenerative disorders. Here, we used TA to investigate possible deep gray nuclei (DGN) abnormalities in a cohort of ALS patients. METHODS: Thirty-two ALS patients and 32 healthy controls underwent MRI in a 3T scanner. The T1 volumetric sequence was used for DGN segmentation and extraction of 11 texture parameters using the MaZda software. Statistical analyses were performed using the Mann-Whitney non-parametric test, with a significance level set at α = 0.025 (FDR-corrected) for TA. RESULTS: Patients had significantly higher values for the parameter correlation (CO) in both thalami and in the right caudate nucleus compared to healthy controls. Also, the parameter Inverse Difference Moment or Homogeneity (IDM) presented significantly smaller values in the ALS group in both thalami. CONCLUSIONS: TA of T1 weighted images revealed DGN alterations in patients with ALS, namely in the thalami and caudate nuclei. Keywords: MRI, texture analysis, ALS, basal ganglia, ROI. Acceptance: Received January 25, 2015, and in revised form April 14, 2015. Accepted for publication April 15, 2015. Correspondence: Address correspondence to Marcondes C. Franc¸a Junior, MD, PhD, Department of Neurology, University of Campinas – UNICAMP. Rua Tessalia ´ Vieira de Camargo, 126. Cidade Universitaria “Zeferino Vaz”, Campinas, SP, Brazil- 13083–887. E-mail: [email protected]. Funding: Supported by FAPESP (Grants 2011/21521-3, 2012/24363-2, 2013/07559-3); CNPq; BRAINN (2013/07559-3). Acknowledgments: None. J Neuroimaging 2016;26:201-206. DOI: 10.1111/jon.12262
Introduction Amyotrophic Lateral Sclerosis (ALS) is the most frequent motor neuron disease and invariably fatal.1 Men have a higher incidence and peak age at onset varies from 50-70 years for sporadic disease.1,2 The disease is characterized classically by damage to the upper and lower motor neurons at the cerebral cortex, brainstem, and spinal cord.1 Besides, microstructural cortical changes in patients with ALS were found both at motor and extramotor areas in a recent study using Magnetization Transfer Imaging.4 Lately, manifestations unrelated to motor neurons have been increasingly recognized, such as executive dysfunction, apathy, impaired social cognition, and extrapyramidal motor symptoms.3 The pathological findings in the motor cortex are well established so far. A recent study that combined pathology and MRI analyses revealed neuronal loss and reduced axonal density in the primary motor cortex of ALS patients.5 These abnormalities correlated with T1 hypointensity in the cortex seen on MRI scans. More recently, besides corticospinal impairment, postmortem studies confirmed the presence of extensive subcortical gray matter involvement at globus pallidus, thalamus, amygdala, red nucleus, striatum, hippocampus, and substantia nigra.6,7 In line with these findings, a recent MRI-based study
Copyright
found widespread basal ganglia involvement in ALS patients using advanced volumetric techniques.3 It is indeed important to ascertain that deep gray matter structures are really damaged in ALS because this might help to explain the neuropsychiatric manifestations of the disease. In this scenario, we propose the use of texture analysis (TA), applied to T1-weighted MR images of the brain, to analyze the differences in deep gray nuclei (DGN) texture parameters between ALS patients and control subjects matched for age and gender. TA has been successfully employed in several neurodegenerative diseases to characterize structural damage, such as Alzheimer’s disease,8 juvenile myoclonic epilepsy,9 focal cortical dysplasia,10 and Machado-Joseph disease,11 but not yet fully explored in ALS patients. In fact, to the best of our knowledge, up to now there is only one study using TA in ALS (https://www.cs.ualberta.ca/research/thesespublications/technical-reports/2014/TR14-01); however, only an abstract has been published, with no accessible information on which structures were evaluated or how exactly were the TA parameters computed. Therefore, there is still need for exploring TA applied to ALS. Texture in images may be understood as a set of intrinsic properties such as brightness, color, size, and so on.12 The
◦ 2015 by the American Society of Neuroimaging C
201
textural characteristics can be analyzed in several aspects, such as undulations, softness, roughness, regularity, or linearity.8,9,12 The concept of texture is related to the distribution of gray levels between pixels in a given region of interest (ROI).8 Our study used a statistical approach, which extracts texture information from the image on the basis of the gray level distribution of pairs of pixels, as demonstrated previously.8
Methods Subjects’ selection We enrolled 32 consecutive patients with possible, probable, or definite ALS according to the El Escorial13 revised criteria (13 women and mean age of 55.3 ± 10.8 years) and 32 healthy controls (12 women and mean age of 53.5 ± 15.4 years). Patients were recruited from the Neurology Outpatient Clinic at UNICAMP hospital between 2012 and 2013. Patients were excluded if they had other associated neurological disorders. This study was approved by the Ethics Committee of our institution (University of Campinas – UNICAMP, Brazil), and all subjects signed a written informed consent. Patients were also evaluated with the clinical scales Amyotrophic lateral sclerosis functional rating scale (ALSFRSr - revised version)14 and Amyotrophic lateral sclerosis severity scale (ALSSS)15 to quantify disease severity in the same day of MRI acquisitions. The ALSFRSr instrument is a questionnaire that measures physical function in activities of daily living (ADL).14 Scores range between 0 and 48 points, and lower scores indicate more severe disability. ALSSS assesses the upper and lower extremity motor function during activities of daily life, as well as speech and swallowing ability. It does not evaluate breathing-related issues.15 ALSSS scores range between 4 and 40, and lower scores also indicate more severe disease.
MRI acquisition protocol All patients and controls were scanned on a 3 T Philips MRI scanner. Volumetric T1-weighted images (sagittal orientation, voxel matrix 240×240×180, voxel size 1×1×1 mm3 no gap, TR/TE 7/ 3.201 ms, flip angle 8º) of the whole brain were obtained using a standard eight channel head coil and used for TA. We also obtained axial T2-weighted and FLAIR images of each subject to exclude unrelated abnormalities.
Segmentation and Texture Analysis For TA, the structures of interest (bilateral thalami, caudate nuclei, putamen, and cerebral peduncles) were manually and blindly segmented in T1 images by a single experienced evaluator (LGVA) using the MaZda software12 (Fig 1). All of these structures were segmented in the left and right hemispheres. Next, TA was performed for each segmented ROI also using the MaZda suite.12,16 The statistical approach adopted to extract texture parameters from the MRI images was based on the Grey Level Co-ocurrence (GLC) matrix, as explained in a previous study.11 Briefly, the GLC matrix is an N×N order matrix, with N being the number of grey levels present in the image. Every element (i,j) of the GLC matrix accounts for the number of times that grey level i co-occurs with gray level j, for a given distance and direction between pixels (generally, distances of 1 to 5 pixels, and directions of 0o , 45o , 90o , and 135o are used).17 The 11 texture parameters used in this work, calculated from the GLC matrix, were: Uniformity (also known 202
Fig 1. Axial T1-weighted MR images showing segmentation of the regions of interest (ROI): green – right putamen, pink – left putamen, brown – right caudate nucleus, yellow – left caudate nucleus, orange – right thalamus, light blue – left thalamus, blue – right cerebral peduncle, light green – left cerebral peduncle.
as Angular Second Moment), Contrast, Correlation, Variance (also known as Sum of Squares), Homogeneity or Inverse Difference Moment, Sum Average, Sum Variance, Sum Entropy, Entropy, Difference Variance, and Difference Entropy, as described in a previous study.9 The MaZda Program calculates the GLC matrices for each ROI and the corresponding texture parameters for each matrix. Given that the ROI size is not necessarily the same in all slices selected, a weighted average of all texture parameters was calculated using a Matlab routine, where the ROI size was used as weight. At the end of the process, four groups of parameters (one for each direction) for every subject were obtained. The parameters were then averaged over directions, following the hypotheses that the analyzed tissues were isotropic, or at least, that the head position in the images should not bias the texture measurements. In the end, there was one parameter set for every distance (and structure). It is important to mention that, given the small size of most of the analyzed structures, it was not possible to compute GLC matrices for all the aforementioned pixel distances. The pixel distances used for each structure were: peduncule d = 1, 2; caudate d = 1, 2, 3; putamen d = 1, 2, 3, 4; thalamus d = 1, 2, 3, 4, 5. The corresponding sets of parameters were used in the statistical analysis. We further evaluated whether TA abnormalities were related to volumetric changes in DGN. To accomplish that, we employed the FreeSurfer software v5.3 to perform automatic volumetry of the DGN. In brief, MRI images underwent correction for magnetic field inhomogeneity; alignment to a specific atlas; skull removal; and the segmentation of the voxels into gray matter (GM), white matter (WM), and cerebro-spinal fluid (CSF).18,19 Estimated Intracranial Volume (eTIV) and the volume of subcortical structures were then recorded. FreeSurfer enables the comparison of subcortical volume measurements according to anatomical atlases such as proposed by Desikan.20
Statistical methods All statistical analyses were performed with the Systat 12 software (SigmaPlot, San Jose, CA). Texture Analysis
The statistical analysis was performed separately for each distance in each ROI. Correlations of texture parameters with
Journal of Neuroimaging Vol 26 No 2 March/April 2016
Fig 2. Box Plot shows the distribution of texture parameters that presented significant difference between patients (group 1) and controls (group 2): A) Right Caudate Nucleus; .i) Correlation for distance d = 2; .ii) Correlation for distance d = 3. B) Left thalamus: .i) Correlation for distance d = 1; .ii) Inverse Difference Moment for distance d = 4; .iii) Inverse Difference Moment for distance d = 5. C) Right thalamus: .i) Correlation for distance d = 4; .ii) Correlation d = 5; .iii) Inverse Difference Moment for distance d = 5.
clinical variables were also performed using the Spearman coefficient (only those parameters that presented significant differences between patients and controls were included). The groups were compared using the Mann-Whitney U test. All results were corrected for multiple comparisons using the FDR (False Discovery Rate) correction (level of significance α = 0.025). To further evaluate how different and able to separate both groups TA parameters were, we performed ROC curve analyses. It was performed with the parameters and distances that presented significant differences between patients and controls. The area under the curve (AUC) measures how well a parameter can distinguish between two diagnostic groups (AUC close to 1 indicate a very informative test). FreeSurfer
The statistical analysis was done using a general linear model (GLM) with age, gender and eTIV as covariates to assess subcortical volume differences between patients and controls for each region. All results were corrected for multiple comparisons
using the FDR correction (level of significance α = 0.029). Next, another GLM with age, gender, and eTIV as covariates was performed to assess possible correlations between TA parameters and DGN volumes. In this last analysis, we only evaluated those DGN that presented any TA abnormality in comparison to controls.
Results Mean disease duration (months), ALSFRSr, and ALSS scores were 36.0 ± 41.2, 30.8 ± 8.8, and 27.4 ± 7.0, respectively. Seven patients had bulbar onset and 25 limb onset. Fifteen were wheelchair bound, but none of them required continuous positive pressure ventilation. Values of texture parameters of the cerebral peduncles and putamen were similar in both groups. We only found between-group differences in TA regarding the right caudate nucleus and both thalami (Figs 2 and 3). The areas under the curve obtained from the ROC curves were calculated and obtained values were near 0.7 (Fig 2) in all analyses. The Correlation parameter in the right caudate, d = 3, presented the highest AUC ( = 0.7134). For this parameter, de Albuquerque et al: Texture Analysis in ALS
203
Fig 3. Plots showing the area under the ROC curves for the texture parameters that presented significant difference between patients and controls: A) Right Caudate Nucleus: .i) Correlation for distance d = 2; .ii) Correlation for distance d = 3. B) Left thalamus: .i) Correlation for distance d = 1; .ii) Inverse Difference Moment for distance d = 4; .iii) Inverse Difference Moment for distance d = 5. C) Right thalamus: .i) Correlation for distance d = 4; .ii) Correlation d = 5; .iii) Inverse Difference Moment for distance d = 5.
204
Journal of Neuroimaging Vol 26 No 2 March/April 2016
Table 1. Texture Parameters that Presented Significant Difference Between Patients with ALS and Controls Region
Right Caudate Left thalamus
Right thalamus
Distance (d)
Texture Parameter
2 3 1 4 5 4 5 5
CO CO CO IDM IDM CO CO IDM
Mean ± SD (Controls)
0.0895 −0.0152 0.5405 0.2072 0.1908 0.0008 −0.0984 0.2081
± ± ± ± ± ± ± ±
0.0727 0.0775 0.0782 0.0303 0.0359 0.060 0.0783 0.0378
Mean ± SD (ALS)
0.1527 0.0434 0.5877 0.1886 0.1659 −0.0310 −0.1411 0.1870
± ± ± ± ± ± ± ±
0.0858 0.0881 0.0739 0.0217 0.0228 0.0594 0.0591 0.0253
P-value
0.0075 0.0032 0.0111 0.0107 0.0064 0.0168 0.0111 0.0150
CO = Correlation; IDM = Inverse Different Moment
Table 2. Volumes of DGN in Patients with ALS and Matched Controls Region
Left Caudate Right Caudate Left Thalamus Right Thalamus Left Putamen Right Putamen
Mean Controls +/− SD (mm3 )
Mean Patients +/− SD (mm3 )
P value
3349.71 +/− 407.79 3564.04 +/− 470.38 7334.65 +/− 906.88 6598.25 +/− 868.10 5487.45 +/− 913.01 5519.78 +/− 742.81
3464.10 +/− 524.47 3583.51 +/− 541.37 7010.00 +/− 984.83 6112.01 +/− 762.46 5415.55 +/− 829.66 5309.93 +/− 803.95
0.099 0.423 0.347 0.015* 0.661 0.559
*significant difference (FDR-corrected p
ABSTRACT BACKGROUND AND PURPOSE: Amyotrophic Lateral Sclerosis (ALS) is characterized by extensive corticospinal damage, but extrapyramidal involvement is suggested in pathological studies. Texture analysis (TA) is an image processing technique that evaluates the distribution of gray levels between pixels in a given region of interest (ROI). It provides quantitative data and has been employed in several neurodegenerative disorders. Here, we used TA to investigate possible deep gray nuclei (DGN) abnormalities in a cohort of ALS patients. METHODS: Thirty-two ALS patients and 32 healthy controls underwent MRI in a 3T scanner. The T1 volumetric sequence was used for DGN segmentation and extraction of 11 texture parameters using the MaZda software. Statistical analyses were performed using the Mann-Whitney non-parametric test, with a significance level set at α = 0.025 (FDR-corrected) for TA. RESULTS: Patients had significantly higher values for the parameter correlation (CO) in both thalami and in the right caudate nucleus compared to healthy controls. Also, the parameter Inverse Difference Moment or Homogeneity (IDM) presented significantly smaller values in the ALS group in both thalami. CONCLUSIONS: TA of T1 weighted images revealed DGN alterations in patients with ALS, namely in the thalami and caudate nuclei. Keywords: MRI, texture analysis, ALS, basal ganglia, ROI. Acceptance: Received January 25, 2015, and in revised form April 14, 2015. Accepted for publication April 15, 2015. Correspondence: Address correspondence to Marcondes C. Franc¸a Junior, MD, PhD, Department of Neurology, University of Campinas – UNICAMP. Rua Tessalia ´ Vieira de Camargo, 126. Cidade Universitaria “Zeferino Vaz”, Campinas, SP, Brazil- 13083–887. E-mail: [email protected]. Funding: Supported by FAPESP (Grants 2011/21521-3, 2012/24363-2, 2013/07559-3); CNPq; BRAINN (2013/07559-3). Acknowledgments: None. J Neuroimaging 2016;26:201-206. DOI: 10.1111/jon.12262
Introduction Amyotrophic Lateral Sclerosis (ALS) is the most frequent motor neuron disease and invariably fatal.1 Men have a higher incidence and peak age at onset varies from 50-70 years for sporadic disease.1,2 The disease is characterized classically by damage to the upper and lower motor neurons at the cerebral cortex, brainstem, and spinal cord.1 Besides, microstructural cortical changes in patients with ALS were found both at motor and extramotor areas in a recent study using Magnetization Transfer Imaging.4 Lately, manifestations unrelated to motor neurons have been increasingly recognized, such as executive dysfunction, apathy, impaired social cognition, and extrapyramidal motor symptoms.3 The pathological findings in the motor cortex are well established so far. A recent study that combined pathology and MRI analyses revealed neuronal loss and reduced axonal density in the primary motor cortex of ALS patients.5 These abnormalities correlated with T1 hypointensity in the cortex seen on MRI scans. More recently, besides corticospinal impairment, postmortem studies confirmed the presence of extensive subcortical gray matter involvement at globus pallidus, thalamus, amygdala, red nucleus, striatum, hippocampus, and substantia nigra.6,7 In line with these findings, a recent MRI-based study
Copyright
found widespread basal ganglia involvement in ALS patients using advanced volumetric techniques.3 It is indeed important to ascertain that deep gray matter structures are really damaged in ALS because this might help to explain the neuropsychiatric manifestations of the disease. In this scenario, we propose the use of texture analysis (TA), applied to T1-weighted MR images of the brain, to analyze the differences in deep gray nuclei (DGN) texture parameters between ALS patients and control subjects matched for age and gender. TA has been successfully employed in several neurodegenerative diseases to characterize structural damage, such as Alzheimer’s disease,8 juvenile myoclonic epilepsy,9 focal cortical dysplasia,10 and Machado-Joseph disease,11 but not yet fully explored in ALS patients. In fact, to the best of our knowledge, up to now there is only one study using TA in ALS (https://www.cs.ualberta.ca/research/thesespublications/technical-reports/2014/TR14-01); however, only an abstract has been published, with no accessible information on which structures were evaluated or how exactly were the TA parameters computed. Therefore, there is still need for exploring TA applied to ALS. Texture in images may be understood as a set of intrinsic properties such as brightness, color, size, and so on.12 The
◦ 2015 by the American Society of Neuroimaging C
201
textural characteristics can be analyzed in several aspects, such as undulations, softness, roughness, regularity, or linearity.8,9,12 The concept of texture is related to the distribution of gray levels between pixels in a given region of interest (ROI).8 Our study used a statistical approach, which extracts texture information from the image on the basis of the gray level distribution of pairs of pixels, as demonstrated previously.8
Methods Subjects’ selection We enrolled 32 consecutive patients with possible, probable, or definite ALS according to the El Escorial13 revised criteria (13 women and mean age of 55.3 ± 10.8 years) and 32 healthy controls (12 women and mean age of 53.5 ± 15.4 years). Patients were recruited from the Neurology Outpatient Clinic at UNICAMP hospital between 2012 and 2013. Patients were excluded if they had other associated neurological disorders. This study was approved by the Ethics Committee of our institution (University of Campinas – UNICAMP, Brazil), and all subjects signed a written informed consent. Patients were also evaluated with the clinical scales Amyotrophic lateral sclerosis functional rating scale (ALSFRSr - revised version)14 and Amyotrophic lateral sclerosis severity scale (ALSSS)15 to quantify disease severity in the same day of MRI acquisitions. The ALSFRSr instrument is a questionnaire that measures physical function in activities of daily living (ADL).14 Scores range between 0 and 48 points, and lower scores indicate more severe disability. ALSSS assesses the upper and lower extremity motor function during activities of daily life, as well as speech and swallowing ability. It does not evaluate breathing-related issues.15 ALSSS scores range between 4 and 40, and lower scores also indicate more severe disease.
MRI acquisition protocol All patients and controls were scanned on a 3 T Philips MRI scanner. Volumetric T1-weighted images (sagittal orientation, voxel matrix 240×240×180, voxel size 1×1×1 mm3 no gap, TR/TE 7/ 3.201 ms, flip angle 8º) of the whole brain were obtained using a standard eight channel head coil and used for TA. We also obtained axial T2-weighted and FLAIR images of each subject to exclude unrelated abnormalities.
Segmentation and Texture Analysis For TA, the structures of interest (bilateral thalami, caudate nuclei, putamen, and cerebral peduncles) were manually and blindly segmented in T1 images by a single experienced evaluator (LGVA) using the MaZda software12 (Fig 1). All of these structures were segmented in the left and right hemispheres. Next, TA was performed for each segmented ROI also using the MaZda suite.12,16 The statistical approach adopted to extract texture parameters from the MRI images was based on the Grey Level Co-ocurrence (GLC) matrix, as explained in a previous study.11 Briefly, the GLC matrix is an N×N order matrix, with N being the number of grey levels present in the image. Every element (i,j) of the GLC matrix accounts for the number of times that grey level i co-occurs with gray level j, for a given distance and direction between pixels (generally, distances of 1 to 5 pixels, and directions of 0o , 45o , 90o , and 135o are used).17 The 11 texture parameters used in this work, calculated from the GLC matrix, were: Uniformity (also known 202
Fig 1. Axial T1-weighted MR images showing segmentation of the regions of interest (ROI): green – right putamen, pink – left putamen, brown – right caudate nucleus, yellow – left caudate nucleus, orange – right thalamus, light blue – left thalamus, blue – right cerebral peduncle, light green – left cerebral peduncle.
as Angular Second Moment), Contrast, Correlation, Variance (also known as Sum of Squares), Homogeneity or Inverse Difference Moment, Sum Average, Sum Variance, Sum Entropy, Entropy, Difference Variance, and Difference Entropy, as described in a previous study.9 The MaZda Program calculates the GLC matrices for each ROI and the corresponding texture parameters for each matrix. Given that the ROI size is not necessarily the same in all slices selected, a weighted average of all texture parameters was calculated using a Matlab routine, where the ROI size was used as weight. At the end of the process, four groups of parameters (one for each direction) for every subject were obtained. The parameters were then averaged over directions, following the hypotheses that the analyzed tissues were isotropic, or at least, that the head position in the images should not bias the texture measurements. In the end, there was one parameter set for every distance (and structure). It is important to mention that, given the small size of most of the analyzed structures, it was not possible to compute GLC matrices for all the aforementioned pixel distances. The pixel distances used for each structure were: peduncule d = 1, 2; caudate d = 1, 2, 3; putamen d = 1, 2, 3, 4; thalamus d = 1, 2, 3, 4, 5. The corresponding sets of parameters were used in the statistical analysis. We further evaluated whether TA abnormalities were related to volumetric changes in DGN. To accomplish that, we employed the FreeSurfer software v5.3 to perform automatic volumetry of the DGN. In brief, MRI images underwent correction for magnetic field inhomogeneity; alignment to a specific atlas; skull removal; and the segmentation of the voxels into gray matter (GM), white matter (WM), and cerebro-spinal fluid (CSF).18,19 Estimated Intracranial Volume (eTIV) and the volume of subcortical structures were then recorded. FreeSurfer enables the comparison of subcortical volume measurements according to anatomical atlases such as proposed by Desikan.20
Statistical methods All statistical analyses were performed with the Systat 12 software (SigmaPlot, San Jose, CA). Texture Analysis
The statistical analysis was performed separately for each distance in each ROI. Correlations of texture parameters with
Journal of Neuroimaging Vol 26 No 2 March/April 2016
Fig 2. Box Plot shows the distribution of texture parameters that presented significant difference between patients (group 1) and controls (group 2): A) Right Caudate Nucleus; .i) Correlation for distance d = 2; .ii) Correlation for distance d = 3. B) Left thalamus: .i) Correlation for distance d = 1; .ii) Inverse Difference Moment for distance d = 4; .iii) Inverse Difference Moment for distance d = 5. C) Right thalamus: .i) Correlation for distance d = 4; .ii) Correlation d = 5; .iii) Inverse Difference Moment for distance d = 5.
clinical variables were also performed using the Spearman coefficient (only those parameters that presented significant differences between patients and controls were included). The groups were compared using the Mann-Whitney U test. All results were corrected for multiple comparisons using the FDR (False Discovery Rate) correction (level of significance α = 0.025). To further evaluate how different and able to separate both groups TA parameters were, we performed ROC curve analyses. It was performed with the parameters and distances that presented significant differences between patients and controls. The area under the curve (AUC) measures how well a parameter can distinguish between two diagnostic groups (AUC close to 1 indicate a very informative test). FreeSurfer
The statistical analysis was done using a general linear model (GLM) with age, gender and eTIV as covariates to assess subcortical volume differences between patients and controls for each region. All results were corrected for multiple comparisons
using the FDR correction (level of significance α = 0.029). Next, another GLM with age, gender, and eTIV as covariates was performed to assess possible correlations between TA parameters and DGN volumes. In this last analysis, we only evaluated those DGN that presented any TA abnormality in comparison to controls.
Results Mean disease duration (months), ALSFRSr, and ALSS scores were 36.0 ± 41.2, 30.8 ± 8.8, and 27.4 ± 7.0, respectively. Seven patients had bulbar onset and 25 limb onset. Fifteen were wheelchair bound, but none of them required continuous positive pressure ventilation. Values of texture parameters of the cerebral peduncles and putamen were similar in both groups. We only found between-group differences in TA regarding the right caudate nucleus and both thalami (Figs 2 and 3). The areas under the curve obtained from the ROC curves were calculated and obtained values were near 0.7 (Fig 2) in all analyses. The Correlation parameter in the right caudate, d = 3, presented the highest AUC ( = 0.7134). For this parameter, de Albuquerque et al: Texture Analysis in ALS
203
Fig 3. Plots showing the area under the ROC curves for the texture parameters that presented significant difference between patients and controls: A) Right Caudate Nucleus: .i) Correlation for distance d = 2; .ii) Correlation for distance d = 3. B) Left thalamus: .i) Correlation for distance d = 1; .ii) Inverse Difference Moment for distance d = 4; .iii) Inverse Difference Moment for distance d = 5. C) Right thalamus: .i) Correlation for distance d = 4; .ii) Correlation d = 5; .iii) Inverse Difference Moment for distance d = 5.
204
Journal of Neuroimaging Vol 26 No 2 March/April 2016
Table 1. Texture Parameters that Presented Significant Difference Between Patients with ALS and Controls Region
Right Caudate Left thalamus
Right thalamus
Distance (d)
Texture Parameter
2 3 1 4 5 4 5 5
CO CO CO IDM IDM CO CO IDM
Mean ± SD (Controls)
0.0895 −0.0152 0.5405 0.2072 0.1908 0.0008 −0.0984 0.2081
± ± ± ± ± ± ± ±
0.0727 0.0775 0.0782 0.0303 0.0359 0.060 0.0783 0.0378
Mean ± SD (ALS)
0.1527 0.0434 0.5877 0.1886 0.1659 −0.0310 −0.1411 0.1870
± ± ± ± ± ± ± ±
0.0858 0.0881 0.0739 0.0217 0.0228 0.0594 0.0591 0.0253
P-value
0.0075 0.0032 0.0111 0.0107 0.0064 0.0168 0.0111 0.0150
CO = Correlation; IDM = Inverse Different Moment
Table 2. Volumes of DGN in Patients with ALS and Matched Controls Region
Left Caudate Right Caudate Left Thalamus Right Thalamus Left Putamen Right Putamen
Mean Controls +/− SD (mm3 )
Mean Patients +/− SD (mm3 )
P value
3349.71 +/− 407.79 3564.04 +/− 470.38 7334.65 +/− 906.88 6598.25 +/− 868.10 5487.45 +/− 913.01 5519.78 +/− 742.81
3464.10 +/− 524.47 3583.51 +/− 541.37 7010.00 +/− 984.83 6112.01 +/− 762.46 5415.55 +/− 829.66 5309.93 +/− 803.95
0.099 0.423 0.347 0.015* 0.661 0.559
*significant difference (FDR-corrected p
