The corticospinal tract profile in amyotrophic lateral sclerosis

Subjects and Clinical Parameters

ALS patients were enrolled from the Neurology unit of the University “Magna Graecia” of Catanzaro. An expert neurologist (P.V.), blind to the aim of this study, with 20 years of experience in motor neuron disorders performed a complete clinical examination. All patients with ALS met the revised El Escorial criteria for probable, probable laboratory supported or definite ALS [Brooks et al., 2000]. All ALS patients were right-handed and riluzole treatment-naïve. Motor skills were assessed according to the ALS functional rating scaled revised (ALSFRS-R) [Cedarbaum et al., 1999]. Moreover, to define and quantify the presence of bulbar involvement, we used the score on the three bulbar items of the ALSFRS-R (speech, sialorrhoea, and swallowing, a score of four per item representing normal function). We considered as exclusion criteria: the presence of: (a) multifocal motor neuropathy and paraneoplastic neuropathy, according to nerve conduction examinations; (b) forms of dementia, according to the structured clinical interview of the DSM-IV [Diagnostic and Statistical Manual of Mental Disorders, 4th Edition, American Psychiatric Association, 1994]; (c) history of cerebrovascular disease, head trauma, hypertension or diabetes. Only two patients with a co-morbid diagnosis of frontotemporal dementia according to the Neary Criteria [Neary et al., 1998] were excluded because of the confounding effects of imaging changes associated with this phenotype. Finally, 24 ALS patients were enrolled and underwent neuroimaging exam. All clinical details are reported in Table 1.

Twenty-four age- and sex-matched, right-handed, non-demented, healthy controls [mean age ± SD 61 ± 8.3 (t-test P-level= 0.79); 62.5% female, (χ², P-level= 0.558)] with no previous history of neurological or psychiatric diseases and with normal MRI of the brain were matched for demographical variables with patients. All participants gave written informed consent, which was approved by the Ethical Committee of the University “Magna Graecia” of Catanzaro, according to the Helsinki Declaration.

Data Acquisition

Brain MRI was performed according to our routine protocol [Sarica et al., 2014] by a 3 T scanner with an 8-channel head coil (Discovery MR-750, General Electric, Milwaukee, WI). The protocol included: (a) whole-brain T1-weighted scan MRI scanning (SPGR; TE/TR = 3.7/9.2 ms, flip angle 12°, voxel- size 1 × 1 × 1 mm³); (b) diffusion-weighted volumes, acquired by using spin-echo echo-planar imaging (TE/TR = 87/10,000 ms, bandwidth 250 kHz, matrix size 128 × 128, 80 axial slices, voxel size 2.0 × 2.0 × 2.0 mm³) with 27 equally distributed orientations for the diffusion-sensitizing gradients at a b-value of 1,000 s/mm². Particular care was taken to restrain the subject's movements with cushions and adhesive medical tape. Moreover, before further processing, we visually check the images to exclude all scans with artifacts, largely caused by subject motion.

Data Pre-Processing

Diffusion images were first converted from DICOM to NifTI format by using the script dcm2nii from the MRIcron1 tool. The FMRIB's v5.0 (Oxford Centre for Functional MRI of the Brain2) Diffusion Toolbox (FDT) was used for correcting eddy current and motion distortion. The non-diffusion-weighted images, that is, the first volume image used as reference in DTI data, were skull stripped using the FMRIB's brain extraction tool (BET)3 and used to mask all diffusion-weighted images [Smith, 2002]. A diffusion-tensor model was fitted at each voxel using Diffusion Toolkit, generating besides fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD) and radial diffusivity (RD) maps, a raw T2 image with no diffusion weighting (S0). The script dtiMakeDt6FromFSL from the Mr Diffusion toolkit4 was used for aligning the T1 image—the anatomical reference—to the S0 image, obtaining a dt6 MATLAB format file for further analysis, described in the next section.

Automated Fiber Quantification

The Automated Fiber Quantification (AFQ, version 1.2 on MATLAB R2014b) tool [Yeatman et al., 2012] was used for reconstructing the CST and for evaluating diffusion metrics along the tract length. AFQ reconstructs the main white matter tracts, for both hemispheres, and measures FA, MD, AD, and RD along their trajectories. AFQ has been validated against manual tracings in healthy controls [Yeatman et al., 2012] and applied in several clinical realms, such as children born preterm [Yeatman et al., 2012], elderly individuals [Yeatman et al., 2014], autism disorders [Libero et al., 2015] and major depression [Sacchet et al., 2014].

In particular, AFQ uses a three-step procedure to identify 24 major fiber tracts in an individual's brain. The procedure is based on a combination of the methods described by Hua et al. [2008] and Zhang et al. [2008]: (1) whole-brain fiber tractography, (2) waypoint ROI-based fiber tract segmentation, and (3) cleaning and refinement of fiber tracts based on a probabilistic fiber tract atlas. An example of CST reconstruction has been showed in Figure 1. For the first step, the whole-brain fiber tractography, AFQ uses a deterministic streamline-tracing algorithm (STT) [Mori et al., 1999] with a fourth-order Runge–Kutta path integration method with 1-mm step size. Voxels with FA greater than 0.3 are selected for generating a white matter mask used as a seed. Streamlines are traced bidirectionally along the principal diffusion axis, until the stopping criteria are reached (FA < 0.2 or angle between steps > 30°) [Yeatman et al., 2012]. The output of this first step is a whole-brain tractography that is used in the next step for segmenting the fiber tracts, using a waypoint ROI-based method [Wakana et al., 2007]. The pairs of waypoint ROIs define the fibers that pass through stereotypical locations and they are drawn on a group-average DTI dataset in MNI space, which are then non-linearly warped to each participant's brain [Friston and Ashburner, 2004]. In the case of the CST, the first ROI defines the entire cerebral peduncle in an axial plane at the level of the decussation of the superior cerebellar peduncle, while the second ROI is placed in the central sulcus, right after the bifurcation to the motor and sensory cortex [Wakana et al., 2007].

In the third step, involving the refinement of segmented fiber tracts, each candidate fiber is compared with a fiber tract probability map created by Hua et al. [2008]. This fiber tract probability map specifies the probability that a voxel is associated with a given fiber tract. Thus, estimated fibers are discarded if their probability of belonging to a fiber is low. Since the tractography process is noise prone, fiber tracts are further cleaned by iteratively discarding outlier fibers that are more than four standard deviations above the mean fiber length or that deviate more than five standard deviations from the core of the fiber tract. The Mahalanobis distance is calculated to estimate the single fiber's contribution to the core, and it corresponds to the probability that a given point belongs to the distribution. Diffusion properties are assessed at each node of each fiber, on an individual subject basis, using spline interpolation of the diffusion properties: FA, MD, AD, and RD.

Statistical and Machine Learning Analysis

To compare the diffusion properties between ALS and controls, mean values and standard deviations (±SD) were plotted [Yeatman et al., 2012] and visually analyzed and then t-tests were conducted point-wise along the CST for 100 voxels. A permutation-based multiple-comparison correction was applied to determine a statistically significant threshold for P-values (P < 0.05) [Nichols and Holmes, 2002].

For exploring the relationship between diffusion properties in ALS patients and clinical parameters, a Pearson's correlation analysis was performed point-wise between voxel diffusion values and ALSFRS-R/bulbar scores and between voxel diffusion values and disease duration, with the aim of quantifying the degree of this relationship.

Although statistical separability is useful, it gives only a general evaluation of the differences between groups, and it is not able to help predict the diagnosis as well as the machine learning approach [Libero et al., 2015]. Furthermore, machine learning techniques offer the possibility to assess the clinical utility of diffusion parameters. One of the emerging methods to evaluate the prediction power of variables (i.e., how well a variable could discriminate between classes) is Random Forests [Breiman, 2001]. Random Forests is a collection or ensemble of Classification and Regression Trees (CART) [Breiman et al., 1984] trained on datasets of the same size as training set, called bootstraps, created from random resampling of the original training set. Once a tree is constructed, a set of bootstrap datasets, which does not include any particular record from the original dataset (out-of-bag examples), is used as test set. The error rate of the classification on test sets is the out-of-bag (OOB) estimate of the generalization error. Breiman [1996] showed by empirical evidence that, for the bagged classifiers, the OOB error is accurate as using a test set of the same size as the training set. To classify new input data, as depicted in Figure 2, each individual CART tree votes for one class and the forest predicts the class that has the plurality of votes.

Random Forest follows specific rules for tree growing, tree combination, self-testing and post-processing, it is robust to overfitting and it is considered more stable in the presence of outliers and in very high dimensional parameter spaces than other machine learning algorithms [Caruana and Niculescu-Mizil, 2006; Menze et al., 2009]. The concept of variable importance is an implicit feature selection performed by Random Forest with a random subspace methodology, and it is assessed by the Gini impurity criterion index [Ceriani and Verme, 2012]. The Gini index is a measure of prediction power of variables in regression or classification, based on the principle of impurity reduction [Strobl et al., 2007]; it is non-parametric and, therefore, does not rely on data belonging to a particular type of distribution. The Gini index is calculated as follows:

(1)

where D is a dataset with n classes and p_j is the relative frequency of class j in D. If a dataset D, with size N and n classes, is split into two subsets D₁ and D₂ with sizes N₁ and N₂, the Gini index is defined as:

(2)

For splitting a node, the attribute value that provides the smallest Gini_split(D) is chosen. A low Gini (i.e., a greater decrease in Gini) means that a particular predictor variable plays a greater role in partitioning the data into the defined classes.

In this study, the analyses were conducted by using R language (version 3.1.2) and the Random Forest package [Liaw and Wiener, 2002]. The dataset contained the CST voxel values of each subject, for each diffusion metric (FA, MD, AD, and RD). The left and right CST were kept separated in order to investigate any asymmetry and neither age nor sex were added as features owing to the almost perfect match of the three groups. In particular, this dataset has 48 rows (subjects), one diagnosis column—ALS or CTRL (healthy control)—and 800 numerical columns, that is, values of 200 voxels for each metric, 100 for each hemisphere. After a data cleaning—for removing columns with missing values (19)—the dimension of the feature space was 781. A binary classifier was then trained by Random Forest on this set, with 10,000 trees in the forest (ntree) and a recommended value for the number of variables considered at each split of [Diàz-Uriarte & Alvarez, 2006], where number of features = 781 (after the data cleaning) and mtry = 27. Voxels were then ranked according to their Mean Decrease Gini and all voxels with a mean decrease gini ≥ 0.1 were selected and extracted for creating a new training set with the most important voxels only. This cutoff was selected so to construct a new classifier with the lowest OOB error. The accuracy of the new classifier was evaluated by the 5-fold cross validation approach, which estimates the performance by randomly dividing for five times the set into five folds, where four folds are used for training and the remaining one for testing.

Univariate Tract-Profile Analysis

Figure 3 showed comparisons of diffusion parameters between ALS patients and controls for the CST profile. We found significant differences, surviving corrections for multiple comparisons, in the FA, MD, and RD metrics. The highest differences were detected using FA and RD metrics where DTI values in some parts of CST were not overlapping between groups (Fig. 3a,d). Indeed, ALS patients showed significant bilateral FA reduction and RD increase in the cerebral peduncle, posterior limb of internal capsule, corona radiata, and the primary motor cortex with respect to controls. Otherwise, MD and AD measurements showed small regions of differences (Fig. 3b,c). Considering the former, we depicted significant differences only in the right cerebral peduncle, whereas analysis of AD revealed a very restricted—but not statistically significant—pattern of difference in the right primary motor cortex.

Correlation Analysis

We performed correlation analyses between DTI metrics and disease severity/duration scales to evaluate the clinical relevance of the detected microstructural changes in the CST (Fig. 4). We found non-uniform and weak relationships. Indeed, none of the correlation coefficients reached the statistically significant threshold for multiple comparisons, expect for bulbar scores. Indeed, increasing of disease severity correlated with a reduction of microstructural integrity (as measured by AD values) within the right cerebral peduncle (Fig. 4).

Machine Learning Applied on Tract Profile

Random Forest classifier, trained on all the features, showed an OOB error of 25% and the confusion matrix of the classification of the training set is reported in Table 2.

After the feature selection, the classifier trained on the most important voxels (mean decrease gini ≥ 0.1), showed an OOB error of 16.67% (Table 2). The 5-fold cross validation revealed a mean accuracy of 80%, with a range between 75% and 87.5%.

The analysis of the variable importance, as extracted from the classifier trained on all voxels, showed that voxels influencing the classification of ALS with respect to CTRL were localized mainly in tracts along the cerebral peduncle until the corona radiata in the FA and RD metrics (Fig. 5). The MD-related measurements of the internal capsule in the right tract presented one significant voxel contributing to ALS classification. Furthermore, no AD voxels resulted to have a significant role in the class prediction, as it could be seen in the plot of the first fifty voxels on the left of Figure 5.