Subjects and MRI acquisition

Thirty-three patients with a diagnosis of nfPPA and 29 patients with a diagnosis of svPPA were retrospectively included in the study. Controls (n = 39) included both 10 healthy subjects and 29 subjects with subjective cognitive decline [60] and negative CSF biomarkers for neurodegeneration (amyloid and tau pathology). For clarity, the term ‘controls’ will henceforth refer collectively to both healthy volunteers and individuals with subjective cognitive decline (SCD) and negative CSF biomarkers.

NfPPA and svPPA were diagnosed according to established diagnostic criteria [5, 61] by specialized board-certified neurologists in all patients. Data acquisition was performed at two study sites (Ulm, Germany and Tricase, Italy). Participants scanned in Ulm are designated as cohort “A”, and those scanned in Tricase as cohort “B”. The sum of boxes of the FTLD-Clinical Dementia Rating (FTLD-CDR) as a validated score for measuring progression of functional decline [62] was available for all patients. The study was approved by the local ethics committees (Ethics Committee of the University of Ulm, reference 39/11 and Institutional Review Board of Azienda Sanitaria Locale Lecce, report n. 6, 25 July 2017, respectively), and written informed consent was obtained from each participant or the primary caregiver in accordance with the Declaration of Helsinki.

Among the participants of cohort A, the 30 nfPPA, 25 svPPA patients, and 15 controls underwent MR scanning on a 3.0T scanner (Ulm: Allegra, Siemens Medical). Among the participants of the cohort B, 3 nfPPA, 4 svPPA patients, and 24 controls underwent MR scanning on a 3.0T scanner (Tricase: Philips Ingenia). DTI and T1-weighted data were acquired using center-specific protocols. The DTI protocol used for cohort A consisted of 31 gradient directions (GD), including one b = 0 GD (80 slices, 112 × 128 pixels; slice thickness was 2.0 mm, in-plane pixel size was 2.0 × 2.0 mm2). The echo time (TE) and repetition time (TR) were 88 and 11,100 ms; b was 1000 s/mm2. The DTI protocol used for cohort B consisted of 65 GD, including one b = 0 GD (60 slices, 96 × 96 pixels; slice thickness was 2.5 mm, in-plane pixel size was 2.5 × 2.5 mm2); TE and TR were 85 and 6852 ms; b was 1000 s/mm2.

The structural Fast-Field Echo (FFE) T1-weighted data acquired for cohort A included 144 sagittal slices with 1.2 mm thickness, 1.0 × 1.0 mm2 in-plane resolution in a 256 × 248 matrix, TE = 4.2 ms, and TR = 1640 ms. The FFE T1-weighted sequences for cohort B were acquired with the following parameters: 200 slices, TR = 8.2 ms, TE = 3.8 ms, field of view = 256 × 256 mm2, flip angle = 8°, and voxel size = 1 mm3 isotropic.

Among the participants from the cohort A, 10 nfPPA and 6 svPPA patients had a follow-up scan including DTI with a time-interval of in average 12 months. The average time-interval between baseline and follow-up MRI was 11.80 ± 1.40 months for nfPPA patients and 12.20 ± 1.60 months for svPPA patients, with no significant difference between groups (p = non-significant). The remaining patients and controls were not available for a second MRI investigation due to progression of the clinical symptoms and associated inconveniences.

Microstructural MRI analysis—DTI—standardized data pre-processing

The DTI analysis software Tensor Imaging and Fiber Tracking (TIFT) [63, 64] was used for post-processing and statistical analysis.

MNI normalization

After motion correction of individual DTI data sets, baseline and follow-up DTI data were aligned by fitting the (b = 0) volumes to minimize intensity differences, using halfway linear registration matrices to avoid baseline data bias [65]. Subsequently, baseline and follow-up data underwent stereotaxic Montreal Neurological Institute (MNI) transformation using identical parameters. After quality control including visual examination, 4 nfPPA and 2 svPPA subjects were discarded due to low data quality.

Spatial normalization to the MNI stereotaxic standard space was performed iteratively [66]. This process utilized a study-specific (b = 0) template and an additional FA template for the second iteration, as the FA template provides greater contrast than (b = 0) images [67]. The correlation between individual FA maps and the FA template exceeded 0.7 after two iterations, so the iterative process was halted [68]. Directional information during the normalization process was preserved using techniques described by Alexander et al. [69]. Eventually, 29 nfPPA and 27 svPPA patients were included in the analyses of the current study. An isotropic three-dimensional 8 mm full-width at half-maximum Gaussian smoothing filter was applied to the individual normalized FA maps. This filter size, approximately two-to-three times the recording voxel size depending on the protocol, offers a good balance between sensitivity and specificity [64].

Inter-center correction

Fractional anisotropy maps from the different protocols were corrected for age using regression models based on datasets from 15 and 24 controls, separately for the two contributing centers. Subsequently, FA maps of patients with nfPPA, svPPA, and controls were harmonized for center by applying respective 3-D correction matrices (linear first-order correction). These 3-D correction matrices were derived as linear adjustments based on differences in the DTI scans of controls from each center [52, 70]. No residual site effects could be detected after inter-center correction.

Definition of tract structures

To apply group-based fiber tracking (FT) algorithms [69], an averaged template DTI dataset was generated from 24 controls using the same DTI protocol. This involved arithmetic averaging of the MNI-transformed data. Eigenvectors and eigenvalues were calculated for each voxel position, representing the average of the 24 controls. Only controls were used to avoid bias from pathology-induced alterations. Directional information from each dataset was preserved during normalization and incorporated into template creation [69].

This averaged DTI dataset from controls was then utilized to identify pathways for defined TOIs for the four groups of patients. Given that no neuropathological staging hypothesis has been put forward for individual clinical PPA syndromes due to the heterogeneity of their underlying neuropathology [10, 71], the previous DTI studies in nfPPA [13, 17, 18, 29, 33, 42, 72, 73] and svPPA [13, 17, 18, 29, 33, 42, 72,73,74] were used as references for TOI definitions. White matter tracts were selected if they were described to be affected in >75% of the selected neuroimaging papers [75].

A seed-to-target approach was used [76, 77]. Seed regions were defined for both the seed and target regions. For fiber tracking, only voxels with an FA value above 0.2 were considered. A modified probabilistic streamline tracking approach, which accounts for the directional information of neighboring fiber tracts, was used for fiber tracking [69]. All fiber tracts originating in the seed regions and terminating in the target regions of the respective pathway defined the corresponding TOI.

The technique of tract-wise fractional anisotropy statistics (TFAS) [63] was employed to quantify the tractography results using the TOIs. The FA values of the specific tracts were arithmetically averaged for each stereotaxically normalized DTI dataset of each subject. The following TOIs were thus defined for the nfPPA and svPPA clinical syndromes: left and right uncinate fasciculus (UF), genu, section II, III IV and splenium of the corpus callosum (CC), left and right superior longitudinal fasciculus (SLF), left and right inferior longitudinal fasciculus (ILF), inferior fronto-occipitalis fasciculus (IFOF), cingulum, pontine projections, anterior thalamic radiation, corticostriatal projections, corticospinal tract (CST), optic radiation, fornix, and left and right tapetum.

Structural MRI analysis: atlas-based volumetry

The T1-weighted data were processed using MATLAB (version R2014b, The Mathworks, USA) and the Statistical Parametric Mapping 12 (SPM12) software (Wellcome Trust Center for Neuroimaging, London, UK, www.fil.ion.ucl.ac.uk/spm), following a standardized processing pipeline for ABV; ABV has already been successfully employed in cross-sectional and longitudinal studies [54, 56,57,58, 78]. The processing steps included: (1) segmentation into gray matter, white matter, and cerebrospinal fluid (CSF) compartments, (2) stereotaxic normalization into MNI space, and (3) volumetric assessment using voxel-by-voxel multiplication and subsequent integration of normalized modulated component images (GM, WM, or CSF) with predefined masks from various brain atlases.

To improve the quality of atlas space mapping, high-dimensional registration methods were introduced, showing intrascanner variability of volumetric results to be less than 1% for most investigated structures [55]. All volumetric results were linearly standardized to the mean intracranial volume (ICV) of controls. Group-level differences were tested for significance after correction for multiple comparisons.

Two controls and 1 subject with nfPPA had to be excluded from the analysis as their T1-weighted data were compromised by artifacts.

A series of standard cortical and subcortical SOIs were chosen for ABV analysis [78]. These were the cerebrum gray and white matter, the frontal lobes, the temporal lobes, the parietal lobes, the occipital lobes, the insulae; the cerebellum, the brainstem, the left and right hippocampus, the left and right amygdala, the left and right caudate, the left and right putamen, and the left and right thalamus.

DTI post-processingWhole brain-based voxel-wise statistics at the group level

Whole brain-based spatial statistics (WBSS) [52, 79] was used to calculate cross-sectional differences in FA maps. Statistical comparisons between the patients with nfPPA (n = 29), svPPA (n = 27) and 39 controls were performed voxel-wise using the Welch’s test, with the FA threshold set at 0.2 [79, 80]. Statistical results were corrected for multiple comparisons using the false discovery rate (FDR) algorithm at p < 0.05 [81]. Additionally, Type 1 error was further reduced by applying a spatial correction algorithm that eliminated isolated voxels or small clusters of voxels within the size range of the smoothing kernel, resulting in a cluster-size threshold of 256 voxels (256 mm3).

WBSS was also performed to calculate longitudinal differences in FA maps. White matter FA values were corrected for age [82], and statistical voxel-wise comparisons of FA for 10 patients with nfPPA and 6 patients with svPPA were performed versus the 39 control datasets acquired at baseline. The results were corrected for multiple comparisons using the FDR algorithm at p < 0.05, with an additional cluster-size correction for type 1 error as previously described.

Cross-sectional tract-wise comparison at the group level

To quantify the directionality of the underlying tract structures, the technique of tract-wise FA statistics (TFAS) [63] was applied. Age and scanner-corrected FA maps from baseline scans of patients with nfPPA (n = 29), svPPA (n = 27), and 39 controls were used to calculate mean FA values for the investigated tracts. Cross-sectional comparisons of mean FA values between patients and controls were performed.

Cross-sectional region-wise comparison at the group level

For the detection of volumetric alterations, group-level differences in ABV of nfPPA and svPPA compared to controls, respectively, and nfPPA compared to svPPA, were assessed for statistical significance following FDR correction for multiple comparisons. Z-scores were calculated as the difference between the subject’s mean and the control group’s mean, divided by the control group’s standard deviation.

Longitudinal tract-wise comparison at the group level

Fractional anisotropy maps from patients with nfPPA (n = 10) and svPPA (n = 6) and controls (n = 39) who had received at least one follow-up scan were analyzed to calculate group-averaged differences in the staging-associated tracts.

Longitudinal region-wise comparison at the group level

For volumetric SOIs, group-level differences for longitudinal data were assessed for statistical significance versus controls, following correction for multiple comparisons with FDR. Z-scores were calculated as the difference between the subject’s mean and the control group’s mean, divided by the control group’s standard deviation.

Cross-sectional correlation of FA maps to clinical scores

Fractional anisotropy maps from 29 patients with nfPPA and 27 patients with svPPA were voxel-wise correlated to the sum of boxes of the FTLD-CDR scores. Results were corrected for multiple comparisons with the FDR and cluster-size approach described above. For the correlation analyses, spherical regions of interest (ROIs) were placed bi-hemispherically within the peak results clusters of WBSS. The average FA values underlying the respective fiber tracts (as estimated by TFAS analysis) were also correlated to sum of boxes of the FTLD-CDR scores.

Classification by machine learning: random forest model

For the selection of the most important features to distinguish patients with nfPPA and svPPA from controls, two random forest algorithm were implemented. This model allows an understanding of the hierarchical order of the selected features by means of the Gini importance. As the model was trained using FA values from the TOIs employed in the TFAS analyses along with SOIs employed in the ABV analysis, its application was instrumental in identifying key SOIs and TOIs that significantly contributed to the classification task [83,84,85]. Furthermore, random forest models have been used successfully in a former study with a similar data structure combining DTI and T1-weighted data [86].

For model development, the dataset was divided into 80% training and 20% validation cohorts. Two controls and one subject with nfPPA had to be excluded from the analysis as their volumetric data were compromised by artifacts. The training cohort included 30 controls, 22 patients with nfPPA and 22 patients with svPPA. For validation the remaining 7 controls, 5 patients with nfPPA and 5 patients with svPPA were used. To reduce the risk of overfitting because of the limited sample size, a fivefold cross-validation was applied [87]. This means that each model for a defined feature selection was implemented five times: in each iteration, training was done on fourfold (80% of the data), and validation was performed out on the remaining fold (20%). This iteration was repeated until each fold had served once as the validation set. For all models in an iteration, cross-validated average accuracy, sensitivity, and specificity was calculated (Supplementary Table 7).

The random forest classifiers were implemented using the Scikit-learn library [88]. They were configured with the following key hyperparameters: the number of trees was fixed at 100, the Gini index was used as the splitting criterion, and the maximum depth of the trees was set to 4. Additionally, the minimum number of samples required to split a node was set to 2, and the minimum number of samples required at a leaf node was set to 1. For each split, a random subset of features (square root of the total number) was considered (max_features = ‘sqrt’), and bootstrap sampling was enabled. All models were trained using a fixed random state to ensure reproducibility. Feature selection was performed iteratively by removing the least important features based on the Gini index, retaining only those that maintained or improved classification accuracy [83, 87].