Given the absence of a gold standard for the pPFT population, we validated our tract reconstructions through a qualitative assessment by a panel of neuroanatomical experts. As a quantitative measure, we calculated the relative tract volume to whole brain white matter. The dMRI data used for the fiber tractography were obtained from our institution’s protocol optimized for intraoperative MRI (ioMRI) acquisition [18]. Data from healthy volunteers were included to assess whether our intraoperative dMRI protocol is of sufficient quality to reconstruct tracts in a healthy brain. We also examined whether a control tract outside the resection area maintained consistent reconstructions between the pre- and intraoperative acquisitions.

Ten pPFT patients (aged 3–17 years; four females) and two healthy volunteers (aged 25–27 years; one female) were included in this study (Table 1). The local ethics committee approved this study. All subjects and/or caregivers provided written informed consent.

We obtained all MR images of the pPFT patients with a 3-Tesla intraoperative MRI scanner (Philips Ingenia ElitionX MR-OR system, 70 cm bore, Philips Healthcare, Best, The Netherlands) with two RF coils (Fig. 2b). Both pre- and intraoperative images (Fig. 3a, b) were acquired on the day of surgery under the same anesthesia with the patient in a prone position (i.e., lying on their stomach with their chest lifted and chin-down), and the head secured in the DORO head clamp (Black Forest Medical Group, Fig. 2a). The preoperative scan is acquired before craniotomy. The intraoperative scan is acquired with an open skull at the time when most of the tumor is resected and the neurosurgeon decides to check for residual tumor tissue. If there is any remnant tumor visible on the intraoperative scan, the neurosurgeon has the option to continue to resect additional tumor tissue. MR images of the healthy volunteers were obtained with another 3-Tesla MRI scanner (Ingenia, 70 cm bore, Philips Healthcare, Best, The Netherlands) using the same two RF coils. Healthy volunteers were scanned once in the supine position (i.e., lying on their back with their head in a neutral position) without anesthetics or the head clamp.

Example of patient preparation for MR acquisition before surgery. A After positioning the patient in the surgical prone position (i.e., laying on their stomach with their chest lifted and chin-down), the head is secured in the DORO surgical head-frame. B The two RF coils are then placed anterior and posterior of the patients’ head and fixated with tape before sliding the patient into the intraoperative MRI scanner that is in a room adjacent to the operation theater

Example of T1-weighted images before and during surgery and placement of the regions of interest used for fiber tractography. T1-weighted (T1w) images of a 14-year-old girl with a 4th ventricle tumor before resection (A, preoperative) and after surgical resection but before closure of the skull (B, intraoperative). The red arrow in A indicates the tumor in situ. To reconstruct the dentato-rubro-thalamic tract, the dentate nucleus (C), thalamus (D), and primary motor cortex (E) were included as regions of interest (ROIs), indicated in green. For detailed segmentation results see appendix 3, Fig. 1 (Supplementary material)

The scan protocol consisted of a T1-weighted (T1w) and three dMRI acquisitions (Table 2). The T1w images were used for registration and region of interest (ROI) segmentation purposes. The dMRI acquisition was a spin-echo EPI using Stejskal-Tanner gradients. The first dMRI acquisition contains one non-weighted diffusion image (b = 0 s/mm2, b0 image) and twenty images with different diffusion directions at a b-value of 1000 s/mm2 (b1000). This acquisition is sent to the clinical PACS system as a standard diffusion tensor image (DTI) series. The second acquisition completes the multi-shell dMRI dataset and contains five b0 and 32 images with different diffusion directions at a b-value of 2000s/mm2 (b2000). The gradient directions were optimized for the second shell on a half sphere using QMRITools [19]. For the first shell, we had to rely on the predefined gradient table from Philips to ensure compatibility with the neuronavigation system (Brainlab, Munich, Germany). The optimization of this gradient table was done by electrostatic repulsion on half a sphere with antiparallel gradients. In healthy volunteers, susceptibility distortions were partly corrected by affine registration to the T1w image [20]. In patients, we added a third acquisition with reversed phase encoding direction settings to correct for the susceptibility distortions that are expected during the ioMRI acquisition using FSL topup [20]. The time of the three dMRI acquisitions was ~ 9 min and has been optimized for the surgical setting [18]. The repetition time (TR) was set on a range from 7000 to 8200 ms for all dMRI acquisitions to allow for variability in head orientation and to ensure thermal safety. Certain head angulations can lead to increased gradient load, and the system automatically adjusts the TR within the specified range to prevent overheating. However, as can be seen in Appendix 7 (Supplementary material), the TR values for all patients and measurements in the study varied very little; all TR values were between 7720 and 7734 ms.

We segmented brain regions that are used as either inclusion regions, through which a streamline must pass (AND), and exclusion regions through which streamlines cannot pass (NOT). To reconstruct the DRTT, we used the dentate nucleus, thalamus, and primary motor cortex as inclusion regions and the corpus callosum and preoperative tumor volume as exclusion regions [4]. All regions were segmented based on the T1w images. Automatic segmentation of these regions, especially the dentate nucleus, was challenging due to the strong brain deformations that result from the preoperative tumor or intraoperative resection cavity. To address this, we performed separate segmentations on the pre- and intraoperative datasets, avoiding misregistration across sessions. Whole-brain gray and white matter masks were generated using the online brain segmentation pipeline of volBrain [21] combined with the CERES pipeline for improved detail of the cerebellum [22]. Thalamus segmentations were retrieved from the volBrain pipeline. We then registered the whole brain gray and white matter masks to the SUIT template [23]. The resulting registration matrix was then used to register the SUIT cerebellar atlas, which included the left and right dentate nuclei [24], back to the native gray and white matter masks (see appendix 3 in Supplementary material for detailed segmentation results). Tumor tissue of the pPFT patients was semi-manually segmented using the Smart Brush tool in the Brainlab neuronavigation system (Smart Brush, Brainlab, Munich, Germany). The anatomical accuracy of the dentate nucleus and tumor volume were verified and adjusted by a pediatric neuroradiologist (W.N., 5 years of experience), using both T1w and T2-weighted (T2w) scans available in the patient archiving system. The T2w images offer improved visibility of the dentate nucleus [25]. The T1w image was linearly registered to the dMRI data using an affine transformation and was then used for the co-registration of the ROIs to subsequently use them in the fiber tractography pipeline (Fig. 3c, d). The primary motor cortex was directly registered from the Desikan atlas [26] to the dMRI data, and the corpus callosum from the Destrieux atlas [27] in a similar way.

To reconstruct the corticospinal tract (CST) as a control tract outside the resection area, we included the primary motor cortex and brainstem regions. Exclusion regions for the CST were the corpus callosum, thalamus, cerebellum, contralateral hemisphere white matter, superior frontal gyrus, and occipital and temporal lobes (all regions derived from the Desikan atlas [4, 26]. For pPFT patients, preoperative tumor tissue was additionally used as an exclusion region for preoperative CST reconstruction.

Data of the dMRI acquisitions were used to reconstruct the DRTT and CST. Preprocessing was done with the fully automated MRIToolkit pipeline (https://github.com/delucaal/MRIToolkit.git) and ExploreDTI [28, 29]. Data preprocessing involved correcting for signal drift [30], denoising with the Marchenko-Pastur Principal Component Analysis (MPPCA) [31], correction for eddy currents and subject motion utilizing affine registration, and b-matrix correction [29]. Brain masking was done with the Brain Extraction (BET) FSL toolbox [32]. We corrected susceptibility distortions by means of the reversed phase encoding direction data using FSL topup [20]. After preprocessing, the DKI model was fitted [33, 34], as in our previous work [35]. Given that DKI is prone to spurious mean kurtosis (MK) values, we employed the MK-curve correction method [33, 34]. The Generalized Richardson Lucy spherical deconvolution method [28] was applied to reconstruct fiber orientation distributions in white matter while accounting for partial volume with gray matter and cerebrospinal fluid.

The tractography analysis was performed in two steps. In the first step, we aimed to optimize the tracking parameters specifically for reconstructing the DRTT in our dataset. To this end, we seeded from the dentate nucleus and the thalamus. In the second step, we applied the optimized parameters identified in step one to perform whole-brain seeding. This allowed us to reconstruct control tracts located outside the DRTT region.

To identify the best deterministic fiber tractography parameters in our dataset, we created eight different combinations of parameter combinations varying the angle threshold (45° or 60°) and the fiber orientation distribution (FOD) amplitude threshold (0.1, 0.05, 0.01, or 0.005). Streamlines were terminated if they changed direction by an angle exceeding the angle threshold or if the local FOD amplitude was below the FOD threshold. Tractography was performed using deterministic FACT propagation (ExploreDTI, [29] with seed points placed every 0.5 mm isotropically. The step size was fixed at 1 mm for all experiments, given its intrinsic relation with the angle threshold. For parameter optimization, streamlines were seeded from the dentate nucleus and contralateral thalamus (d-DRTT), or ipsilateral thalamus (nd-DRTT). We then gated the tracts through the inclusion and exclusion regions defined in section “Region of interest segmentation”.

Given the lack of a gold standard, such as postmortem tractography or intraoperative electrical stimulation, we validated the tracts by consulting an anatomical expert panel to identify the parameter combination that aligns best with the neuroanatomy of the DRTT. This qualitative assessment has been detailed in paragraph 2.6.

After determining the best deterministic fiber tractography parameters for DRTT reconstruction in our dataset, we repeated the tractography using whole-brain seeding to enable reconstruction of control tracts beyond the DRTT region. To contain data size, the spacing between seed points was set to 1 mm isotropic. Streamlines were terminated at the gray-white matter interface. For the final DRTT and CST reconstructions, we gated through the above-mentioned (paragraph 2.4) inclusion and exclusion regions.

For quantitative analysis of the tracts, we computed the volume of the pre- and intraoperative DRTT and CST by counting the voxels that contain at least one streamline normalized to the whole brain white matter volume. The tract volumes of the pre- and intraoperative data are summarized in a violin/boxplot format, generated using the default settings in R’s ggplot2 [36]. The violin plots show a smoothed estimate of the data distribution, derived with a Gaussian kernel density estimator (frequency histogram) [37]. Differences in relative tract volumes of patients were analyzed by means of a paired Wilcoxon signed-rank test (significance level p < 0.05). To generate an SNR map, we divided the mean of the signal intensity for each voxel of all b0 images by their standard deviation. We then computed the average SNR across all voxels within the white matter mask.