Radiomic white matter parameters of functional integrity of the corticospinal tract in high-grade glioma

Tractography has become a widely available tool for the planning of neurosurgical operations as well as for neuroscientific research. The absence of patient interaction makes it easily applicable. However, it leaves uncertainty about the functional relevance of the identified bundles. We retrospectively analyzed the correlation of white matter markers with their clinical function in 24 right-handed patients who underwent first surgery for high-grade glioma. Morphological affection of the corticospinal tract (CST) and grade of paresis were assessed before surgery. Tractography was performed manually with MRTrix3 and automatically with TractSeg. Median and mean fractional anisotropy (FA) from manual tractography showed a significant correlation with CST affection (p = 0.008) and paresis (p = 0.015, p = 0.026). CST affection correlated further most with energy, and surface-volume ratio (p = 0.014) from radiomic analysis. Paresis correlated most with maximum 2D column diameter (p = 0.005), minor axis length (p = 0.006), and kurtosis (p = 0.008) from radiomic analysis. Streamline count yielded no significant correlations. In conclusion, mean or median FA can be used for the assessment of CST integrity in high-grade glioma. Also, several radiomic parameters are suited to describe tract integrity and may be used to quantitatively analyze white matter in the future.

White matter tracts in the brain have always been of interest to neuroscientists, who studied both their anatomy and their function 1,2 .Due to their organization in networks, connectomic analyses have been carried out to further understand brain function and plasticity 3 .While these approaches carry great potential to unravel new insights, it is of utmost importance to choose physical quantities that actually describe the connections that are made by axons in white matter tracts between different brain regions.However, little data exists to address this question, especially in the presence of brain tumors 4,5 .
Information in the brain travels along axons on an electrical basis and is transferred in synapses, which are mostly chemical 6 .When a larger number of axons travels together, they form a white matter tract, which gains its volume due to the surrounding myelin sheaths.These sheaths electrically isolate the axons and form a diffusion barrier that causes diffusion tensors to point in the tract direction when measured with diffusion tensor imaging (DTI) sequences in magnetic resonance imaging (MRI).Diffusion tensors can be used to create tractograms by using highly sophisticated algorithms that are able to consider crossing fibers as well as the presence of tumors, while still relying on expert interaction in the workflow 7 .Further, diffusion markers like fractional anisotropy (FA) can be calculated per voxel.FA is a measure of the directedness of diffusion and can take on values between 0 and 1.As of now, it is the most robust marker of white matter integrity [8][9][10][11][12][13][14] .
Tractograms usually consist of streamlines, describing the probability that a white matter tract is present along their coordinates 12 .This spatial information can then be used to define volumes of the depicted tracts 15 .While a great number of connectomic studies use streamline count or its derivates to measure structural connectivity, its biological relevance has not yet been proven 16 .On the other hand, diffusion measures like the mean FA of voxels in a tract volume have been shown to relate to outcome and function [17][18][19] .
Since automatic tractography methods are within reach to become feasible for routine data analysis, machine learning algorithms are of growing interest to handle the expansive amount of data 15 .However, feature selection is essential for the successful analysis of high-dimensional data 20 .

Imaging
MRI was conducted within seven days prior to surgery with a 3 T system (Ingenia, Philips, Eindhoven, Netherlands) using a single-shot echo-planar imaging diffusion tensor imaging (DTI) sequence (TR/TE = 7010/102 ms; FOV = 222 × 222 mm 2 ; matrix 112 × 112; 50 slices without gap; slice thickness 2.7 mm; 32 non-collinear directions, b-value = 1000 s/mm 2 ) and contrast-enhanced T1 weighted MPRAGE 3D dataset (TR/TE = 8.1/3.7 ms; FOV = 222 × 222 mm 2 ; matrix = 512 × 512; 170 slices without gap, thickness 1 mm) using a dedicated head coil.The data was preprocessed with the dwipreproc 23 command in MRtrix3 (www.mrtrix.org) 24 including eddy correction and underwent a visual quality check.CST tractography was performed first manually with MRtrix3 17 , and also automatically in TractSeg. 15Tractography parameters included an FA cutoff of 0.15, a maximum length of 250 mm, a minimum length of 60 mm, and 10 6 seeds.The seed region was placed in the mesencephalon, and the target region in the internal capsule.An experienced neurosurgeon and an experienced neuroradiologist removed false-positive streamlines that were identified as collaterals to other tracts or showed clear signs of noise.CST affection was assessed in comparison to the contralateral CST by an experienced neuroradiologist and an experienced neurosurgeon, who were blinded for patient data, with: 0-unaffected (symmetric), 1-dislocated (asymmetric), 2-compressed (reduced volume), 3-infiltrated (contact or overlap with tumor).We recorded mean and median FA and streamline count from manual tractograms, and radiomic parameters from automatic tractograms 25 .

Radiomics
Radiomics were calculated upon the volumes of the respective tractograms of the CST.The space was defined in the FA map, so that voxel based calculations used the respective FA.A complete description is available under https:// pyrad iomics.readt hedocs.io/ en/ latest/ featu res.html 25 .

FA energy
Energy is a measure of the magnitude of all voxel values within the respective volume, where a larger value marks a larger sum of the squares 25 .

FA kurtosis
Kurtosis measures the shape of the distribution of values, where a lower kurtosis marks a high concentration of values in form of a peak near the mean.

Maximum 2D diameter column
This is the maximum slice diameter of the CST in the axial plane.It is defined as the largest pairwise distance between volume surface mesh vertices in this plane.

Maximum 3D diameter
This is defined as the largest pairwise Euclidean distance between volume surface mesh vertices, also known as the Feret Diameter.

Elongation
Elongation reflects the ratio of the two largest principal components of the respective volume, where λ minor and λ major are the lengths of these components.It can take values between 0 and 1 and is defined as the inverse of true elongation for computational reasons.

Surface area
Surface area is calculated with a mesh of triangles around the respective volume.

Voxel volume
This volume is calculated through multiplication of the single voxel volume V k with the number of voxels in the respective volume.It is less precise than mesh volume and is not used in further calculations.

Mesh volume
This volume is calculated from the mesh of triangles around the respective space by calculating the volume V i of the tetrahedrons within that space defined by the origin O and the points a i , b i , and c i .

Surface volume ratio
This ratio is calculated with surface area A and mesh volume V.

Statistics
Statistical analysis was carried out with SPSS Statistics 27 (IBM, Armonk, NY, USA).FA values of all tracts were tested for normal distribution after D' Agostino-Pearson yielding non-normal distribution in all cases.Correlations were tested by bivariate non-parametric Spearman's rank order for significance.P-values < 0.05 were considered statistically significant.Receiver-operating-characteristics analysis was applied to report area under the curve (AUC), sensitivity, and specificity for each variable.For further investigation, the analyzed variables were sorted by their respective p-value.Significant correlations were then plotted in a diagram with the dependent variable.

Results
Screening of 42 patients yielded 24 patients (42% female) with high-grade glioma (five grade 3, 19 grade 4).Median time between preoperative imaging and surgery was one day (range 1-5 days).Six patients suffered from preoperative contralateral paresis.Epidemiological data is shown in Table 1.Example tractograms are shown in Fig. 1.In morphological affection rating, CSTs were dislocated by glioma growth in four patients, compressed in six and considered to be infiltrated in five.In the remaining nine patients CSTs were marked unaffected.

Discussion
This work analyzed the relation of different white matter parameters with clinical presentation in both radiological and clinical examination.The corticospinal tract (CST) was chosen due to its anatomical size and clinical importance in a large portion of neurological functions 26 .Damage to the CST is easily detectable and can be quantified in clinical presentation while being routinely demonstrated in MRI and more so in MRI tractography.www.nature.com/scientificreports/Furthermore, neurological deficits show significant correlations with overall survival in high-grade glioma, since a certain level of neurological function is always required for adjuvant treatment 27,28 .Neurological function and explicitly motor function is on the other hand important for maintaining quality of life in the treatment of high-grade glioma 29 .
The growing field of tractography studies motivates both clinicians and researchers to find better tools for patient evaluation and patient counseling by depicting and measuring different tracts or whole connectomes in patients with stroke, multiple sclerosis, or brain tumors 5,30,31 .The depiction of tract anatomy is becoming increasingly precise, which enables analyses of higher orders like connectomics 32 .However, surrogate variables of tract damage should be evaluated for their respective value before drawing conclusions from them in higher order analyses.Although FA has been shown to be a feasible marker of white matter integrity, streamline count is often used in the analysis of tractography or connectomes 33 .
In the presented analysis of high-grade glioma patients, we found no significant correlations of streamline count with radiological or clinical CST affection, which is well visible in the scatter plots of our data.Interestingly, different measurements of FA and CST volume showed different correlations with CST affection.Mean and median are easily applicable in both clinical and experimental setups, and have been demonstrated in our screening approach to indicate functional CST affection in imaging and neurological examination.This holds also true for mesh volume, voxel volume, surface area, and energy.While the latter is derived from FA values, it is interesting that its correlation with both radiological and clinical CST affection is stronger than mean and median.Since energy is the sum of the squared values of every voxel, it is volume-confounded, which may be the cause of its stronger correlation with CST integrity.
It is visible in the scatter-plots that voxel volume, maximum 3D diameter, mesh volume, and surface area correlate foremost with CST compression.This is to expect, since a reduced volume was the definition for compression in our data.Nevertheless, these variables correlate also with contralateral paresis, which points both towards intervariable correlations and overlapping effects of compression, infiltration, and function.
Surprisingly, three geometric variables correlate significantly only with paresis.Maximum 2D diameter column, minor axis length, and elongation seem to indicate CST function by its shape, independent from radiological CST affection.They may therefore be useful in automatic categorization of CST affection that is not easily assessable by expert interaction.
The fact that FA kurtosis is a strong indicator of CST function, but not radiological affection, is in line with previously published studies that mentioned postoperative FA in the CST to be more condensed around the mean.Interpreted together, it could indicate higher kurtosis and better motor function 16 .

Limitations
The data was retrospectively collected.Further, different MRI scanners and different tractography algorithms may deliver different results, which remains to be disproven.Therefore, prospective multicenter studies with comparison of different scanners and algorithms would be needed for confirmation or rebuttal of comparability.Also, CST affection was measured manually, which leaves room for uncertainty of the measurement.However, we believe, that reproducibility is given by following the thorough description of affection rating.Nevertheless, further studies which may apply machine learning approaches should analyze larger cohorts.

Conclusions
Our results confirm earlier studies that found FA to be a good indicator of white matter integrity and expand this finding for high-grade glioma.It is further shown that FA-derived radiomics like energy and kurtosis, as well as geometric radiomics like volume and elongation, are suitable indicators of tract integrity and can be used in further research on the functional relevance of white matter tracts.Streamline count should not be used for functional analysis.

Figure 1 .
Figure 1.Example tractograms of the CST.Top row: Manual tractography result with MRTrix3 in coregistered contrast-enhanced T1 weighted MPRAGE 3D dataset.Bottom row: Automatic tractography result with TractSeg in a coregistered skull-stripped FA map.Left: coronal view.Right: sagittal view.Orientation in radiological convention.

Figure 2 .
Figure 2. Scatter plots of significant correlations and streamline count with CST affection.

Figure 3 .
Figure 3. Scatter plots of significant correlations and streamline count with paresis.
Minor axis length This is the second largest axis of the enclosing ellipsoid and calculated with the largest principal component λ minor .

Table 2 .
Markers of radiological affection.Significant correlations (two-tailed) of radiomics and CST affection with the respective p-values.A negative correlation coefficient r indicates less affection.

Table 3 .
Markers of motor function.Significant correlations (two-tailed) of radiomics and CST function (paresis) with the respective p-values.A positive correlation coefficient r indicates better motor function.