On PTV definition for glioblastoma based on fiber tracking of diffusion tensor imaging data

Radiotherapy (RT) is commonly applied for the treatment of glioblastoma multiforme (GBM). Following the planning target volume (PTV) definition procedure standardized in guidelines, a 20% risk of missing non-local recurrences is present. Purpose of this study was to evaluate whether diffusion tensor imaging (DTI)-based fiber tracking may be beneficial for PTV definition taking into account the prediction of distant recurrences. 56 GBM patients were examined with magnetic resonance imaging (MRI) including DTI performed before RT after resection of the primary tumor. Follow-up MRIs were acquired in three month intervals. For the seven patients with a distant recurrence, fiber tracking was performed with three algorithms and it was evaluated whether connections existed from the primary tumor region to the distant recurrence. It depended strongly on the used tracking algorithm and the used tracking parameters whether a connection was observed. Most of the connections were weak and thus not usable for PTV definition. Only in one of the seven patients with a recurring tumor, a clear connection was present. It seems unlikely that DTI-based fiber tracking can be beneficial for predicting distant recurrences in the planning of PTVs for glioblastoma multiforme.


Introduction
A good outcome in radiation therapy requires a good definition of the planning target volume by means of imaging techniques like computer tomography (CT) or magnetic resonance imaging (MRI). For brain tumors, the use of MRI is advisable because of the good soft tissue contrast [1]. In case of glioblastoma (GBM, annual incidence of approximately 5000 cases in Germany, which has a population of roughly 83 million people [2]), MRI images are acquired before and after surgical removal of the visible tumor area for defining the clinical target volume (CTV). The "American Society of Clinical Oncology" recommends resection with consequent external beam radiation therapy (RT) [3,4]. The target volume should include the area of the tumor before resection and the cavity after resection. Moreover, anatomical limits like edema, bones etc. should be taken into account when defining the target volume [3,4] five year survival rate is between 25% and less than one percent depending on age and therapy [5,6]. This stresses the need for better therapeutic approaches including, for example, a better target volume definition [7]. In about 20% of the glioblastoma patients, their recurrence was diagnosed as distant [8] and invasive glioma cells are found to migrate along myelinated fiber tracts of white matter [9]. The orientation of these tracts can be revealed with magnetic resonance diffusion tensor imaging (DTI) [10]. In DTI, the average water diffusion properties in an image voxel are detected and the larger diffusion parallel to white matter fiber tracts indicates their direction. Based on these data, white matter tracts can be reconstructed with fiber tracking techniques [11]. In a pioneering work, Krishnan et al. found that DTI-based fiber tracts connected primary glioma tumors with secondary lesions in eleven out of fourteen patients. They stated that this approach may be used to modify stereotactic radiotherapy target volumes to provide elongated treatment margins along the paths of elevated water diffusion, thereby creating a biologically better treatment plan that may reduce the incidence of progression [12].
Following this idea, we hypothesized that fiber tracking might be helpful for the definition of PTVs for radiation therapy of GBMs, and that the advances in scanner hardware and fiber tracking algorithms might lead to improved results. The aim of this work was to test this hypothesis.

Patients
One hundred fifty-two patients were treated for brain tumors at the Department of Radiation Oncology at the University Hospital Erlangen from January 2015 until March 2017. Fifty-six patients were diagnosed with GBM, 30 of which developed a recurrence. Seven out of these 30 patients had a distant recurrence. These seven patients were included in this retrospective study.
The department of Neuroradiology and the institute of Radiology, which were involved in the patients´treatment, agreed on using routine clinical data. All patients provided written informed consent when enrolled to the IRB approved study (No. 265_14B, Clinicaltrial ID: NCT02600065, ethics committee of the Friedrich-Alexander University Erlangen-Nürnberg).

CT imaging
The planning CT for initial tumors as well as for recurrences was acquired at Sensation Open (Siemens Healthineers, Erlangen, Germany) with acquisition matrix 512 × 512. The voxel size depended on the reconstruction type. The slice thickness varied between 1 mm and 3mm and the pixel spacing varied between 0.7 × 0.7 mm 2 and 0.98 × 0.98 mm 2 .

Image registration
The radiation therapy treatment was planned with the treatment planning system (TPS) Pinnacle 3 (Philips Healthcare, Eindhoven, The Netherlands), with the target volume (gross tumor volume according to ICRU Reports 50 and 62) definition performed in iPlan RT Image 4.1.2 (Brainlab, Munich, Germany) and based on rigidly registered CT and MRI acquired before RT (CT plan and MRI plan ) by experienced oncologists in the routine workflow. The initial gross tumor volume (GTV), PTV, edema, and organ at risk (OAR) (ROI plan ) structures, as well as the CT plan , were exported to the research platform.
Before starting fiber tracking, the DTI data after surgical removal of the tumor (DTI plan ) and the MP-rage of the recurrence (MRI recurrence ) were registered rigidly on CT plan by a radiation oncologist in iPlan RT Image 4.1.2 (Brainlab, Munich, Germany). The recurrence was also defined by a physician on CT plan based on MRI recurrence . The recurrence structure (ROI recurrence ) was exported. Then, CT plan was rigidly registered to DTI plan using the open source software Plastimatch (www.plastimatch.org) to get the transformation matrix (M t ). M t was applied on the ROI plan . As ROI recurrence was defined on CT plan for initial tumor treatment, it was also propagated to DTI plan by M t .

DTI processing, DTI analysis and fiber tracking
iPlan. Streamline tractography was conducted using the fiber tracking module of iPlan Image 4.1.2 (Brainlab AG, Munich, Germany). The DTI images were imported into iPlan and corrected for eddy currents [13]. iPlan uses the tensor deflection (TEND) fiber tracking algorithm with fractional anisotropy (FA) and value of minimum fiber length as stopping criteria [13]. We set the criteria to FA = 0.3 (recommendation made in a personal communication by Brainlab staff) and a minimum length of 50 mm. Three different seed regions were used: (i) GTV enlarged by 2 mm (seed GTV +2mm ), (ii) GTV enlarged by 4 mm (seed GTV +4mm ), (iii) recurrence volume (seed b). If more than one recurrence volume was present, the recurrence volumes were investigated separately. It was evaluated whether a connection between extended GTV volume and recurrence volume existed. If one or more of the tracked fibers entered both volumes, a connection was assumed to be present. Additionally, the strength of the connection was rated subjectively to be weak, middle, or strong.
MITK. Treatment planning CT (CT plan ), initial GTV, recurrence structure (ROI recurrence ) and the DTI (DTI plan ) were co-registered and imported into the open source toolkit MITK-Diffusion (www.mitk.org). A discrete Gaussian noise filter with variance of 2 arbitrary units (a.u) was applied to the DTI raw data. Afterwards, a tensor map was generated via the module "Tensor" with a threshold of 50 a.u. on the b0 image to eliminate regions outside of the brain.
MITK Diffusion provides the opportunity of Gibbs tracking [14] and streamline tracking [11,[15][16][17][18][19]. The fibers were generated by streamline tracking with a FA threshold of 0.3 and the fiber assignment by continuous tracking (FACT) algorithm (MITK parameters f = 1 and g = 0, see also [17]) with minimal tract length of 50 mm. The same seed volumes as in iPlan were used.
In contrast to streamline tracking, Gibbs tracking is a global method for generating all fibers in the brain. Fibers are found by minimizing a cost function that guarantees consistency with the measured data and smoothness of fibers. This approach allows fibers to traverse fiber crossing regions. In contrast, using streamline techniques often results in premature fiber terminations in crossing regions because these regions are associated with small FA values in DTI. Gibbs tracking is able to overcome this limitation and can thus potentially unveil connections that are not found with streamline tracking.
Gibbs tracking in MITK features several parameters which can be set individually. The parameter particle length, particle width, particle weight and random seed were set to "auto". Start Temperature was 0.1 a.u. and End Temperature was 0.001 a.u. The parameter balance of In/Ex Energy was set to 0. The parameter curvature threshold was 45˚and the minimum fiber length was 50 mm. 10 8 iterations were calculated.
After generating fibers by streamline and Gibbs tracking, they were extracted with help of the MITK "Fiber extraction" tool to find the fibers that pass the volumes of interest. I.e., the fibers generated by streamline tracking with seed GTV + enlargement were extracted for passing the recurrence volume. The fibers generated by seed b were extracted for passing GTV +2mm or GTV +4mm . For Gibbs tracking, those fibers were extracted that passed GTV + enlargement and recurrence volume. The strength of the connection was rated subjectively in consensus by BW and FBL to be weak, middle, or strong. The rating "weak" was given if only a few connecting fibers were present, which subjectively could easily be attributed to origin from tracking errors. The rating "strong" was given if a doubtless connection was present. The rating "middle" was given otherwise. Fig 1 shows the fiber tracking results obtained with iPlan with seed GTV +4mm . For patient 6, a strong connection between seed volume and recurrence volume 1 (displayed in cyan color) is visible. However, in all other cases, either only a weak connection is present, where the fibers are only touching the rim of the recurrence region (e.g. for patient 3) or no connection is present (e.g. for patient 2). Fig 2 shows the fiber tracking results obtained with iPlan with different seed volumes for patient 6. Concerning the volume of the second recurrence displayed in purple, no connection is visible for seed GTV +4mm (Fig 2a). However, a connection is present for seed b (purple volume, Fig 2b); i.e. it matters whether the tractography is started in the GTV or in the recurrence volume. showing all fibers that enter or pass the GTV +4mm . In comparison to the other two tracking approaches, much more fibers are generated often penetrating most of the brain volume. Consequently most recurrences feature some connection to the GTV +4mm . Fig 5 also shows Gibbs tracking results. Unlike in Fig 4, only those fibers are shown that enter or pass the GTV +4mm and the recurrence volume. Thus, much less fibers are shown than in Fig 4. Still, a strong connection is again visible for patient 6. However, unlike in Figs 1 and 3, patients 1, 5, and 7 also have strong connections between GTV +4mm and recurrence volumes. Table 1 shows the quantitative and subjective evaluation of strength of the connection between GTV +2mm (upper part) and GTV +4mm (lower part), and the recurrences. For all recurrences, existence and strength of a detected connection strongly depended on the used algorithm and target volumes.

Discussion
Of the 152 considered patients, seven had a distant recurrence. For one of these patients (patient 6), all used tracking approaches with seed GTV +4mm found a strong connection between the GTV and one of the recurrences (recurrence 1). For the second recurrence of The results indicate that the usefulness of fiber tracking methods for inclusion of distant recurrences in the definition of the PTV is rather limited for GBM patients. We used three different tracking approaches and observed the well-known strong dependency of tracking results on the used algorithm [20][21][22][23]. It may be argued that the variance of obtained results may be minimized by properly harmonizing the whole pipeline from acquisition protocol, seed volume definition, tracking method and tracking criteria, to the evaluation of the results. But the problem that the position of the recurrences was not predictable with our tracking data was present for each of the tested methods. For the two streamline tracking techniques, connections were mostly not present or only very weak. The Gibbs tracking approach yielded clear connection for more patients, which were, however, only identifiable retrospectively ( Fig  5). At the time point of PTV delineation, only the GTV is known (as in Fig 4). However, the Gibbs tracking based on the GTV resulted in connections to almost all brain regions, which limits its predictive value. Thus, it seems unlikely that an optimization of parameters or the use of more advanced tracking algorithms-like machine learning techniques [15]-might help in this regard.
The reason for this negative finding is presumably twofold. First, it seems that a connection via large white matter fiber bundles is not a necessity for lesion spreading (as, for example, for patient 4). Second, the lesion may occur at quite distant position with respect to the primary tumor like in patient 2. Even if all distant lesions were connected to the primary tumor, knowing the fiber structure would not help for treatment planning because the lesions could still appear basically anywhere in the brain. This makes a well-conducted radiation dose boosting extremely challenging.
The results of Krishnan et al. were more promising than ours in terms of the predictive value of fiber tracking [12]. They included several glioma types with different therapies including only three patients post resection out of 14 patients in total, while we focused on GBMs after resection. This is, however, unlikely to explain the observed difference, since two out of three GBM patients had a strong connection in [12] for the distant secondary tumor group, which comes closest to our patient population. It seems more likely that the low number of patients in the two studies (14 in [12] and 7 in our study) and the associated large statistical uncertainty is the major reason for the observed differences.
DTI data have been used successfully in the context of glioma evaluation in several regards. For example, they have been used to adapt target volume definitions for GBM treatment  planning [24][25][26], for modeling of brain tumor growth [27][28][29][30], to detect early malignant transformation of low-grade glioma [31], to differentiate between GBM and brain metastases [32], to use DTI-derived fiber tracking in surgery planning [33], and to detect the infiltration of the corpus callosum [34]. Our inability to properly connect most primary and secondary tumors does not diminish the value of DTI for these applications.
Potentially, other MRI techniques such as resting state functional MRI might represent potential venues for future research directions to relate the functional connectivity of the brain in GBM patients with the occurrence of new lesions [35].
One limitation of this study is the small patient number (seven patients with ten recurrences), which may be explained by the rareness of GBM and the low percentage of distant recurrences in GBM. Because of the surgical removal, a cavity occurred, which most likely resulted in a shift of brain anatomy [36][37][38]. Several physicians were involved in the definition of GTV and PTV treatment. If only one physician had planned GTV and PTV, the associated well-known inter-user variance could have been avoided [39]. One further limitation was the limited available scan time. Using 30 or more gradient directions would have stabilized the diffusion tensor calculation [40]. Advanced diffusion techniques that use higher b-values and generally require more scan time might also have improved the tractography results. These techniques include e.g. spherical deconvolution [41], neurite orientation dispersion diffusion imaging (NODDI) [42], higher order tensor reconstructions [43], or q-ball imaging [44]. The use of a higher field strength of 3 T or even 7 T would have been beneficial with respect to the signal to noise ratio and thus might have improved the tracking results [45].

Conclusion
The study investigated the use of DTI based fiber tracking to predict the position of potential recurrences in GBM patients, which might be helpful for the definition of the PTV. The observed fibers were dependent on the used tracking algorithm and most of the recurrences could not be properly connected to the GTV in a prospective, i.e. predictive, manner. Based on the results of this study, it seems not recommendable to adapt the PTV definition for GBM patients including potential distance recurrences with fiber tracking based on the clinical DTI data acquisition techniques described in this study. Observed connections between GTV extended by 2 mm / 4 mm and recurrence volume considering fibers generated by streamline algorithm in iPlan and MITK and by the Gibbs tracking algorithm in MITK with subjective rating of the strength of connection (weak, middle, strong). For the Gibbs tracking algorithm, the number of fibers passing GTV +2mm / GTV +4mm and recurrence is stated. https://doi.org/10.1371/journal.pone.0227146.t001