Spatial distribution of malignant transformation in patients with low-grade glioma

A radiological transformation was considered in the event of a new significant contrast enhancement where repeated scans, clinical course or histopathology from reoperation separated this from cases of pseudoprogression, radionecrosis or unspecific post-treatment changes due to for instance ischemia. To determine the spatial transformation, we relied on MRI findings. Thus, malignant histology from re-operation in the absence of enhancement was despite the transformation excluded from analyses of the spatial distribution of transformation since we had no reliable data on biopsy location. Also, other measures that could be taken as signs of malignant transformation prior to contrast enhancement, such as for instance FET-PET, was not used in this study [27]. Such information would presumably affect the timing of malignant transformation, but we believe to a lesser degree influence the spatial information.

In terms of spatial distribution, we used one method with clinical judgement (A.S.J) with visual inspection and crude one-dimensional measures, and one method based on tumor volume segmentations. A radiological transformation site < 2 cm from the tumor on high-intensity T2w and/or FLAIR signal abnormalities on the pre-transformation MRI scan defined local transformation. Distant malignant transformation was used if clearly separated (> 2 cm) from the high-signal abnormalities from the pre-transformation scan.

In the method with tumor volume segmentation, semi-automatic segmentations were performed using the open source medical imaging platform 3D Slicer (version 4.8.1, www.​slicer.​org). A radiologist (H.K.B) performed preoperative segmentations and the follow-up segmentations were performed by a neurosurgeon with extensive experience in radiological LGG assessment (A.S.J). First, we segmented the preoperative tumor volume using T2 or FLAIR MRI sequences. Next, we identified the scan where malignant transformation (i.e. new contrast enhancement) was detected and segmented the contrast enhancement using T1 with gadolinium enhancement. Then, we segmented the T2 or FLAIR volume from the scan prior to the scan where malignant transformation was detected. Together, these segmentations built the fundament to the processing pipeline as described below.

Key MRI scans being the pre-operative, pre-transformation, and transformation timepoints were selected, and to determine the relative locations of transformation sites they needed to appear in the same referential space. A processing pipeline, illustrated in Fig. 1, was therefore developed to generate the results for each patient. As input, the pair of original images (i.e., pre-operative, pre-transformation, and transformation) and corresponding ground truth volumes (original tumor, pre-transformation, transformation) were used, shown as step 1 in Fig. 1. Then, the brain was automatically segmented using a neural network model pre-trained with over 300 samples, shown as step 2 in Fig. 1. The neural network follows a U-Net architecture and has been implemented in Python using Keras and TensorFlow [28]. Using the brain segmentation, the skull was stripped before performing registration of the pre-transformation and transformation images (as long as the corresponding ground truth volumes) toward the pre-operative MRI volume. This process, illustrated as step 3 in Fig. 1, was performed using a symmetric diffeomorphic technique (named SyN) from the Advanced Normalization Tools [29]. In the end, the three volumes of interest were displayed in an overlap fashion over each MRI volumes, all expressed in the pre-operative MRI space, represented by step 4 in Fig. 1.

In this study we provide only descriptive statistics. This was chosen since focus was to describe patterns of transformation.

Studies indicate an association between tumor size or size of tumor remnant and earlier malignant transformation [7, 30‐34]. The dose–response relationship of tumor size and malignant transformation in astrocytomas and oligoastrocytomas was demonstrated by Shaw et al. where 28% recurred with < 1 cm of remnant, 88% where remnant was 1–2 cm and 100% if > 2 cm remnant [33]. For oligodendrogliomas, the dose response relationship was also present, but to lesser extent with 23%, 43% and 75% depending on largest diameter of remnant. Of note, in the study from Shaw et al. the definition of progression was a clear increase in T2/FLAIR or contrast enhancement, and this differs from our definition focusing on the detrimental event of malignant transformation. Others have demonstrated that oligodendrogliomas are more likely to progress without malignant transformation [30, 35]. Since the existing literature is a mix of progression and transformation, the numbers would likely differ even more between astrocytomas and oligodendrogliomas if only malignant transformation was analyzed.

In the surgical literature several studies have reported that remnant less than 10–15 ml have a better prognosis [36, 37] although no visible postoperative remnant is clearly superior [38‐40]. In our series, smaller residual tumor volumes than 10 ml did not protect against transformation although it was seen in only 7% of cases. Also, in none of the patients transformation occurred without any preceding T2 hyperintensities, albeit in almost 10% of cases the transformations were distant and seemingly unrelated to earlier T2w hyperintensities. This association with MRI defined volume and transformation is further corroborated by a study that focused on recurrent surgery of a previous LGG, where in stable lesions being WHO grade II the median volume was 15.6 ml while for tumors with transformations to WHO grade III and grade IV the volumes were 30.9 ml and 69.7 ml volume respectively [34].

In perceived low-risk patients undergoing surgical resection with aim of gross-total removal, and where no adjuvant therapy was provided, the spatial distribution of progression was found to be within 2 cm from the resection cavity in 82% of cases, more than 2 cm away in 16% and 2% had truly distant progression [33]. In our cohort that was not a typical low-risk profile and being based partly upon historical data where biopsies where frequently performed [25], we observed a higher proportion of local progressions and transformations. In low-risk patients undergoing extensive resection the pattern of transformation may be somewhat different, as observed by others that more aggressive therapy leave room for more atypical and distant progression and transformation patterns in the longer-term [19]. We included only patients with radiological transformations, excluding patients with transformation based upon histopathology alone from reoperation without preceding change in MR phenotype. These surgical transformations are also local in origin, hence the vast majority of patients with LGG will present a local transformation. Overall, the patterns of distributions together with the volume associations mentioned above conceptually favor aggressive locoregional therapy, and presumably also repeated surgery whenever possible.

An early study from North et al. demonstrated that all treatment failures were within the radiation field when 2–3 cm margin was used [21]. Similarly, Shaw et al. described that all failures were within the radiation field [24]. Also, survival was not improved by more extensive fields such as whole brain radiotherapy [24]. We observed slightly more distant recurrences in the subgroup treated with radiotherapy prior to malignant transformation. This may be a survivor effect if radiotherapy to some degree prevents local malignant transformation, and since there is no cure, transformations will therefore eventually more often be distant. A recent study in patients with anaplastic gliomas treated with intensity-modulated radiation therapy suggested that radiotherapy may prevent local recurrence. In that study, a relapse pattern with components of marginal and distant pattern were observed in 19% and 37%, respectively [41]. In our view, the majority of studies indicate that the majority of progression and transformation follow a local pattern, and this can argue for proton-beam radiotherapy in patients with LGG. However, to date there is very limited clinical evidence [42]. It has been speculated if the more conformal field would create a risk for more distant recurrences. However, one very recent larger retrospective study found that most recurrences following proton beam-radiotherapy were indeed local, with only 12% being “out of field”, a comparable figure to our “distant” transformations [43]. Thus, the “dose-bath” beyond the targeted areas are perhaps not needed since most recurrences and transformations are local. That larger areas of the presumably functional brain more often receives no radiation based upon comparative proton plans, means also less risk in the longer-term for cognitive decline. This is potentially important as we otherwise can transform long-term treatment successes in terms of survival to long-term failures speaking of cognitive function and quality of life.

Previous studies of malignant transformation have not used the WHO 2016 classification, and this may affect results. One recent study on failures following radiotherapy in anaplastic gliomas demonstrated a distant pattern of failure in 45% of IDH mutated patients compared to 25% in those with IDHwt [41]. In our study we had > 1/3 with IDH mutated astrocytomas and almost 30% molecularly defined oligodendrogliomas, however we did not find any differences in patterns of recurrence with respect to the molecular profile of the tumors. However, across subgroups the most common pattern of transformation was local.

Ideally all areas should have been sampled to verify malignant transformation, although either new histopathology or the clinical course ensured that only patients with malignant transformation were included (and not pseudoprogression). The T2/FLAIR volume at time of progression was segmented to illustrate the growth of the tumor and use this in relation to the newly developed contrast enhancement to demonstrate the tumor evolution. However, in some cases the T2w images showed the occurrence of gliosis following surgery and hyperintensities in patients undergoing radiotherapy. In these cases, this volume is associated with inherent uncertainty. Also, even the intra-observer variability in LGG segmentations can be significant [44]. The overlay segmentation model also holds some limitations when comparing with the preoperative volumes in patients undergoing resection, as the cavity may shrink/collapse causing areas outside the cavity to appear within the cavity perhaps increasing the proportion of recurrences within the preoperative tumor volume. Further, this study is not equipped to answer effectiveness of therapies and the sample size did not allow for comparisons for time to transformation. Also, the small sample limits the subgroup analyses. Finally, the centroid model can provide erroneous results in multifocal tumors.