Discussion

In this study, we retrospectively analyzed the distances between the GTV and tumor recurrence in 201 patients with glioblastoma. Recurrence distance analysis demonstrated that 80% of the recurrences occurred within 10 mm from the GTV. Furthermore, GTV volume and involvement of the SVZ and SGZ were significantly associated with the distance between the GTV and tumor recurrence.

Recurrence distance analysis showed that the majority of recurrences in our cohort were located within or in close proximity to the GTV. For 80% of the patients, the RD80 was ≤10.4 mm. Notably, the RD80 represents a relatively strict metric compared to a pattern of failure classification based on the 95% isodose line. Although we did not generate theoretical radiotherapy plans with a 10 mm CTV-margin, it is reasonable to hypothesize that a pattern of failure analysis using theoretical radiotherapy plans with a 10 mm CTV-margin would have resulted in more than 80% of patients having an in-field recurrence in our cohort. This estimate falls within the range of previously published retrospective data on pattern of failure in patients treated with a 10 mm CTV-margin [7], [8], [9]. However, recurrence distance analysis and pattern of failure analysis based on the high-dose radiation treatment field are not directly interchangeable. Hence, while our work demonstrates that the majority of recurrences are located in close proximity to the GTV, prospective evaluation in clinical trials remains necessary to validate the feasibility of reduced CTV-margins.

In our cohort, smaller GTV volumes were significantly associated with greater distances between the GTV and the recurrence volume. One possible explanation for this association is that smaller tumors exhibit lower hypoxia burden [29]. As a consequence, radiotherapy may have a better local efficacy [30] in patients with smaller GTVs, resulting in a higher likelihood of developing a distant recurrence. Another plausible reason is that the surrounding brain tissue for tumors with larger GTVs becomes more limited, potentially constraining the distance over which tumor infiltration can migrate [31]. It is important to acknowledge that methodological and geometric aspects may also contribute to the observed association. The RD80 represents a specific distance from the outer edge of the GTV; hence, it may be influenced by the size and shape of the GTV. The observation that smaller GTV volumes had a higher rate of distant recurrence, is consistent with the results from Langhans et al. [31], who raised the possibility that a CTV-margin depending on GTV size may allow similar (or even improved) tumor control while having less side effects.

Furthermore, contact with a neurogenic zone was a significant predictor of a greater recurrence distance in our cohort, consistent with results from prior studies. Tumors contacting the SVZ have repeatedly been associated with atypical behavior, including a higher likelihood of distant recurrence, multifocal disease, and worse prognosis [14], [15], [16], [17], [25], [26], [27], [32], [33]. Chen et al. [28] further demonstrated that tumors contacting the SVZ or SGZ had a higher incidence of distant recurrence outside the high-dose radiation treatment field. The biological mechanism linking contact with a neurogenic zone and recurrence distance remains unclear. It has been hypothesized that tumor cells in neurogenic zones may co-opt the microenvironment and existing migratory pathways of these regions, potentially enhancing their ability to migrate throughout the brain [26], [34], [35]. Investigating the mechanisms underlying the observed association was beyond the scope of this study and future research aimed at elucidating these mechanisms may provide valuable insights into the biological processes driving glioblastoma recurrence patterns. Interestingly, although involvement of the SGZ was not significantly correlated with the log-transformed RD80 in our univariable analysis, it emerged as a significant predictor in the final multiple linear regression model. This change may be explained by confounding. As can be observed in Supplementary Fig. S3c, tumors with SGZ involvement tend to have larger GTV volumes. Since our results also indicate that larger GTV volumes are associated with a shorter RD80, the association between SGZ involvement and RD80 may have been masked by differences in GTV volume. In the multiple linear regression model, this association became apparent after adjusting for differences in GTV volume.

Although our final multiple linear regression model was able to identify significant variables associated with recurrence distance, the model only explained 12% of the variance. This low percentage indicates that it is challenging to reliably predict the recurrence distance with linear regression and the variables that we examined. Nonetheless, the significant variables found in our work may offer a potential direction for personalized CTV-margins.

In our sensitivity analyses, the number of completed adjuvant chemotherapy cycles prior to PD and time to progression were independently associated with the RD80 when added separately to the final RD80 regression model. The latter finding is in line with previous reports [36], [37]. However, when both variables were added simultaneously, only time to progression remained independently associated, potentially reflecting a positive correlation between time to progression and the number of adjuvant chemotherapy cycles (see Supplementary Fig. S4e). Hence, the association between the number of adjuvant chemotherapy cycles and RD80 may partially be reflecting underlying differences in disease course. Further research on the potential impact of adjuvant chemotherapy on recurrence patterns may therefore be of interest. Importantly, the significant predictors in our final model remained unchanged after adjustment for variations in treatment, supporting the robustness of the main findings.

In this study, we introduced the RD80 as a novel metric to provide additional quantitative information for pattern of failure analysis in glioblastoma. The 80% coverage threshold for the RD80 is not based on a clearly defined biological rationale. Instead, it was chosen to maintain consistency with existing pattern of failure analyses based on the high-dose radiation treatment field, where recurrences are typically classified as in-field when at least 80% of the recurrence volume is located within the 95% isodose line [7], [22], [23], [24]. Lee et al. [38] proposed this cutoff based on an evaluation of the relationship between dose-volume histogram data and the recurrence volume, theorizing that recurrences with ≥80% of their volume located within the 95% isodose line most likely originated within the high-dose radiation treatment field, and subsequently expanded outward. The significant predictors of the final multiple linear regression model remained unchanged when we performed the regression analysis using the RD95.

There were several limitations in our study. First, our recurrence distance analysis was designed to explore the potential for reduced CTV-margins. However, this approach does not take into account how recurrence patterns might have changed if a different CTV-margin was actually employed. Nevertheless, our work provides a valuable hypothetical framework for optimizing CTV-margins, laying the groundwork for further prospective exploration. Second, this study only included patients with IDH-wildtype glioblastoma who subsequently developed radiologically confirmed recurrence on follow-up MRI. While this may have introduced selection bias, these criteria were required to allow for recurrence distance analysis in a homogeneous cohort in accordance with the most recent 2021 WHO classification [12]. Third, data on the MGMT methylation status was unknown in 29% of patients, potentially decreasing statistical power to find an effect of MGMT methylation status on the RD80 and introducing a selection bias. A proportion of missing data may be attributed to changes in routine molecular diagnostics between 2012 and 2022. Some studies have suggested that MGMT methylation in glioblastoma may be associated with distant recurrences [14], [22], [39]. Data on MGMT methylation status was still available for 143 patients, but did not show a significant association with recurrence distance in our univariable regression. It is also important to acknowledge that we did not include other alterations that may impact glioblastoma recurrence, like EGFR amplification or TERT mutation. Fourth, variability across institutions warrants careful generalizability of our findings. Specifically, our single-institution retrospective design does not account for different contouring philosophies across centers. As a result, our findings may not be directly generalizable to institutions that adopt the cone-down approach outlined in the 2025 ASTRO Clinical Practice Guideline [40]. There were also variations in target delineation within our cohort. While the GTV was typically defined as the resection cavity plus residual contrast-enhancing tumor, deviations from this definition may have occurred in individual cases at the discretion of the treating radiation oncologist (e.g. for non-enhancing tumors). Additionally, 34% of our patients were irradiated with a 20 mm CTV-margin, which could have had an effect on the site of tumor recurrence. Di Perri et al. [41] reported similar pattern of failure and progression-free survival in patients treated with a 20 mm CTV-margin compared to patients treated with a 15 mm CTV-margin, implying the effect may be inconsequential. To obtain further evidence to this effect, we performed a sensitivity analysis confirming that the difference in treated CTV-margin did not significantly impact our results. Fifth, as tumor progression can be accompanied with mass effect and midline shift, registration of the MRI-scan at the time of recurrence to the radiotherapy planning CT can be challenging, and introduce an inaccuracy when calculating the RD80. We attempted to minimize these effects by using the MRI-scan with first radiological evidence of recurrence for tumor recurrence delineation. Sixth, we performed multiple linear regression modeling with a backward elimination procedure. This approach may be sensitive to sample-specific variations, potentially leading to model instability. Therefore, the results from our regression analysis should be interpreted with appropriate caution. Finally, we computed the RD80 using the Euclidean distance, which does not take anatomical barriers into account. While it may introduce an error in the estimation of the RD80, we do not expect a substantial impact on our analysis. From Fig. 2, we can derive that the proportion of patients with an RD80 ≤ 15 mm (∼86%) closely matches the proportion of patients with an in-field recurrence (87%) in our cohort, indicating that our approach provides a reasonable estimate.

Our work provides an incentive for prospective clinical trials to assess the effect of reduced CTV-margins on pattern of failure, progression-free and overall survival, and quality of life. While personalized CTV-margins could theoretically benefit patients, our results suggest that the realization of this concept is challenging. Advanced prediction methods, e.g. radiomics [42], may better capture the complex spatial and biological characteristics that drive glioblastoma recurrence patterns. Another approach for more optimal CTV delineation of glioblastoma was introduced by Detsky et al. [43], who demonstrated that MRI-guided adaptive radiotherapy may enable a reduction of the CTV-margin without compromising tumor control. In their study, patients with glioblastoma were treated with a 1.5 Tesla MRI-guided linear accelerator with weekly online adaptation and a 5-mm CTV-margin, resulting in a low rate of marginal failure. Furthermore, future strategies could leverage the preferential infiltration patterns of glioblastoma along white matter tracts and perivascular spaces [44], [45], [46], and shift from isotropic CTV-margins toward patient-specific anisotropic CTVs that more accurately reflect the direction of tumor infiltration. This approach may further benefit from integrating multimodal imaging, like advanced MRI [47] or amino acid PET [48], [49], to enable more accurate target delineation of glioblastoma.

In conclusion, we present the distribution of recurrence distances and significant variables impacting recurrence distance in a large cohort of patients with glioblastoma. Our work demonstrates that the majority of recurrences are located within or in close proximity to the GTV. Prospective evaluation remains necessary to assess the impact of reducing the CTV-margin for patients with glioblastoma.