Improved dosimetric accuracy with semi‐automatic contour propagation of organs‐at‐risk in glioblastoma patients undergoing chemoradiation

Abstract Background We study the changes in organs‐at‐risk (OARs) morphology as contoured on serial MRIs during chemoradiation therapy (CRT) of glioblastoma (GBM). The dosimetric implication of assuming non‐deformable OAR changes and the accuracy and feasibility of semi‐automatic OAR contour propagation are investigated. Methods Fourteen GBM patients who were treated with adjuvant CRT for GBM prospectively underwent MRIs on fractions 0 (i.e., planning), 10, 20, and 1 month post last fraction of CRT. Three sets of OAR contours — (a) manual, (b) rigidly registered (static), and (c) semi‐automatically propagated — were compared using Dice similarity coefficient (DSC) and Hausdorff distance (HD). Dosimetric impact was determined by comparing the minimum dose to the 0.03 cc receiving the highest dose (D0.03 cc) on a clinically approved reference, non‐adapted radiation therapy plan. Results The DSC between the manual contours and the static contours decreased significantly over time (fraction 10: [mean ± 1 SD] 0.78 ± 0.17, post 1 month: 0.76 ± 0.17, P = 0.02) while the HD (P = 0.74) and the difference in D0.03cc did not change significantly (P = 0.51). Using the manual contours as reference, compared to static contours, propagated contours have a significantly higher DSC (propagated: [mean ± 1 SD] 0.81 ± 0.15, static: 0.77 ± 0.17, P < 0.001), lower HD (propagated: 3.77 ± 1.8 mm, static: 3.96 ± 1.6 mm, P = 0.002), and a significantly lower absolute difference in D0.03cc (propagated: 101 ± 159 cGy, static: 136 ± 243 cGy, P = 0.019). Conclusions Nonrigid changes in OARs over time lead to different maximum doses than planned. By using semi‐automatic OAR contour propagation, OARs are more accurately delineated on subsequent fractions, with corresponding improved accuracy of the reported dose to the OARs.


| INTRODUCTION
Glioblastoma (GBM) is the most frequent primary malignant brain tumor in adults with an overall survival of 5.1% at 5 years from time of diagnosis despite aggressive therapy. In appropriate patients, current standard of care is maximal safe resection, followed by adjuvant concurrent chemoradiation therapy (CRT) to a total of 4000-6000 cGy. 1,2 During the 4-6 weeks between surgery and the initiation of adjuvant CRT, approximately 50% of patients develop tumor growth or changes in the resection cavity contrast enhancement pattern. 3 In a computed tomography (CT)-based imaging study, the median resection volume reduction was approximately 35% at week 4 of treatment. 4,5 Definition of the tumor and surrounding organs-at-risk (OARs) can be better visualized with magnetic resonance imaging (MRI) and OAR contours for plan adaptation each day, need to be addressed before the MRL can be used clinically. 6 There can also be considerable inter-observer variability in the contours for different OARs. For example, one study suggests that the inter-observer Dice similarity coefficient (DSC) for the brainstem is 0.83 and the DSC for the optic nerves is 0. 5. 7 The purpose of this study is to investigate the changes that occur in GBM OARs on serial prospectively acquired MRIs during adjuvant CRT. We investigate the accuracy and dosimetric implication of assuming non-deformable OAR changes, in addition to the feasibility of semi-automatic contour propagation for the purpose of an adaptive workflow in MR-guided radiotherapy.

2.A | MRI schedule and parameters
Under an institutional research ethics board approved research protocol, between June 2016 and January 2017, patients who have undergone a surgical resection for GBM were recruited to undergo prospective serial multiparametric MRIs on fractions 0 (i.e., planning), 10, and 20 of CRT and 1 month post last radiotherapy fraction. All images were acquired on a 3T Philips Achieva scanner (Philips, Best, Netherlands) and we studied the T1-weighted post-Gadolinium MRI, which consisted of a full 3D acquisition using the Philips Fast Field Echo (FFE) gradient echo sequence (TR = 9.5 ms, TE = 2.3 ms) with a voxel size of 0.49 × 0.49 × 1.50 mm.

2.B | OAR contouring
OARs including the brainstem, globes, optic nerves, and chiasm were manually contoured on each MRI by a senior radiation oncology resident (SLL) and verified by a board-certified radiation oncologist (CLT).
These contours served as a ground truth for contour evaluation.
Manually contoured OARs from the planning scan were propagated to the MRIs of subsequent fractions using a deformable registration algorithm through ADMIRE software version v2.0.0.1, which was run from a research version of the Monaco treatment planning system version v5. 19.03 (Elekta AB, Stockholm, Sweden). The current clinical version of Monaco in use on the MRI-linac uses a deformable image registration algorithm that is identical to the atlas-based algorithm used in the present study. The method simulated a semi-automatic workflow in which the contours for the current time point are automatically segmented using the manual contours from prior time point as the reference atlas. Manual contours were propagated to the fraction 10 MRI with a single atlas segmentation method using the fraction 0 manual contours as the reference atlas. The segmentation involved several steps including linear registration, poly-smooth nonlinear registration, and dense hybrid deformable registration as outlined in work by Han et al. 8 The deformably registered contours were evaluated for accuracy by comparing to the manual contours. Similarly, the fraction 10 manual contours were used as a reference atlas to automatically segment the OARs on the fraction 20 MRI and the fraction 20 manual contours were used to automatically segment the OARs on the post 1 month MRI (Fig. 1). The proposed propagation scheme reflects an anticipated workflow with the MRL in which the contours are propagated each day and are manually corrected before the radiation plan is reoptimized and delivered.

2.C | Static contours vs manual contours
In order to assess the presence and degree of interfractional deformation in OARs during treatment, the manual contours of each fraction were compared to the manual contours of fraction 0. To make the comparison, the fraction 0 MRI was rigidly coregistered to the MRI of each subsequent time point and the resulting overlap between the manual contours and fraction 0 "static" contours was assessed using the Dice similarity coefficient (DSC) and Hausdorff distance (HD). 9 For two three-dimensional regions, A and B, the DSC measures the degree of overlap and is defined as where jAj is the volume of region A, and jA∩Bj is the volume of the with perfectly overlapping contours giving HD = 0.

2.D | Propagated contours vs manual contours
Similarly, the propagated contours of each time point were compared to the corresponding manual contours using DSC and HD

2.E | Comparison of static contours vs propagated contours
To compare the performance of the static contours vs the propagated contours, the manual contours were used as a ground truth.
The DSC, HD, and ΔD0.03 cc between the static contours vs manual contours were compared to the DSC, HD, and ΔD0.03 cc between the propagated contours vs manual contours.

2.F | Statistical analysis
Statistical significance on the differences observed was determined using the Wilcoxon signed-rank test, with significance level defined as P < 0.05. All statistical analyses were done on MATLAB and Statistics Toolbox (Release 2015, The MathWorks Inc., Natick, United States).

| RESULTS
Serial MRIs were prospectively obtained on 14 recruited patients.

3.A Static contours vs manual contours:
The DSC between the static fraction 0 contours vs the manual contours from each fraction varied across all time points and structures with mean ± 1 SD of 0.77 ± 0.17. DSC for each structure across all patients and time points are summarized in Table 2 (Fig. 2).

| DISCUSSION
In the present study, we found significant nonrigid changes in OARs in GBM patients on prospectively acquired serial MRIs during and after adjuvant CRT that lead to different maximum doses than planned. By using semi-automatic OAR contour propagation, the OARs were more accurately delineated, which increased the accuracy of the reported dose to the OARs.
Using manual contours as a ground truth, the static contours had Some of the results can be explained by the type of metric used.

DSC is a reliable evaluation method for volumetric segmentations
where high-quality contours and a high degree of overall agreement is expected. HD is sensitive to outliers and is more reliable when the contours are small and overlaps are small. 9 The results showed a greater number of significant differences in DSC between static and propagated contours compared to the number of significant differences using the HD metric as the OARs compared typically had a high degree of overlap. As shown in Fig. 4, for organs that had a high degree of overlap, such as the eye, the separation of the DSC between the static and propagated contours over time was clear.
However, for organs that had a lower degree of overlap and greater uncertainty for the contour edges such as the optic nerve, the separation of the DSC between static and propagated contours over time was less obvious.
Delineation of OARs in the brain is detailed in guidelines such as those by Scoccianti et al and Niyazi et al. 13,14 However, despite such guidelines, inter-observer variation in the delineation of OARs in the T A B L E 2 Summary of performance comparisons between static fraction 0 contours vs propagated contours.  brain is significant. 15,16 The agreement between manual ground truth contours and both the static and propagated contours in this work was greater than the inter-observer agreement in the literature. For example, the static contour mean DSC for the brainstem was 0.93 and for the optic nerve was 0.6 compared to the reported inter-observer DSC of 0.83 and 0.5, respectively. 7 Intra-observer variability may also contribute to uncertainty of OAR delineation, with intra-observer variability of 20% being quoted in brain tumor contouring. 17 Both the intra-and inter-observer OAR contour variation as well as the change in OARs over time results in a significant variation in the maximum dose estimates to the OARs. In previous studies, the inter-observer variability contributed to a range of maximum doses to the optic apparatus equivalent to 70% of the prescribed dose in stereotactic radiosurgery. 15,18 In our study, the range of dose differences to the optic chiasm ranged from −1500 to 1700 cGy for a 6000 cGy plan, corresponding to a variability range of 53% of the prescribed dose. The wide variation in maximum dose in our study can be explained by the high-dose gradients surrounding the OARs.
Typical gradients in the high-dose fall-off region were 300-400 cGy/ mm. A second contributing factor was the variability in the contouring of the chiasm, where it demonstrated the highest average HD compared to other OARs, with a mean of 5.9 mm for the propagated contours. Similarly, in a study for oropharyngeal carcinoma, inter-observer contour variability contributed to a maximum dose increase of 23% to the brainstem. 16 In our study, the greatest change in maximum dose to the brainstem was 400 cGy (7% of the maximum dose of 5400 cGy). Therefore, the variations in dose between the propagated contours and the ground truth observed in our study were smaller than the variations in dose due to inter-observer variability studied in published literature.

Use of automatic tumor and OAR contouring can help reduce
inter-and intra-observer variability. Several algorithms can be used to automatically contour the OARs including atlas-based methods, statistical models, and deformable models. 19 The use of automatic tumor volume segmentation is an active area of research. Studies evaluating automated segmentation of the OARs and target volumes have shown that although automated segmentation can reduce the amount of time required to contour, manual editing of automated contours is still required. 20,21 In this study, we show the feasibility of semi-automated contour propagation, using deformable registration of the manual contours from the prior time point. Using manually corrected semi-automatically propagated contours is similar to our daily clinical workflow in which a radiation oncologist reviews and corrects the work of a resident, which has been reported previously. 22 The ability to visualize and adapt radiation plans according to changing tumor and OAR volumes may result in greater therapeutic ratios. The majority of recurrences in GBM patients occur locally in the field treated with high-dose radiation. 23 The possibility of reducing local recurrences with dose escalation and concurrent temozolomide is being investigated in the phase 2 NRG BN001 trial. 24