Comparison of Volumetric and 2D Measurements and Longitudinal Trajectories in the Response Assessment of BRAF V600E-Mutant Pediatric Gliomas in the Pacific Pediatric Neuro-Oncology Consortium Clinical Trial

Trial Data and Measurements

We conducted a retrospective analysis of 31 participants (292 total follow-up scans) enrolled in the PNOC-002 trial, a phase I/II study of vemurafenib in BRAF V600E-mutant pediatric gliomas (Table 1).¹² Before trial enrollment, all participants had recurrence of tumor after surgery and at least one other therapy (chemo- or radiation therapy) and were not on any other treatment other than vemurafenib during the trial period. A board-certified neuroradiologist (M.S.A.) unblinded to follow-up scans performed 2D measurements on the pretreatment sequence (T2/FLAIR, T1 contrast-enhancing, or T1 pre-contrast) in a plane that best captured solid tumor, with cystic components included only if inseparable from solid components. The same sequence and plane were used to perform 2D measurements in all subsequent follow-up images to ensure consistency. T2/FLAIR sequences were used for most participants. However, in cases in which 2D measurements could not reliably and consistently capture tumor burden (eg, tumors with irregular or diffuse borders on T2/FLAIR) or the solid tumor portion was difficult to clearly visualize on T2/FLAIR, alternative sequences (ie, T1 or T1 contrast-enhancing) were used according to the neuroradiologist’s clinical judgment. The T1 contrast-enhancing sequence was only chosen in cases in which there was consistent tumor enhancement without transient or variable patterns of change. The Online Supplemental Data show the 2D measurements performed for each participant on the pretreatment image. According to RAPNO guidelines,⁵ the minimal bidimensional length of a lesion was 10 mm. Lesions that did not meet these criteria were not included in 2D or volumetric measurements.

3D segmentations of whole tumor (ie, including solid and cystic components) on FLAIR and cystic tumor on T2 were performed by a fourth-year medical student (D.R.) and corrected by the same board-certified neuroradiologist who performed the 2D measurements (M.S.A.). Segmentations were performed on a PACS-integrated annotation platform whereby a volumetric tool automatically extrapolated manual segmentations performed at the outermost tumor borders.¹⁴ These extrapolated segmentations were then manually adjusted slice by slice. Solid tumor volume was calculated by subtracting the cystic volume on T2 from the whole volume on FLAIR.

Comparison of 2D and Volumetrics with the Visual Assessment of Neuroradiologists

We compared 2D and volumetric RAPNO assessments with a reference standard, which we defined as the consensus score between 2 neuroradiologists’ visual assessments of the tumor response using the BT-RADS criteria.¹³ This comparison was performed for all participants at the end of treatment or last available on-treatment scan (for participants still on therapy at the time of analysis). BT-RADS is used in several clinical neuroradiology practices and represents the clinical interpretation of imaging findings with scores corresponding to definitive recommendations in clinical monitoring frequency and management.¹³ Standard and volume-extrapolated RAPNO criteria (Table 2) were applied to percentage change from pretreatment in the 2D area (sum of the product of maximal bidimensional diameters of all lesions) and whole volume, respectively. We used the pretreatment scan, which was also postsurgical, as the baseline.⁵ BT-RADS scores from a pretreatment baseline were assigned to each follow-up scan on the basis of mutual consensus between 2 board-certified neuroradiologists (M.S.A. and F.M.). We classified a BT-RADS score of 1a as partial response (PR), 1b to 3b as stable disease (SD), and 3c to 4 as progressive disease (PD) (Table 2). Paired log-rank tests were used to compare the time to the first response (confirmed on a consecutive scan at least 4 weeks later) between BT-RADS and each of the following criteria: 2D area change using the RAPNO minor response (MinR) threshold and whole and solid volume change using the volume-extrapolated RAPNO MinR threshold.

Empirical receiver operating characteristic (ROC) curves were plotted to compare the performance of 2D, solid-volume, and whole-volume change in classifying participants as having BT-RADS PR at 2 time points: 1) the end of treatment or last available on-treatment follow-up scan, and 2) the time of the BT-RADS first-confirmed response or first follow-up scan if no response. The DeLong test from the pROC¹⁵ package in R (https://www.rdocumentation.org/packages/pROC/versions/1.18.5) was used to compare whether the area under the curve (AUC) values were statistically different between measurement methods at both time points. A 95% confidence interval for the AUC was computed using 2000 stratified bootstrap replicates. All statistical analyses were performed in R studio, Version 2022.12.0 + 353 (http://rstudio.org/download/desktop).

Modeling 2D and Volumetric Longitudinal Treatment Response

We describe tumor growth mechanistically as the action and counteraction between treatment response and the onset of therapy resistance to parameterize the obtained tumor growth curves.^16,17 This model uses 3 patient-specific parameters characterizing the tumor growth rate (λ), treatment-induced tumor shrinkage (γ₀), and therapy resistance (∈). The tumor volume V is approximated as an exponential function in time at rate λ. In contrast, systemic therapy with vemurafenib is modeled to lead to an exponential reduction in tumor volume at rate γ: 1)

Given the onset of therapy resistance, we, in turn, describe the treatment efficacy as time-dependent with an exponential decay at rate ϵ: 2)

The analytical solution to this system of ordinary differential equations is 3)where γ₀ indicates the initial condition at therapy initiation at time t₀. V₀ is the tumor volume (or 2D area) assessed at the first scan. Patient data in either 2D or whole-volume tumor growth curves were fitted to Equation 3 using the python lmfit library (https://lmfit.github.io/lmfit-py/) in a bootstrap approach in which each data point was resampled n = 500 times from a normal distribution around the observed value (mean) with a 10% standard deviation as an estimate of random error stemming from contouring uncertainty. Given the choice of 3 patient-specific model parameters, only growth trajectories comprising at least 3 data points (ie, 2 follow-up scans) were included in this analysis.

Assessing Agreement between 2D and Volumetric Response Curves

From the model-obtained parametrizations, the estimated time to best response, ie, the minimal tumor volume (t_Vmin), was used for further analysis as a notion of the overall trend in the data: 4)

We assigned data points preceding V_min (including a 10% uncertainty window) as PR, those around (within a 10% uncertainty window) V_min as SD, and those beyond V_min as PD. We quantified the correlation between t_Vmin obtained from 2D and volumetric fits using the Spearman rank correlation coefficient ρ. The time to transition between response categories (ie, PR to SD or SD to PD) was also compared for all patients in whom such transitions were observed. RAPNO criteria require confirmation of response assessment on 2 consecutive scans,⁵ and the median time interval between 2 consecutive scans in this data set was 90 days. Therefore, we defined “transition agreement” as a difference in the time to transition between response categories of <90 days. We used the Wilcoxon rank-sum test to compare the following quantitative features between participants with agreement and disagreement in 2D-versus-whole-tumor trends: FLAIR sequence slice thickness, spacing, pretreatment whole and cystic volume, and the relative model fit performance of 2D-versus-whole-tumor curves quantified by normalized root mean squared error (nRMSE). The latter was interpreted as a proxy for uncertainty in measurement/segmentation. All statistical tests were performed using the scipy stats module in python (https://docs.scipy.org/doc/scipy/tutorial/stats.html).