Automated RANO 2.0-Aligned Response Assessment for Glioblastoma Clinical Trials: External Evaluation Against Expert...

Reliable and reproducible imaging-response assessment is essential for endpoint evaluation in glioblastoma clinical trials. The Response Assessment in Neuro-Oncology (RANO) criteria provide a rule-based framework for standardisation, but their implementation can still yield variable outputs between readers and across timepoints and trial sites. This partly reflects differences in lesion definition, measurement geometry, longitudinal tracking, and new-lesion adjudication. We externally evaluated an automated pipeline for auditable, RANO 2.0-aligned imaging-only response assessment using pre-existing segmentation masks.

This retrospective feasibility and pilot evaluation used longitudinal glioblastoma magnetic resonance imaging (MRI) studies from the public single-centre LUMIERE dataset. Dataset-provided tumour segmentation masks were retained unchanged. An expert neuroradiologist visually reviewed the masks against the corresponding MRI images, and timepoints with segmentations clearly non-representative of tumour tissue were excluded. Deterministic rule-based processing performed lesion definition, admissibility-filtered measurement extraction, longitudinal tracking, new-lesion detection, and provisional response classification. Expert RANO imaging annotations and bidimensional measurements served as comparators rather than complete clinical RANO 2.0 adjudications. Analyses assessed agreement in the sum of products of diameters (SPD), longitudinal burden change, response classification, progression timing, new-lesion agreement, and failure modes, with patient-level clustered bootstrap resampling.

Visual segmentation quality control excluded 57 timepoints across 42 patients, leaving 542 evaluable timepoints, including 506 matched to expert annotations. Across 84 expert-measured timepoints from 47 patients, the constrained maximum product of perpendicular diameters (Max-PPD-c) correlated with expert SPD (Pearson r = 0.753; Spearman ρ = 0.861), although the concordance correlation coefficient (CCC) was 0.680 and the median absolute percentage difference was 36.8%. Overall, 73.8% of measurements were within ±50% of expert SPD. Across 31 consecutive pairs from 17 patients, correlations for longitudinal percentage change were r = 0.715 and ρ = 0.874. Per-timepoint four-class response agreement was 48.1% (κ = 0.173), whereas final-timepoint binary progressive disease (PD) agreement was 76.0%. Among patients with PD by both methods, the pipeline assigned PD earlier more often than later, although the median timing difference was 0.0 weeks and the paired comparison was not statistically significant (p = 0.068). New-lesion detection had a positive predictive value (PPV) of 22.3% and negative predictive value (NPV) of 89.0%. Conservative patient-level later-positivity estimates were 10.8% within 12 weeks and 27.0% within 24 weeks, without same-lesion verification.

This study supports the value of a deterministic pipeline for standardising imaging-derived response outputs in glioblastoma clinical trials. Future work will focus on prospective multi-institutional validation incorporating clinical variables, non-enhancing disease assessment, and adjudicated outcomes.