Abstract
<title>Abstract</title> <p>Purpose Contrast-enhanced MRI is the cornerstone of neuro-oncological imaging, but concerns regarding sustainability and cost prompted interest in the development of gadolinium-based contrast agent (GBCA)-free alternatives. We investigated the ability of human raters to guess the contrast-enhancing tumor borders in glioblastoma using GBCA-free MRI and whether a deep learning algorithm can improve this delineation. Methods A segmentation model was trained using BraTS (n = 989) and UCSF-PDGM (n = 501) datasets. Ground truth segmentations of the enhancing tumor regions were generated using an established algorithm and corrected by an independent neuroradiologist. T1-weighted (T1w), T2-weighted (T2w), T2-FLAIR, DWI-B0 and B-1000, ADC MRI of 117 glioblastoma, IDH-wildtype, patients were independently re-segmented by three radiologists and one non-clinician. Another radiologist (R-AI) improved segmentations produced by an AI algorithm without post-contrast sequence input. Statistical analysis compared AI segmentations with human performance. Results All raters showed a moderate-to-strong correlation with the ground truth (r = 0.79; P < 0.001). AI assistance increased Dice similarity by 0.09 (0.82 vs. 0.72; P < 0.001), greatly reduced Hausdorff distance (9.97 vs. 16.16 mm; P < 0.001), and decreased absolute volume error by 10.3 ml (6.75 vs. 17.07 ml; P < 0.001). Rater confidence increased with larger tumor volume, despite nonsignificant volumetric error metrics after correction (all P ≥ 0.24). Conclusion Human raters predicted enhancing tumor areas from GBCA-free sequences with moderate accuracy, showing considerable inter-rater variability in glioblastoma. AI-based GBCA-free segmentation assistance significantly improved human segmentations.</p>