2026 Proffered Presentations
S082: PREOPERATIVE FLAIR RADIOMICS PREDICT CRANIOPHARYNGIOMA HISTOLOGIC SUBTYPE
Briana A Santo, PhD; Omar Selim, BS; Mazin Elshareif, BS; Austin Carmichael, BS; Taha Khalilullah, BS; Yves Greatti, MS; Ritvik Pulya, BS; Calixto-Hope Lucas, MD; Connor Liu, MD; A. Karim Ahmed, MD; Debraj Mukherjee, MD; Johns Hopkins University School of Medicine
Introduction: Craniopharyngiomas (CPs) are surgically complex tumors arising in the sellar and suprasellar region, causing significant morbidity due to mass effect on surrounding neurovasculature. They occur in two main histologic subtypes, adamantinomatous (ACP) and papillary (PCP), which differ in both behavior and treatment response. Notably, over 90% of PCPs harbor the BRAFV600E mutation, an actionable genomic alteration that, when targeted, may reduce tumor burden without surgical intervention. However, preoperative genetic testing is costly and technically challenging. The identification of interpretable imaging biomarkers capable of reliably distinguishing CP subtypes prior to surgery is critical to enable personalized treatment strategies, guide surgical planning, and improve patient outcomes.
Objective: To identify an interpretable preoperative MRI radiomic signature predictive of craniopharyngioma histologic subtype.
Methods: This study included pathologically confirmed craniopharyngioma cases that underwent surgical resection at our institution between January 1, 2015, and January 1, 2025. Preoperative T1CE, T2, and FLAIR MRI scans were collected and registered using the BRAINS module in 3D Slicer. Tumors were manually annotated by experienced raters, with inter-rater agreement assessed via the Dice coefficient and Hausdorff distance. Radiomic features were extracted from the tumor region in each MRI series using PyRadiomics varying the bin size parameter to capture tumor texture at multiple resolutions. Two-sample tests with FDR correction identified features differing significantly between ACP and PCP. A logistic regression (LR) model was developed to classify ACP versus PCP using stratified 3-fold cross-validation, hyperparameter tuning, and feature selection. Model-selected features were ranked by relative feature importance, and the most predictive features were visualized as heatmaps using PyRadiomics and Python. Our workflow is summarized in Figure 1.
Results: This study analyzed 71 craniopharyngiomas, including 59 ACPs and 12 PCPs. High agreement among annotators was observed, with a mean dice coefficient of 0.91±0.04 and mean Hausdorff distance of 2.50±2.19 mm (Figure 2A-B). Radiomic features that differed significantly between subtypes included both morphological and textural characteristics (Figure 2C-E). For example, ACPs exhibited larger tumor surface areas (q=0.009) in alignment with their characteristic irregular, multi-lobulated morphology. Meanwhile, PCPs demonstrated higher skewness on T2-weighted images (q=0.001), highlighting their mostly solid, homogeneous appearance on T2, with few, small regions of elevated intensity. An optimized logistic regression model achieved a mean AUC of 0.905 using six selected features (Figure 3). These model-selected features were all derived from FLAIR images, and included GLCM SumSquares 128 (feature rank 1, q<0.001) and GLCM Inverse Variance 48 (feature rank 6, q=0.001) (Figure 4). Here GLCM SumSquares 128 was elevated in ACPs and when visualized as a feature map emphasized hyperintensity on the peripheral rim, perhaps capturing the margin adjacent reactive gliosis. In comparison, GLCM Inverse Variance 48, which measures homogeneity, was increased in PCPs highlighting their predominantly solid, isointense FLAIR appearance.
Conclusions: Preoperative FLAIR radiomics can accurately distinguish ACP from PCP, and visualized features reflect characteristic imaging patterns of each subtype. These FLAIR-derived textural features may capture biologically meaningful tumor attributes and warrant further radiogenomic analysis to investigate associations with BRAF mutation status to advance non-invasive, preoperative tumor characterization.
