2026 Poster Presentations
P214: RADIOMIC CLUSTERING OF FIRST-ORDER FEATURES PREDICTS POSTOPERATIVE OUTCOMES FOLLOWING PITUITARY ADENOMA RESECTION
Mohammad Hamo, BS1; Dhanush Amin, MD1; Yifei Sun, BS1; Alexander Mullen, MD1; Hariteja Ramapuram, BS1; Michael E Ivan, MD, MBS2; Carolina G Benjamin, MD2; Kristen Riley, MD1; Elizabeth Liptrap, MD1; Dagoberto Estevez-Ordonez, MD, PhD2; 1University of Alabama, Birmingham; 2University of Miami
Background: Transsphenoidal resection of pituitary adenomas can result in adverse outcomes, including diabetes insipidus (DI), hypopituitarism, hyponatremia, and unplanned readmissions. Radiomic analysis offers a means to characterize tumor heterogeneity through first-order intensity features. Unsupervised clustering of first-order radiomic features may help identify patients at risk of postoperative complications.
Methods: We conducted a single center, retrospective review of 77 patients from 2012 to 2024 who underwent transsphenoidal resection of pituitary adenomas. Preoperative and outcome variables were obtained. T1- and T2-weighted images we acquired as part of routine preoperative care. Scans were obtained by 1.5T or 3T scanners using standard pituitary protocols, and image slice thickness ranged from 1-3mm. Exact acquisition parameters varied based on scanner and year of acquisition, and outputs were studied for variations in quality (e.g., slice thickness). Masks were segmented or verified by a neuroradiology fellow (DA) near the completion of fellowship training. Radiomic feature extraction was performed using PyRadiomics (version 3.0.1, Python 3.11.4), extracting first-order features. Data was preprocessed, and k-means clustering with elbow method was used to define subgroups. Subgroups were statistically compared for outcomes using Chi-square, with Fisher’s exact test for pairwise comparisons.
Results: First-order statistics included 18 features reflecting tumor heterogeneity through uniformity, brightness, and pixel distribution. Clustering generated 3 subgroups: cluster 1 with low-intensity tumors (n=33), cluster 2 with heterogenous or intermediate intensity (n=36), and cluster 3 with hyperintense tumors (n=8). Postoperative diabetes insipidus was significantly different between groups (p = 0.002), with pairwise comparisons revealing higher rates in cluster 3 compared to 1 and 2 (p = 0.007, 0.001, respectively). Other outcomes, including hypopituitarism, hyponatremia, and readmissions within 30, 60, or 90 days revealed no differences between groups (p = 0.409, 0.943, 0.483, 0.509, 0.130, respectively).
Conclusion: Unsupervised clustering of radiomic features may help define higher-risk subgroups following transsphenoidal resection. Large, multi-institutional cohorts could help confirm generalizability and guide postoperative monitoring strategies.
Figure 1. Principle component analysis plot showing separation of patients into three radiomic subgroups based on first-order features
Figure 2. Example of MRI input and tumor mask segmentation
