Ilies Djebbara, Morten Winkler Møller, Ivar Yannick Christensen, Bo Halle, Christian Bonde Pedersen, Frantz Rom Poulsen, Jan Saip Aunan-Diop
Preoperative MRI radiomics was associated, at the group level, with a composite methylation-defined risk label previously associated with postoperative regrowth in the same source cohort. The study was clinically outcome-anchored, but the classifier did not use individual regrowth or progression-free survival as its endpoint and should not be interpreted as an independently validated prognostic model. Pooling k3, k4, and k5 does not establish a shared imaging phenotype of aggressiveness. External outcome-linked validation is required before clinical implementation.
PURPOSE: DNA methylation profiling identifies clinically relevant subgroups of non-functioning pituitary adenomas (NFPAs), but requires tumour tissue and is unavailable preoperatively. We investigated whether reported outcome-associated methylation-defined NFPA phenotypes are associated with preoperative MRI radiomic features.
METHODS: We performed a single-centre retrospective radiogenomic analysis nested within a previously published, outcome-characterised NFPA methylation cohort. Seventy-four patients with preoperative gadolinium-enhanced T1-weighted MRI were included. Tumours were automatically segmented using a fine-tuned U-Net. The primary analysis tested binary discrimination between the outcome-associated low-risk (k1/k2) and high-risk (k3/k4/k5) methylation strata using a prespecified L1-penalised logistic-regression model with leakage-controlled repeated stratified cross-validation, 0.632 + bootstrap optimism correction, and 1000-shuffle whole-pipeline permutation testing.
RESULTS: Seven radiomic features were false-discovery-rate significant and four survived Bonferroni correction, predominantly LoG-filtered first-order intensity features. The prespecified model separated high-risk from low-risk tumours with balanced accuracy 0.69 (95% CI, 0.58-0.80), AUC 0.73 (0.61-0.84), 0.632 + AUC 0.71, and permutation p = 0.001. The SF1-restricted analysis remained significant (balanced accuracy 0.73, AUC 0.70, p = 0.008), whereas k3-focused analyses were not significant. Clinical sex-lineage models classified at chance.
CONCLUSIONS: Preoperative MRI radiomics was associated, at the group level, with a composite methylation-defined risk label previously associated with postoperative regrowth in the same source cohort. The study was clinically outcome-anchored, but the classifier did not use individual regrowth or progression-free survival as its endpoint and should not be interpreted as an independently validated prognostic model. Pooling k3, k4, and k5 does not establish a shared imaging phenotype of aggressiveness. External outcome-linked validation is required before clinical implementation.