Radiomics and artificial intelligence in pituitary adenomas (PitNETs): from imaging biomarkers to clinical...

Pituitary neuroendocrine tumors (PitNETs) are common intracranial neoplasms with heterogeneous hormonal activity, invasiveness, and treatment response. Conventional MRI is central to diagnosis, but is limited in quantifying tumor heterogeneity and predicting clinically relevant features such as consistency, molecular subtype, proliferation, and clinical outcomes. To review current evidence on MRI-based radiomics and artificial intelligence (AI) in PitNETs, focusing on clinical applications, methodological quality, and future integration into precision medicine. Narrative review of studies published between 2019 and 2025 evaluating radiomics, machine learning, and deep learning applied to PitNETs, highlighting tumor consistency, molecular and histological subtypes, proliferation, invasiveness, treatment response, recurrence, visual outcomes, and differential diagnosis of sellar lesions. Radiomics has been applied across the PitNET clinical trajectory. Texture- and shape-based features from T2-weighted and multiparametric MRI predict intraoperative tumor consistency with AUCs often > 0.80, outperforming conventional radiology. Radiomic signatures allow non-invasive differentiation of functional vs. non-functional adenomas, somatotroph granulation patterns, silent corticotroph adenomas, prolactinomas, and prediction of Ki-67 and PIT-1 expression. Models also predict cavernous sinus invasion, postoperative regrowth, recurrence, visual outcomes, and treatment response, including dopamine-agonist therapy. Differential diagnosis of hypophysitis and Rathke's cleft cyst has also been explored. Methodological quality remains heterogeneous, with frequent single-center, retrospective designs, limited external validation, and suboptimal adherence to reporting guidelines. Radiomics and AI show promise for risk stratification and personalized therapy in PitNETs, particularly for predicting tumor consistency, molecular subtypes, and recurrence. Clinical implementation requires standardized imaging protocols, multicenter datasets, reproducible segmentation, interpretable models, and prospective validation.