An MRI radiomics approach using invasion-based weak supervision for identifying and evaluating aggressive PitNETs

Pituitary neuroendocrine tumor (PitNET) aggressiveness critically affects treatment and prognosis, yet reliable noninvasive preoperative tools remain lacking. We developed a deep learning radiomics (DLR) model integrating automatic segmentation, feature extraction, selection, and DLR score computation, trained on the training cohort and validated on the remaining cohorts (total n = 1089 from three medical centers). Using nnUnet and a fine-tuned Swin Transformer, 13 key features were identified to construct the model. The DLR score demonstrated strong correlation with Knosp and Hardy-Wilson invasion classifications, while outperforming them in predicting recurrence and indicating aggressive pathological markers (Ki-67, p53, macrophages) and revealing biological pathways (MAPK, TGF-β). The model was further implemented into an online platform, enabling clinical deployment. This noninvasive preoperative approach provides a robust imaging biomarker for identifying and evaluating PitNET aggressiveness and may support individualized treatment strategies.