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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Pituitary neuroendocrine tumors (PitNETs) require accurate assessment of aggressiveness for treatment and prognosis.
• Current noninvasive preoperative tools for evaluating PitNET aggressiveness are limited.
• Distinguishing aggressive PitNETs preoperatively is crucial for personalized patient management.

Purpose of the Study:

• To develop and validate a deep learning radiomics (DLR) model for noninvasive preoperative assessment of PitNET aggressiveness.
• To correlate DLR scores with established invasion classifications and pathological markers.
• To establish a robust imaging biomarker for guiding clinical decisions in PitNET management.

Main Methods:

• Development of a DLR model using nnUnet and Swin Transformer for automatic segmentation and feature extraction.
• Training and validation on a large cohort (n=1089) from three medical centers.
• Identification of 13 key radiomic features to construct the DLR model.

Main Results:

• The DLR score showed strong correlation with Knosp and Hardy-Wilson invasion classifications.
• The model outperformed existing classifications in predicting tumor recurrence.
• DLR scores indicated aggressive pathological markers (Ki-67, p53, macrophages) and revealed biological pathways (MAPK, TGF-β).

Conclusions:

• The developed DLR model provides a reliable, noninvasive preoperative tool for assessing PitNET aggressiveness.
• This imaging biomarker can support individualized treatment strategies by identifying high-risk tumors.
• Clinical deployment via an online platform facilitates the integration of this technology into practice.