A Deep Learning Ensemble Model for the Classification of Pituitary Neuroendocrine Tumors Subtypes Using Magnetic Resonance Imaging

Motivation: Pituitary Neuroendocrine Tumors (PitNET) are a diverse group, representing 15% of intracranial tumors. Preoperative MRI cannot differentiate PitNET subtypes effectively, requiring patients to undergo surgery before classification is performed. Goal(s): Current classification relies on clinical manifestations and hormone tests which can be controversial and inconclusive. The study aims to develop a non-invasive MRI-based method to classify PitNET subtypes. Approach: We developed an MRI-based deep learning ensemble model to classify PitNET subtypes by transcription factors: TPIT, SF1, and PIT1. Results: Our model achieved an accuracy of 76.92%, with class-specific AUC values of 0.88 for TPIT, 0.96 for SF1, and 0.86 for PIT1. Impact: Our model provides a non-invasive method for classifying PitNET subtypes using MRI, potentially enhancing the accuracy of preoperative diagnosis beyond reliance on hormone tests alone ultimately improving patient outcomes.