Deep learning based semi-automated model can predict lineage in patients with pituitary neuroendocrine tumors.

Abstract

Pituitary neuroendocrine tumors (PitNETs) represent the most prevalent category of neuroendocrine neoplasms. Contemporary classification paradigms emphasize transcription factor immunohistochemistry (IHC) as a cornerstone for molecular subtyping and risk stratification. However, the clinical adoption of this approach is hindered by the lack of standardized interpretative thresholds for antibody staining and limited global availability of specialized reagents, particularly in resource-limited settings. To address these challenges, we developed a semi-automated computational framework that predicts PitNET lineages directly from hematoxylin and eosin (H&E)-stained histology slides. The pipeline employs a dynamic confidence threshold: samples below this threshold undergo confirmatory IHC staining and manual pathological review, while those surpassing it are classified automatically. In prospective validation, this approach achieved a 68.9% reduction in diagnostic workload while maintaining 95.9% overall accuracy. Similar efficacy was observed in functional (74.4% workload reduction, 99.0% accuracy) and external (39.3% reduction, 95.1% accuracy) cohorts. Statistical analysis confirmed non-inferiority between semi-automated predictions and fully manual IHC-based evaluations in all the cohorts. Furthermore, we implemented a deep learning-based virtual IHC staining module, generating synthetic transcription factor images demonstrating high morphological concordance with ground-truth IHC slides. Notably, our computational analysis revealed distinct histomorphological correlates of lineages: SF1-lineage tumors exhibited homogeneous cellular architecture characterized by densely packed, compact cells with reduced cytoplasmic volume, whereas PIT1-lineage neoplasms displayed larger cells with expanded intercellular spacing and disorganized spatial arrangements.