Deep Learning Enabled Scoring of Pancreatic Neuroendocrine Tumors Based on Cancer Infiltration Patterns

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Pancreatic neuroendocrine tumors (PanNETs) are diverse neoplasms with varying prognoses.
• Current WHO grading and infiltration pattern scoring for PanNETs rely on subjective human judgment, leading to observer variability.
• Automated systems for PanNET categorization using quantitative metrics are lacking.

Purpose of the Study:

• To develop and validate a novel computational pipeline for the automated categorization of PanNETs.
• To address the need for objective, quantitative metrics in PanNET classification.
• To reduce intra- and inter-observer variability in PanNET diagnosis.

Main Methods:

• A deep learning pipeline utilizing graph neural networks (GNNs) was developed to automatically categorize PanNETs.
• The pipeline constructs entity-graphs representing cells and their relationships within the tumor microenvironment.
• Domain knowledge was integrated into GNN construction and training to enhance the analysis of architectural changes.

Main Results:

• The developed pipeline achieved a 76.70% F1-score on a test set of 105 whole slide images of PanNET tissues.
• The domain knowledge-integrated approach demonstrated significant improvements over existing methods.
• The system quantitatively characterizes PanNETs by analyzing cell-level features and their spatial organization.

Conclusions:

• The proposed computational pipeline represents the first automated system for PanNET categorization using deep learning and GNNs.
• Integrating domain knowledge into the GNN model improves the accuracy of PanNET classification.
• This approach offers a promising solution for objective and reproducible PanNET diagnosis, reducing reliance on subjective interpretation.