Feature-interactive Siamese graph encoder-based image analysis to predict STAS from histopathology images in lung...

Area of Science:

• Oncology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Spread Through Air Spaces (STAS) is a critical prognostic indicator in lung cancer.
• Histopathological analysis for STAS is subjective, time-consuming, and prone to errors.
• Current methods limit large-scale STAS assessment and clinical application.

Purpose of the Study:

• To develop and validate an automated image analysis model for accurate STAS detection in lung cancer.
• To improve the efficiency and objectivity of STAS diagnosis.
• To provide a tool that aids in prognosis assessment and surgical decision-making.

Main Methods:

• Development of VERN, a Siamese graph encoder model for analyzing histopathological images.
• Utilizing feature sharing and skip connections to capture spatial topological features.
• Training and validation on a large dataset of 1,546 lung cancer histopathology slides.

Main Results:

• VERN achieved high performance with an AUC of 0.9215 in internal validation.
• External validation demonstrated robust performance with AUCs of 0.8275 and 0.8829 on different test sets.
• The model showed strong generalizability across multiple datasets, indicating clinical-grade accuracy.

Conclusions:

• VERN offers a reliable and efficient method for automated STAS detection in lung cancer.
• The model's performance suggests its potential to significantly enhance clinical diagnosis and patient management.
• An open platform is provided to facilitate wider adoption and further research in STAS analysis.