Multi-Beholder: Biomarker Prediction for Low-Grade Glioma with Multiple Instance Learning and One-Class...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Molecular diagnostics

Background:

• Biomarker detection is crucial for low-grade glioma (LGG) diagnosis and treatment.
• Current methods involve costly, complex genetic testing with potential for variability.
• There is a need for accessible and reliable biomarker detection in LGG.

Purpose of the Study:

• To develop an interpretable deep learning pipeline, Multi-Beholder, for predicting LGG biomarker status.
• To utilize only hematoxylin and eosin (H&E)-stained whole slide images for biomarker prediction.
• To improve the accuracy and accessibility of biomarker detection in LGG.

Main Methods:

• Developed Multi-Biomarker Histomorphology Discoverer (Multi-Beholder), a deep learning pipeline.
• Integrated one-class classification with multiple instance learning for pseudo-labeling.
• Validated the pipeline on TCGA-LGG and Xiangya cohorts using H&E whole slide images.

Main Results:

• Multi-Beholder achieved high prediction performance, with an Area Under the Receiver Operating Characteristic Curve (AUC) up to 0.973 (TCGA-LGG) and 0.820 (Xiangya).
• The pipeline demonstrated interpretability, correlating biomarker status with histomorphology.
• Accurate instance-level pseudo-labeling improved prediction accuracy.

Conclusions:

• Multi-Beholder offers a novel, accurate, and interpretable approach for LGG biomarker prediction from H&E images.
• This method enhances the applicability of molecular treatments for LGG patients.
• The pipeline facilitates discovery of new mechanisms in LGG progression and molecular functionality.