[Paper Review] Deep Learning Predicts Biomarker Status and Discovers Related Histomorphology Characteristics for Low-Grade Glioma

Zijie Fang, Yihan Liu+11 more|arXiv (Cornell University)|Oct 11, 2023

Digital Imaging for Blood Diseases4 citations

TL;DR

This paper proposes Multi-Beholder, a deep learning pipeline that predicts five low-grade glioma (LGG) biomarkers using only H&E-stained whole slide images, combining multiple instance learning with one-class classification for accurate instance-level pseudo-labeling. It achieves AUCs up to 0.973 on TCGA-LGG and 0.820 on an external cohort, enabling interpretable discovery of histomorphological correlates to biomarker status.

ABSTRACT

Biomarker detection is an indispensable part in the diagnosis and treatment of low-grade glioma (LGG). However, current LGG biomarker detection methods rely on expensive and complex molecular genetic testing, for which professionals are required to analyze the results, and intra-rater variability is often reported. To overcome these challenges, we propose an interpretable deep learning pipeline, a Multi-Biomarker Histomorphology Discoverer (Multi-Beholder) model based on the multiple instance learning (MIL) framework, to predict the status of five biomarkers in LGG using only hematoxylin and eosin-stained whole slide images and slide-level biomarker status labels. Specifically, by incorporating the one-class classification into the MIL framework, accurate instance pseudo-labeling is realized for instance-level supervision, which greatly complements the slide-level labels and improves the biomarker prediction performance. Multi-Beholder demonstrates superior prediction performance and generalizability for five LGG biomarkers (AUROC=0.6469-0.9735) in two cohorts (n=607) with diverse races and scanning protocols. Moreover, the excellent interpretability of Multi-Beholder allows for discovering the quantitative and qualitative correlations between biomarker status and histomorphology characteristics. Our pipeline not only provides a novel approach for biomarker prediction, enhancing the applicability of molecular treatments for LGG patients but also facilitates the discovery of new mechanisms in molecular functionality and LGG progression.

Motivation & Objective

• To develop a cost-effective, interpretable method for predicting LGG biomarkers without relying on expensive molecular testing.
• To overcome limitations of current biomarker detection, including high cost, technical complexity, and inter-rater variability.
• To enable widespread clinical application of molecular diagnostics by leveraging routine H&E whole slide images.
• To discover quantitative and qualitative histomorphological features associated with specific biomarker statuses in LGG.
• To improve the accessibility and accuracy of biomarker prediction in low-grade glioma using deep learning.

Proposed method

• The framework integrates multiple instance learning (MIL) with one-class classification to generate accurate instance-level pseudo-labels from slide-level annotations.
• One-class classification is applied to identify representative patches within each whole slide image that are most likely to reflect the biomarker status.
• The model learns to distinguish between positive and negative patch-level features using only slide-level labels, reducing reliance on precise patch-level annotations.
• A dual-branch attention mechanism enhances feature learning by focusing on morphologically relevant regions across multiple tissue patches.
• The pipeline is trained end-to-end on whole slide images, with inference performed at the patch level to predict biomarker status.
• Interpretability is achieved via attention maps and saliency visualization, linking predicted biomarker status to specific histological patterns.

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict five key LGG biomarkers using only H&E-stained whole slide images?
• RQ2 How can one-class classification improve pseudo-labeling accuracy in a multiple instance learning framework for biomarker prediction?
• RQ3 What histomorphological features are quantitatively and qualitatively correlated with specific biomarker statuses in LGG?
• RQ4 How generalizable is the model across diverse patient populations and scanning protocols?
• RQ5 Can the model’s interpretability reveal biologically meaningful patterns related to molecular functionality and tumor progression?

Key findings

• The model achieved an area under the ROC curve (AUC) of 0.973 on the internal TCGA-LGG cohort, demonstrating high predictive performance.
• On an external, independent Xiangya cohort with diverse race and scanning protocols, the model achieved an AUC of 0.820, indicating strong generalization.
• The integration of one-class classification significantly improved instance-level pseudo-labeling accuracy, enhancing slide-level prediction performance.
• Interpretability analysis revealed distinct histomorphological patterns associated with each biomarker, including nuclear density, chromatin texture, and cellular arrangement.
• The pipeline successfully discovered novel morphological correlates linked to molecular functionality, suggesting potential biological insights into LGG progression.
• The model’s performance was robust across different scanning protocols and patient demographics, supporting clinical translatability.