DMGN

DMGN predicts patient survival by integrating imaging mass cytometry (IMC) single-cell spatial phenotype data with clinical variables using a multimodal graph-based network.

Key Features:

• Integration of Multiplexed Imaging Data: Integrates whole IMC images and all marker channels to capture single-cell–level biomarker expression and the spatial architecture of the tumor microenvironment.
• Graph-Based Analysis: Constructs multimodal graphs that represent spatial relationships between phenotypic features and adaptively models interactions across image regions and clinical variables.
• Clinical Data Embedding: Generates specialized embeddings for each clinical variable to enable multimodal aggregation with imaging-derived features.
• End-to-End Methodology: Implements an end-to-end approach that jointly uses IMC-derived features and clinical embeddings to produce patient-level survival predictions.

Scientific Applications:

• Oncological prognosis: Predicting cancer patient survival by combining spatial phenotypic information and clinical variables.
• Tumor microenvironment analysis: Characterizing spatial architecture and phenotypic heterogeneity of the tumor microenvironment at single-cell resolution.
• Breast cancer prognostic research: Demonstrated improved survival prediction performance on breast cancer datasets compared with prior methods.
• Personalized treatment stratification: Informing patient-specific treatment strategies via survival risk estimation that integrates imaging and clinical data.

Methodology:

DMGN comprises two explicitly stated computational modules: a Multimodal Graph-Based Module that constructs and analyzes graphs representing spatial relationships between phenotypic features from whole IMC images and adaptively integrates these graphs with clinical variables, and a Clinical Embedding Module that generates embeddings for each clinical variable to enable multimodal aggregation in an end-to-end survival prediction framework.