RNA-to-image multi-cancer synthesis using cascaded diffusion models.

Area of Science:

• Biomedical data science
• Computational biology
• Medical imaging

Background:

• Data scarcity hinders robust machine learning in biomedicine.
• Existing synthetic data methods often focus on single data modalities.
• Need for multi-modal synthetic data generation in cancer research.

Approach:

• Proposed RNA-Cascaded-Diffusion-Model (RNA-CDM) for RNA-to-image synthesis.
• Utilized variational auto-encoder for gene expression dimensionality reduction.
• Employed cascaded diffusion model for realistic whole-slide image tile synthesis from RNA-Seq data.

Key Points:

• Generated synthetic image tiles accurately preserve cell type distribution and fractions.
• Synthetic data improved machine learning model pretraining performance.
• Demonstrated gene expression modifications impact synthetic cell type composition.

Conclusions:

• RNA-CDM offers a novel solution for data scarcity in cancer diagnosis.
• Synthetic data can pretrain models, impute missing modalities, and advance precision medicine.
• Highlights potential for robust clinical decision support systems.