Laboratory Test-Guided Medical Image Generation for Multi-Modal Disease Prediction

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Biomedical imaging analysis

Background:

• Accurate disease prediction relies on integrating laboratory tests and medical images.
• Medical imaging data often suffers from temporal sparsity, hindering effective multi-modal analysis and reducing prediction accuracy.
• Laboratory tests are frequently collected, offering a denser temporal signal compared to imaging data.

Purpose of the Study:

• To address the challenge of temporal sparsity in medical imaging for improved disease prediction.
• To develop a novel method for generating synthetic medical images at additional time points, conditioned on laboratory test results.
• To enhance multi-modal interaction between laboratory tests and medical images for more accurate disease forecasting.

Main Methods:

• Proposed an Organ-Centric Modal-Shared Image Generator to synthesize medical images.
• Utilized an Organ-Centric Graph to link laboratory tests and imaging abnormalities, with organs as central nodes.
• Implemented a Knowledge-Guided Modal-Shared Trajectory Module to unify multi-modal features into an organ state trajectory over time.

Main Results:

• The proposed method successfully generates additional medical images, mitigating temporal sparsity.
• Demonstrated significant improvements in multi-modal disease prediction performance across various conditions.
• The organ-centric approach effectively bridges the gap between sparse imaging and dense laboratory data.

Conclusions:

• The Organ-Centric Modal-Shared Image Generator enhances disease prediction by improving multi-modal data integration.
• Generating temporally consistent medical images based on laboratory data is a viable strategy to overcome data sparsity.
• This approach holds promise for advancing AI-driven diagnostic tools in clinical practice.