Uncertainty as a Proxy for Semantic Correctness in Diffusion-Based Medical Image Synthesis
What it is
This paper investigates whether uncertainty metrics can serve as a proxy for anatomical correctness in diffusion models that synthesize contrast-enhanced CT (CECT) from non-contrast CT (NCCT). It compares six uncertainty estimation methods, finding that Monte Carlo Dropout (MCDropout) is particularly effective at flagging severe generation failures and detecting out-of-distribution cases without requiring additional training costs.
The problem it solves
Visually realistic synthesized medical images are not necessarily anatomically correct, and standard similarity metrics do not directly measure semantic or anatomical validity, making it difficult to ensure clinical safety.
What you could use it for
- Quality filtering of synthesized medical images to remove anatomically incorrect outputs
- Reliability assessment of diffusion model generations in clinical workflows
- Detection of out-of-distribution patient cases where the model may fail
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.