Gulluk et al. introduce Transition-Aware best-of-N sampling for longitudinal chest X-ray reports
First training-free method to compare prior and current chest X-rays for better report generation.
Deep Dive
Gulluk et al. developed Transition-Aware best-of-N sampling, a training-free method for chest X-ray report generators that explicitly accounts for changes between prior and current exams. It splits reports into sentences, embeds them as unordered sets, computes a directional vector representing the transition, and scores candidates by cosine similarity to cached ground-truth vectors. Outperforms random selection across three generators, with largest gains on Impression sections.
Key Points
- First training-free sampling scheme for longitudinal chest X-ray reports that considers prior-to-current transitions.
- Uses four directional set distances (mean-shift, novelty residual, directed-Hausdorff anchor, cost-weighted optimal transport) to encode change.
- Outperforms random selection on three vision-language generators, with largest gains on Impression sections (up to 20% improvement).
Why It Matters
Radiologists rely on comparing current and prior X-rays; this AI method generates reports that reflect that longitudinal context.