Transition-Aware Best-of-N Scores X-Ray Reports by Change

Radiologists don't read a chest X-ray in isolation, they read it against the patient's prior exam, and much of what they write is about what has changed. Report-generation models mostly ignore that. A new paper on arXiv proposes a fix that requires no retraining, and instead reranks the candidate reports a pretrained model already produces by how well they describe the transition from one visit to the next.

The mechanic is a best-of-N sampling scheme. Each candidate report is split into sentences, each sentence is embedded into an unordered set, and the set of sentences from the prior exam and the current exam are compared using a directional set-to-set distance. That transition vector is then scored by cosine distance against a cached bank of ground-truth training transitions, aggregated as either a minimum distance or a k-nearest-neighbours score. Four different distance definitions are tried, mean-shift, novelty residual, directed-Hausdorff anchor, and cost-weighted optimal transport. The authors call it transition-aware best-of-N sampling, and claim it is the first training-free scheme of its kind for chest X-ray report generators.

Why this matters if you're not building radiology models: report-generation systems have been getting good at describing the picture in front of them and less good at the part clinicians spend most of their time on, the delta between today's scan and last month's. A wrapper that sits on top of any pretrained generator and reranks its candidates for change-awareness is the kind of thing a hospital or vendor can adopt without touching model weights, which is where a lot of the deployment friction actually lives.

The honest caveat is that the reported comparison is against random selection of candidates, not against stronger reranking baselines, and the abstract does not give hard numbers, only that gains are consistent and largest on the Impression section. Evaluation runs on a multi-visit AP-PA cohort with three vision-language generators under three prompts, so generalisation to other imaging protocols and patient populations is an open question. What the paper doesn't tell you yet is which of the four distances actually wins, or how well the cached transition bank holds up when the patient distribution shifts.

The direction is the interesting part, treating change as the first-class object rather than state. If that framing spreads to other longitudinal medical imaging tasks, the same trick could plug in without touching the underlying generator.