Radiologists reading follow-up chest X-rays rely on comparison with prior studies to spot what has changed. According to a new arXiv preprint, most automated report generation systems either ignore this longitudinal context entirely or incorporate it in a limited way. The authors propose CPVB-RG, a prior-aware framework built around a cross-prior visual bridging mechanism.
The idea is to explicitly connect visual features from the current exam with those from earlier scans, rather than treating prior reports or images as a simple add-on. The stated goal is to better capture new, resolved, or progressing findings in generated reports.
The paper is a preprint, so its claims have not yet been peer-reviewed. Still, it highlights a practical gap in clinical AI: generating accurate reports for follow-up studies may require models that can reason across time, not just within a single image.