Can AI learn to compare medical images like radiologists do?

Radiologists diagnose by comparing current scans to prior studies and similar cases—but medical AI typically interprets images in isolation. This work treats comparison as a core task: build retrieval to find analogous cases and a vision-language model to explain interval changes. Using 690,000 images across eight institutions, they train entity-aware models that decompose reports into anatomical structures and findings. MedReCo retrieval improves external matching by 6 percentage points; MedReCo-VLM boosts longitudinal accuracy by 14–46% on chest radiographs. Code and model released.