Spatially grounded longitudinal comparison in radiology

Establish effective methods for tightly coupling spatial grounding with longitudinal comparison in radiology, such that clinically meaningful changes are identified through the evolution of localized evidence across prior and current chest X-ray studies.

Background

Longitudinal chest X-ray interpretation requires comparing prior and current examinations to determine how findings change over time, while visual grounding connects clinical language to localized image evidence. The paper identifies the integration of these two capabilities as unresolved because clinically important changes may be confined to specific anatomical regions rather than reflected in global image differences. CheXGround addresses this gap by representing corresponding anatomical regions across paired studies and aligning temporally contextualized region tokens with clinical phrases, but the broader problem of tightly coupling spatial grounding with longitudinal radiological comparison remains open.

References

However, tightly coupling spatial grounding with longitudinal comparison in radiology remains open, as clinically meaningful change may depend on how localized evidence evolves across prior and current studies.

— CheXGround: Anatomical Region Tokens for Grounded Longitudinal Chest X-ray Interpretation  (2608.30758 - Gebremedhin et al., 31 Aug 2026) in Section 1, Introduction

These backbones are trained and evaluated one study at a time, so what their representations retain across examinations of the same patient has not been established.

— BrainDiff: Longitudinal Report Generation for Multimodal Brain MRI  (2609.00593 - Patel et al., 1 Sep 2026) in Related Work, paragraph “3D medical foundation models”