Skip to content
Book Open access

ProAlign: Anatomically Aligned Progression Modeling for Longitudinal Chest X-ray Report Generation

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 19 references

Abstract

Longitudinal chest X-ray report generation requires joint understanding of current and prior studies to accurately describe disease evolution. Existing methods often underuse prior images and reports, rely on implicit temporal fusion, and overlook cross-time anatomical misalignment caused by posture and acquisition differences, leading to unreliable progression descriptions. We propose ProAlign, a framework for anatomically aligned and clinically grounded progression modeling. Its Anatomical Feature Alignment module formulates cross-time correspondence as a regularized Optimal Transport problem, enabling anatomically consistent matching without rigid registration and providing a reliable structural basis for temporal comparison. Built on this, a Text-Grounded Progression Estimation module uses the prior report as a semantic prior to suppress irrelevant visual changes and derive a robust Disease Progression Field (DPF) for progression-aware report generation and clinically faithful decoding. By combining anatomically aligned visual comparison with prior-report-guided semantic filtering, ProAlign better distinguishes true pathological evolution from non-pathological variations and reduces clinically unreliable temporal reasoning. On MIMIC-CXR and M-ABN, ProAlign outperforms prior state-of-the-art methods with CE F1 improvements of 3.5% and 3.2%, respectively.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.