S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation

Abstract

S2D-Align introduces a shallow-to-deep auxiliary learning strategy for radiology report generation that progressively aligns generated reports with anatomical structures, moving from coarse global alignment to fine-grained anatomical grounding. This anatomically-grounded generation paradigm improves the clinical accuracy and faithfulness of automatically generated radiology reports.