CXR-CoRe: Two-Stage View- and Section-Aware Study-Level Chest X-ray Report Retrieval

Reliable linkage of chest X-ray examinations to corresponding reports is fundamental to large radiology archives. Yet heterogeneous view composition and similar image-report content across examinations, particularly longitudinal records, make exact content-based linkage difficult. Global vision-language representations support archive-wide search but can obscure view-specific and section-level correspondences, whereas exhaustive fine-grained interaction is computationally costly. We propose CXR-CoRe, a two-stage view- and section-aware framework. Stage 1 combines learned view embeddings with residual gated multi-view fusion, achieving 89.3% exact-pair coverage in its top-50 candidate set. Stage 2 reranks these candidates using patch-to-section correspondence and predicted clinical-label consistency, with same-subject inter-study negatives providing examination-specific supervision. On the full MIMIC-CXR test set of 3269 studies from 293 subjects, CXR-CoRe exceeded a controlled CXR-CLIP retraining by 3.8 and 5.0 percentage points in Recall@1 and Recall@5, respectively, and achieved higher Recall@10, mAP@5, and nDCG@5; the controlled baseline attained a lower MeanRank. Stage 2 improved all six metrics across three evaluation settings, with paired-bootstrap support at deeper cutoffs and in ranking quality. In GPU-resident online matching after shared feature extraction, candidate-limited matching was 16.3 times faster and used 93.6 times fewer estimated operations than full-archive one-layer cross-attention. These results establish coarse-to-fine retrieval as an effective strategy for exact study-level image-report linkage in large radiology archives.