View-Aware Chest X-Ray Report Generation with Relational-Contrastive Alignment

Abstract

One of the most popular diagnostic imaging modalities is the chest X-ray (CXR), however creating radiology reports still takes a lot of time and requires radiologist competence. Current automated techniques limit cross-view comprehension and report quality by frequently using single-view images, discarding fine-grained spatial information through global average pooling, and using inflexible decoding methodologies. This study proposes a view-aware chest X-ray report generation framework that integrates multiple radiographic views, strengthens visual–text semantic alignment, and calibrates decoding automatically. The framework comprises a domain-specific visual representation using a CheXpert-pretrained DenseNet-121 encoder that preserves 7×7 spatial features, a view-aware embedding, a relational-contrastive semantic alignment module, and an automatic decoding calibration mechanism. Evaluated on the IU X-ray dataset using paired frontal–lateral images and a DistilGPT2 decoder, the proposed framework achieved BLEU-1 of 0.4768, BLEU-2 of 0.3004, BLEU-3 of 0.2056, BLEU-4 of 0.1525, METEOR of 0.4001, ROUGE-L of 0.3211, and CIDEr of 0.3812. These results demonstrate that the proposed framework improves the coherence and clinical relevance of generated chest X-ray reports while providing an effective vision-language framework for multi-view radiology report generation.