FASA-Net: A Frequency-Domain-Enhanced and Anatomically Constrained Vision-Language Framework for Automated Chest X-ray Report Generation

Automated radiology report generation can reduce radiologists' workloads by producing preliminary reports from medical images. However, generating chest X-ray reports remains challenging. Subtle lesions and organ boundaries may be underrepresented, visual regions may be poorly aligned with medical terminology, and long reports require coherent clinical descriptions. To address these challenges, we propose FASA-Net, a vision-language framework that combines frequency-domain enhancement with anatomical semantic constraints. Its Frequency-Domain Pyramid Attention Network (FPAN) integrates frequency-domain analysis and a multiscale feature pyramid to enhance local structures and high-frequency visual details. The Anatomical Semantic Constraint Attention (ASCA) module guides cross-modal interactions between visual regions and anatomy-related terms to reduce the risk of regional misalignment and semantic hallucination. The Context Aggregation Module (CAM) incorporates a summary of the text generated so far into decoding to improve coherence and terminological consistency. Experiments on the public IU X-Ray and MIMIC-CXR datasets show competitive performance on natural language generation metrics, including BLEU, METEOR, and ROUGE-L, together with improved clinical label consistency measured using CheXbert-based clinical efficacy metrics. Ablation studies, visualizations, and case studies further illustrate the contributions of FPAN, ASCA, and CAM to visual detail enhancement, anatomical semantic alignment, and contextual aggregation. FASA-Net therefore has potential as an auxiliary tool for chest X-ray report generation, although clinical deployment requires further validation using multicenter data, radiologist assessments, and prospective studies.