Anatomy-Guided Radiology Report Generation With Pathology-Aware Regional Prompts

Goal: Radiology report generation holds significant potential to alleviate clinical workloads and streamline medical care. However, achieving high clinical accuracy remains challenging, as radiographs often feature intricate structures and subtle pathologies. Methods: To address these challenges, this work introduces an innovative approach that explicitly integrates anatomical and pathological information into report decoding by leveraging pathology-aware regional prompts. Specifically, we develop an anatomical region detector that extracts structured visual features from distinct anatomical areas, coupled with a novel multi-label pathology detector that identifies global abnormalities. Results: Our model demonstrates superior report generation performance in natural language generation and clinical efficacy, surpassing previous state-of-the-art methods. It achieved scores of 0.394 in BLEU-1, 0.302 in ROUGE-L, and 0.470 in F1, reflecting substantial improvements in both linguistic fluency and medical accuracy. Formal expert evaluations further affirmed the model's potential to elevate radiology practice. Conclusion: By integrating anatomical and pathological insights to emulate radiologists' workflow, our model achieves superior accuracy and clinical coherence of radiology reporting. It offers remarkable promise to support clinical decision-making and transform patient management.