A Modular Hybrid Transformer Framework for Trustworthy Radiology Report Summarization

Transformer-based language models have demonstrated strong potential for clinical text summarization, yet their applicability in radiology remains limited by hallucination, loss of critical diagnostic content, and inconsistency with reporting conventions. To address these challenges, this work proposes a modular hybrid framework combining extractive sentence scoring, domain-aware clinical entity filtering, and transformerbased abstractive generation. The system is evaluated on the Indiana University Chest X-ray dataset, using paired Findings and Impression sections as supervised targets. Quantitative results indicate substantial improvements over an extractive baseline, with higher ROUGE and BERTScore values, while qualitative inspection suggests stronger semantic alignment, clearer organization of clinical abnormalities, and reduced speculative phrasing. The findings indicate that architectural modularity and domain filtering provide structured mechanisms for constraining generative behavior and improving controllability in medical summarization. Overall, the approach contributes to ongoing efforts toward safe and clinically coherent Natural Language Processing systems for high-stakes healthcare applications.