Prompt Engineering in Large Language Models for BI-RADS Classification of Imaging Reports: A Retrospective Evaluation

Abstract Purpose To evaluate how prompt engineering modulates large language models' (LLMs) accuracy in Breast Imaging Reporting and Data System (BI-RADS) classification of digital breast tomosynthesis (DBT) reports. Materials and Methods This retrospective study collected reports from 216 patients who underwent DBT for breast cancer screening or diagnosis. BI-RADS classifications were independently assigned to all reports by two experts. Three LLMs (GPT-4o, GPT-o3 mini, Qwen-2.5 max) were utilized to classify all reports using different prompts. Besides, six human readers independently assigned BI-RADS classifications. Agreement between experts and LLMs for BI-RADS categories was evaluated using Weighted Cohen’s kappa (κw). Friedman and Nemenyi tests assessed κw differences among three prompt conditions.The frequencies of changed BI-RADS category assignments, which could impact clinical management, were also calculated. Results In prompt III, GPT-4o achieved near-perfect agreement with experts (κw = 0.80), surpassing GPT-o3 mini (0.76) and Qwen-2.5 max (0.79). Its κw was significantly higher in prompt III than in prompt II (0.69, P, P < 0.05) and prompt I (0.63,, P < 0.01). While GPT-4o's κw remained lower than two mid-level radiologists (0.89 and 0.86), it exceeded two entry-level radiologists (0.76 and 0.79). Regarding clinical management changes, prompt III yielded a 14.8% discordance rate with experts, outperforming prompts I (29.6%) and II (28.2%), and aligning with entry-level radiologists (15.3%, 14.4%). Conclusion With optimized prompts, GPT-4o achieved near-perfect agreement and matched the clinical management performance of entry-level radiologists. These findings support the use of LLMs as an auxiliary tool for BI-RADS classification in breast cancer diagnosis by radiologists.