超轻量级DeepSeek-R1大语言模型在CT报告分类中的微调性能研究OA

Objective To address the clinical need for accurate severity stratification in CT report triage by fine-tuning an ultra-lightweight large language model(LLM),DeepSeek-R1-1.5B,and systematically evaluating its per-formance.Methods A total of 6,000 CT reports were retrospectively collected,including 5,000 reports for training,validation,and internal testing,and an additional 1,000 reports for independent external testing.Reports were annotated by experienced radiologists into three categories based on clinical significance:"negative,""routine"(common condi-tions),and"urgent"(severe conditions).Four lightweight models-DeepSeek-R1-1.5B,BERT-base-uncased,Qwen2.5-1.5B,and LLaMA3.2-1B-were fine-tuned and evaluated.An untrained full-scale model,DeepSeek-R1-671B,was included as a zero-shot baseline.Model performance was assessed using classification accuracy and cor-responding 95％confidence intervals(CIs).Results After fine-tuning,DeepSeek-R1-1.5B achieved the best per-formance,with accuracies of 0.964(95％CI:0.962-0.966)on the internal test set and 0.962(95％CI:0.960-0.963)on the external test set.Its performance was significantly superior to that of the other fine-tuned lightweight mod-els(P＜0.001)and the zero-shot DeepSeek-R1-671B model(P＜0.001).Subgroup analysis demonstrated consist-ently high performance across different scan types and anatomical regions.Conclusion The domain-adapted ultra-lightweight model DeepSeek-R1-1.5B demonstrates high accuracy in three-class severity classification of CT reports,highlighting its potential for clinical deployment in resource-constrained settings.