关键词: 大语言模型, 乳腺癌, 超声, 钼靶

Abstract: Objective To evaluate the application effectiveness of ChatGPT 4 and the fine-tuned Llama 3 model in breast cancer diagnosis,particularly in processing unstructured reports and diagnostic imaging of ultrasound,mammography,and their combined modalities. Methods Retrospective data from 689 patients who underwent both breast ultrasound and mammography examinations were collected.The diagnostic performance of the two models was compared across text and image modalities,and the impact of breast density on model performance was explored. Results In the text modality,the fine-tuned Llama 3 model performed excellently,achieving a combined diagnostic accuracy of 91.7%,outperforming 71.7% of ChatGPT 4.In the image modality,both models had accuracies below 70%,but ChatGPT 4 exhibited higher sensitivity（78.3%）,while Llama 3 demonstrated outstanding specificity（98.3%）.Subgroup analysis indicated that mammography performed better in non-dense breasts,whereas ultrasound was more advantageous in dense breasts. Conclusions The large language models still require further optimization in medical image processing and multimodal integration,but fine-tuned large language models in the medical field show potential in handling unstructured clinical texts.

Key words: large language model, breast cancer, ultrasound, mammography