探讨DeepSeek大语言模型在胃癌和直肠癌超声报告结构化及T分期自动评估中的应用价值。

本研究纳入四川省肿瘤医院2023年1月至2024年12月进行的胃癌和直肠癌超声检查报告共121份。由资深超声医师团队制定胃癌和直肠癌超声报告结构化模板，使用DeepSeek R1和V3模型进行结构化信息提取和T分期评估。采用召回率、精确率和F1分数评估结构化报告生成的性能，并以准确性评估T分期性能。邀请3位医师对比评估DeepSeek生成的报告与原始报告，评价其在审阅效率和临床易用性方面的表现。

DeepSeek R1与V3模型在结构化信息提取方面召回率、精确率和F1分数均高于0.9，二者差异均无统计学意义（P均＞0.05）。在T分期评估中，采用推理模式的DeepSeek R1模型准确性最高，达到76.86%，显著优于DeepSeek V3模型的59.50%，二者差异具有统计学意义（χ²=8.51，P＜0.05）。与审阅原始报告所需的平均时间[（60.96±6.11）s/份]相比，审阅DeepSeek R1[（18.12±4.52）s/份]（t=60.38；P＜0.001）和DeepSeek V3[（17.15±2.60）s/份]（t=71.98；P＜0.001）生成的结构化报告所需时间缩短。5分李克特量表评分结果显示，原始报告的评分为3（3，3）分，而DeepSeek R1和V3报告的评分分别为1（1，2）分（Z=-9.72；P＜0.001）和1（1，2）分（Z=-9.95；P＜0.001），差异具有统计学意义。

DeepSeek大语言模型，特别是R1版本，可有效从胃癌和直肠癌超声报告中提取结构化信息，并在T分期评估方面展现出较高的准确性，其生成的报告有助于提高审阅效率，并具有辅助临床决策的潜力。

To investigate the utility of the DeepSeek large language model (LLM) in the structured generation of ultrasound reports and the automatic assessment of T-staging for gastric and rectal cancer.

A total of 121 ultrasound examination reports for gastric and rectal cancer, collected from Sichuan Cancer Hospital between January 2023 and December 2024, were included in this study. A structured template for gastric and rectal cancer ultrasound reports was developed by a team of senior sonographers. The DeepSeek R1 and V3 models were employed to extract structured information and assess T-staging. The performance of structured report generation was evaluated using recall, precision, and F1 score, while T-staging performance was assessed based on accuracy. Three physicians were invited to compare the reports generated by DeepSeek with the original reports to evaluate review efficiency and clinical usability.

Regarding structured information extraction, both DeepSeek R1 and V3 models achieved recall, precision, and F1 scores exceeding 0.9, with no statistically significant differences between the two (P＞0.05). In T-staging assessment, the DeepSeek R1 model (utilizing reasoning mode) achieved the highest accuracy of 76.86%, which was significantly superior to the 59.50% achieved by the DeepSeek V3 model (χ²=8.51, P＜0.05). Compared to the average time required to review original reports [(60.96±6.11) s/report], the review time for structured reports generated by DeepSeek R1 [(18.12±4.52) s/report] (t=60.38; P＜0.001) and DeepSeek V3 [(17.15±2.60) s/report] (t=71.98; P＜0.001) was significantly shortened. The 5-point Likert scale evaluation showed that the score for the original reports was 3 (3, 3), while the scores for the DeepSeek R1 and V3 reports were 1 (1, 2) (Z=-9.72; P＜0.001) and 1 (1, 2) (Z=-9.95; P＜0.001), respectively, indicating a statistically significant difference.

The DeepSeek large language models, particularly the R1 version, can effectively extract structured information from gastric and rectal cancer ultrasound reports and demonstrates high accuracy in T-staging assessment. The generated reports contribute to improved review efficiency and possess the potential to assist in clinical decision-making.