Page 81 - 南京医科大学学报自然科学版

第42卷第2期 南京医科大学学报（自然科学版）
2022年2月 Journal of Nanjing Medical University（Natural Sciences） ·227 ·

·临床医学·

人工智能软件辅助诊断新鲜肋骨骨折的效能评估

朱雅茹，祁 良，徐 磊，王梦悦，徐蕴潮，纪执琳，邹月芬 *
南京医科大学第一附属医院放射科，江苏 南京 210029

［摘 要］ 目的：评估人工智能软件（artificial intelligence system，AI）辅助放射科医师在胸部计算机断层扫描（computed tomog⁃
raphy，CT）新鲜肋骨骨折诊断中的应用效能。方法：收集行胸部CT检查的新鲜肋骨骨折病例共508例。6名影像医师分为低
年资组（≤5年）和高年资组（> 5年），使用PACS系统对CT图像进行独立阅片；在间隔为期4周的洗脱期后，影像医师在结合AI
辅助诊断结果的前提下对胸部CT图像进行第2轮阅片；记录骨折类型、部位及诊断时间。采用配对卡方检验比较AI辅助前后
诊断骨折的灵敏度、特异度有无统计学差异。绘制受试者操作特征（receiver operating characteristic，ROC）曲线，并计算曲线下
面积（area under curve，AUC），应用Medcalc 软件对AUC 进行显著性检验。采用Cohen’s kappa 系数分析低年资医师和高年资
医师诊断肋骨骨折的一致性。应用配对样本t检验比较人工智能辅助前后医师的诊断时间是否具有差异。结果：508例胸部
CT 共包含骨折 2 883 处，AI 辅助后低年资医师和高年资医师的诊断灵敏度由 77.95%、83.96%提升至 88.52%、90.98%（P <
0.001）；平均AUC从0.902增加到0.948（P < 0.001）；独立阅片和AI辅助诊断后低年资医师与高年资医师之间的Cohen’s kappa
系数由0.832提升至0.900；诊断每个肋骨骨折病例的时间平均减少了28.43 s（P < 0.001）。结论：AI辅助医师诊断新鲜肋骨骨
折能够提高检出效能，并减少诊断时间。
［关键词］ 人工智能；新鲜肋骨骨折；计算机断层扫描；效能；阅片时间
［中图分类号］ R814.42 ［文献标志码］ A ［文章编号］ 1007⁃4368（2022）02⁃227⁃06
doi：10.7655/NYDXBNS20220214

Effectiveness evaluation of artificial intelligence system in the diagnosis of fresh rib
fractures

ZHU Yaru，QI Liang，XU Lei，WANG Mengyue，XU Yunchao，JI Zhilin，ZOU Yuefen *
Department of Radiology，the First Affiliated Hospital of Nanjing Medical University，Nanjing 210029，China

［Abstract］ Objective：This study aims to evaluate the effectiveness of using an artificial intelligence system to assist radiologists in
the diagnosis of fresh rib fractures with computed tomography（CT）. Methods：A dataset of 508 cases with fresh rib fracture CT images
were collected and analyzed by 6 radiologists independently with the PACS system，the radiologists were divided into two groups：low
seniority group（≤5 years）and high seniority group（>5 years）. After a washout period of four weeks，the CT images were evaluated
again by the 6 radiologists with the assistance of AI system. The fracture type，site，and time of diagnosis were recorded for further
analysis. The paired chi⁃square test was used to compare the sensitivity and specificity in diagnosis of fresh fracture with and without
AI assistance. The receiver operating characteristic curve（ROC）and the area under curve（AUC）was calculated. Cohen’s kappa
coefficient was used to analyze the consistency between low and high seniority radiologists in diagnosing rib fractures. Paired sample t
test was used to compare the difference in diagnosis time. Results：A total of 2 883 fresh rib fractures were found in the 508 CT
images. With the help of AI⁃assisted diagnosis，the diagnostic sensitivity of low and high seniority radiologists increased significantly
from 77.95%，83.96% to 88.52%，90.98%（P < 0.001）；The average AUC increased from 0.902 to 0.948（P < 0.001）；The Cohen’s
kappa coefficient between low and high seniority radiologists increased from 0.832 to 0.900. The average diagnosis time of each case
decreased 28.43s，P < 0.001. Conclusion：With the AI assistance in the diagnosis of fresh rib fracture，the detection efficiency of rib
fracture is improved and the diagnosis time is reduced.
［Key words］ artificial intelligence；fresh rib fracture；computed tomography；efficiency；diagnosis time
［J Nanjing Med Univ，2022，42（02）：227⁃232］
［基金项目］ 国家自然科学基金（81701652）
∗
通信作者（Corresponding author），E⁃mail：zou_yf@163.com