基于深度学习人工智能辅助系统用于CT检出肋骨新鲜骨折

目的 观察基于深度学习的人工智能(AI)辅助系统用于CT检出肋骨新鲜骨折的效能。方法 由2名高年资影像科医师对1 000例急诊CT所示肋骨新鲜骨折患者进行逐层标注,再由另2名高年资医师进行审核;将图像导入深睿医疗Dr.Wise^®骨折CT影像辅助检测系统(简称Dw_AI)。于1 000例中随机抽取60例(40例肋骨新鲜骨折、20例无骨折)作为数据集2,以其余940例为数据集1(902例肋骨新鲜骨折、38例无骨折)。由1名影像科主任医师(CR)基于数据集1独立评估肋骨新鲜骨折,将其结果与Dw_AI结果进行对比,评估Dw_AI的效能。由2名低年资和2名中等年资医师参与多阅片者多病例(MRMC)临床试验,基于数据集2,分别于病灶、肋骨和患者级别评估Dw_AI辅助不同年资医师诊断的效能。结果 数据集1 全部940例中, 2 946支肋骨存在3 452处新鲜骨折;Dw_AI对各级别肋骨新鲜骨折的敏感度均高于CR(P均＜0.05)。数据集2全部60例中,112支肋骨存在123处新鲜骨折;Dw_AI辅助下,不同年资医师诊断各级别肋骨新鲜骨折的敏感度均有所提高(P均＜0.05)。结论 AI辅助系统用于CT检出肋骨新鲜骨折的效能较佳,且能提高医师、尤其低年资医师的诊断敏感度。

Artificial intelligence assisted system based on deep learning for detecting fresh rib fractures on CT

Objective To observe the efficacy of artificial intelligence (AI) assisted system based on deep learning (DL)for detecting fresh rib fractures on CT. Methods Fresh rib fractures on CT images of 1 000 patients who underwent emergency chest scanning for chest trauma were labeled layer by layer by 2 high qualified radiologists and then reviewed by 2 other high qualified radiologists, and the final annotation results were considered as the gold standards. The images were imported into Dr.Wise^® fracture CT image aided detection system (Dw_AI for short). Among 1 000 cases, 60 were randomly selected as data set 2 (including 40 with and 20 without rib fresh fractures), while the remaining 940 were enrolled as data set 1 (including 902 with and 38 without fresh rib fractures). Fresh rib fractures of data set 1 were independently assessed by a chief radiologist (CR), and the results of Dw_AI and CR were compared to assess the efficacy of Dw_AI. Based on data set 2, 2 residents and 2 attending physicians participated in a multi-reader multi-case (MRMC) clinical trial, and the efficacy of Dw_AI for assisting physicians diagnosing rib fresh fractures at lesion, rib and patient levels were evaluated. Results Among 940 cases in data set 1, 2 946 ribs were found with 3 452 fresh fractures. At lesion, rib and patient levels, Dw_AI demonstrated higher sensitivity in detecting rib fresh fractures in data set 1 than CR (all P<0.05). Among 60 cases in data set 2, there were 112 ribs with 123 fresh fractures. Under the assistance of Dw_AI, the sensitivity of all 4 physicians with different seniorities for diagnosing rib fresh fractures in data set 2 increased at lesion, rib and patient levels (all P<0.05). Conclusion AI assisted system had good efficacy for detecting fresh rib fractures on CT, which could also assist physicians, especially junior physicians to improve detection rate of fresh rib fractures.