人工智能在胸部创伤肋骨骨折CT诊断中应用的初步研究

doi:10.3969/j.issn.1674-8115.2021.07.012

摘要/Abstract

摘要：

目的·评价人工智能（artificial intelligence，AI）软件在胸部电子计算机体层扫描（computed tomograph，CT）图像上自动检测创伤性肋骨骨折的诊断效能。方法·收集2019年8月—9月因急性胸部外伤而行CT扫描的393例连续数据。以3位资深影像专家的共同阅片结果作为评估AI检出肋骨骨折病灶的参考标准。在病灶层面计算全部肋骨外伤改变以及不同类型肋骨骨折（错位型骨折、轻微骨折、骨皮质扭曲）的敏感度。并分别以全部肋骨外伤改变和错位型肋骨骨折为研究对象，在患者和肋骨2个层面分析AI自动检出肋骨骨折的效能。结果·在病灶层面上，AI检出全部肋骨外伤改变的敏感度为81.75%，错位型肋骨骨折检出的敏感度为94.85%，在3种骨折类型中最高（P=0.000）。以全部肋骨外伤改变为研究对象，AI在肋骨层面的骨折检出敏感度、特异度、阳性预测值、阴性预测值分别为82.45%、98.33%、75.30%和98.91%；在患者层面分别为90.91%、76.21%、77.63%和90.23%。以错位型肋骨骨折为研究对象时，AI在肋骨层面的骨折检出敏感度、特异度、阳性预测值、阴性预测值分别为94.57%、98.26%、51.94%和99.89%；在患者层面分别为95.56%、74.59%、52.76%和98.26%。结论·AI软件检出肋骨骨折病灶具有较高的敏感度，可望用于急性胸部外伤CT读片的初筛和风险分层。

关键词: 肋骨骨折, CT成像, 人工智能, 急诊影像学

Abstract:

Objective·To assess the detection rate of an artificial intelligence (AI) software for acute traumatic rib fractures on chest computed tomograph (CT) images.

Methods·A consecutive cohort of CT images of the patients with acute chest trauma were collected from August 2019 to September 2019 (n=393). The reference standard was defined as the consensus reading results of three radiologist experts. At the lesion level, the sensitivity was studied, including all lesions and different types of rib fractures (i.e., displaced rib fractures, mild fractures, and cortical distortion). The sensitivity was also studied at both patient and rib levels.

Results·At the lesion level, the total rib fracture detection sensitivity of AI was 81.75%. For displaced rib fractures, the sensitivity was 94.85%, which was the highest among the three types of fractures (P=0.000). When all types of rib fractures were taken as the object of study, the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of AI were 82.45%, 98.33%, 75.30% and 98.91% respectively at the rib level, and 90.91%, 76.21%, 77.63% and 90.23% respectively at the patient level. When displaced rib fracture were taken as the object, the sensitivity, specificity, PPV and NPV of AI were 94.57%, 98.26%, 51.94% and 99.89% respectively at the rib level, and 95.56%, 74.59%, 52.76% and 98.26% at the patient level.

Conclusion·AI software has high sensitivity in detecting the fracture sites, and it could be potentially used for screening the thorax CT images in patients with acute chest trauma.

Key words: rib fracture, computed tomography, artificial intelligence, emergency radiology

中图分类号:

R445.3

刘想, 谢辉辉, 许玉峰, 陶晓峰, 柳林, 吴迪嘉, 王霄英. 人工智能在胸部创伤肋骨骨折CT诊断中应用的初步研究[J]. 上海交通大学学报(医学版), 2021, 41(7): 920-925.

Xiang LIU, Hui-hui XIE, Yu-feng XU, Xiao-feng TAO, Lin LIU, Di-jia WU, Xiao-ying WANG. Application of artificial intelligence to CT diagnosis of thoracic traumatic rib sites: a preliminary study[J]. JOURNAL OF SHANGHAI JIAOTONG UNIVERSITY (MEDICAL SCIENCE), 2021, 41(7): 920-925.

图/表 9

图1 错位型肋骨骨折Note： Displaced fracture of the left 10th rib (arrow).

Fig 1 Displaced rib fracture

图2 轻微骨折Note： Mild fracture of the right 10th rib (arrow).

Fig 2 Mild rib fracture

图3 骨皮质扭曲Note： Buckle fracture of the left 4th rib (arrow).

Fig 3 Buckle rib fracture

图4 全部肋骨外伤改变与错位型肋骨骨折在不同层面的分布

Fig 4 Distribution of all traumatic and displaced rib fractures in different levels

表1 肋骨骨折病灶分布及检出情况（n）

Tab 1 Distribution and detection of rib fracture lesions （n）

Rib site	True positive /reference standard				False positive	Total detected lesion
Rib site	Displaced RF	Mild RF	Buckle RF	Total	False positive	Total detected lesion
R01	2/2	1/2	0/0	3/4	4	7
R02	5/5	2/4	0/3	7/12	2	9
R03	7/8	5/5	10/12	22/25	5	27
R04	16/16	1/1	16/22	33/39	9	42
R05	15/15	6/8	21/30	42/53	7	49
R06	11/12	3/4	24/27	38/43	5	43
R07	7/9	7/10	21/27	35/46	5	40
R08	8/8	2/5	13/16	23/29	13	36
R09	9/9	5/5	7/10	21/24	9	30
R10	7/7	3/4	4/6	14/17	10	24
R11	3/3	3/3	2/3	8/9	8	16
R12	1/1	0/0	0/0	1/1	4	5
L01	3/3	1/1	1/1	5/5	7	12
L02	3/3	8/10	4/6	15/19	1	16
L03	9/11	2/3	14/21	25/35	6	31
L04	12/12	2/4	14/23	28/39	9	37
L05	10/10	3/4	26/33	39/47	7	46
L06	10/10	7/7	24/26	41/43	10	51
L07	11/11	8/9	19/23	38/43	13	51
L08	8/8	7/7	12/16	27/31	6	33
L09	10/12	6/6	8/10	24/28	6	30
L10	13/14	5/6	3/5	21/25	9	30
L11	3/3	1/3	0/2	4/8	8	12
L12	1/2	0/2	0/1	1/5	11	12

表2 AI软件检出肋骨骨折敏感度（病灶层面）［%（n/n）］

Tab 2 Sensitivity of AI software in detecting rib fracture （lesion level）［% （n/n）］

Type	Sensitivity
All the traumatic changes	81.75 （515/630）
Displaced rib fracture	94.85 （184/194）
Mild rib fracture	77.88 （88/113）
Buckle rib fracture	75.23 （243/323）

图5 AI检出真阳性结果Note： A. A rib fracture lesion was detected by reference standard reading (red box). B. The reading result of AI software was consistent with reference standard reading (green box), and the detected lesion was considered as true positive lesion.

Fig 5 True positive result detected by AI software

图6 AI检出假阳性结果Note： A. No rib fracture lesion was detected by reference standard reading. B. A rib fracture lesion was detected by AI software (green box), and the detected lesion was considered as false positive lesion.

Fig 6 False positive result detected by AI software

表3 AI软件检出肋骨骨折的诊断效能（患者和肋骨层面）

Tab 3 Diagnostic efficiency of AI software in detecting rib fractures （patient and rib level）

Type	Level	SEN/%	SPE/%	PPV/%	NPV/%	PLR	NLR
All the traumatic changes	Patient	90.91	76.21	77.63	90.23	3.82	0.12
	Rib	82.45	98.33	75.30	98.91	49.37	0.18
Displaced rib fracture	Patient	95.56	74.59	52.76	98.26	3.76	0.06
	Rib	94.57	98.26	51.94	99.89	54.35	0.06

参考文献 23

1	Liman ST, Kuzucu A, Tastepe AI, et al. Chest injury due to blunt trauma[J]. Eur J Cardiothorac Surg, 2003, 23(3): 374-378.
2	Ivey KM, White CE, Wallum TE, et al. Thoracic injuries in US combat casualties: a 10-year review of Operation Enduring Freedom and Iraqi Freedom[J]. J Trauma Acute Care Surg, 2012, 73(6 ): S514-S519.
3	Bemelman M, de Kruijf MW, van Baal M, et al. Rib fractures: to fix or not to fix? An evidence-based algorithm[J]. Korean J Thorac Cardiovasc Surg, 2017, 50(4): 229-234.
4	de Jong MB, Kokke MC, Hietbrink F, et al. Surgical management of rib fractures: strategies and literature review[J]. Scand J Surg, 2014, 103(2): 120-125.
5	Chung JH, Cox CW, Mohammed TL, et al. ACR appropriateness criteria blunt chest trauma[J]. J Am Coll Radiol, 2014, 11(4): 345-351.
6	Livingston DH, Shogan B, John P, et al. CT diagnosis of rib fractures and the prediction of acute respiratory failure[J]. J Trauma, 2008, 64(4): 905-911.
7	Talbot BS, Gange CP Jr, Chaturvedi A, et al. Traumatic rib injury: patterns, imaging pitfalls, complications, and treatment[J]. Radiographics, 2017, 37(2): 628-651.
8	李洪涛. 16层螺旋CT与DR胸片对肋骨骨折的诊断价值分析[J]. 中国继续医学教育, 2016, 8(29): 47-48.
9	Gurcan MN, Sahiner B, Petrick N, et al. Lung nodule detection on thoracic computed tomography images: preliminary evaluation of a computer-aided diagnosis system[J]. Med Phys, 2002, 29(11): 2552-2558.
10	Lehman CD, Blume JD, DeMartini WB, et al. Accuracy and interpretation time of computer-aided detection among novice and experienced breast MRI readers[J]. AJR Am J Roentgenol, 2013, 200(6): W683-W689.
11	Lin FC, Li RY, Tung YW, et al. Morbidity, mortality, associated injuries, and management of traumatic rib fractures[J]. J Chin Med Assoc, 2016, 79(6): 329-334.
12	Dunham CM, Hileman BM, Ransom KJ, et al. Trauma patient adverse outcomes are independently associated with rib cage fracture burden and severity of lung, head, and abdominal injuries[J]. Int J Burns Trauma, 2015, 5(1): 46-55.
13	Barnea Y, Kashtan H, Skornick Y, et al. Isolated rib fractures in elderly patients: mortality and morbidity[J]. Can J Surg, 2002, 45(1): 43-46.
14	Holcomb JB, McMullin NR, Kozar RA, et al. Morbidity from rib fractures increases after age 45[J]. J Am Coll Surg, 2003, 196(4): 549-555.
15	Park S. Clinical analysis for the correlation of intra-abdominal organ injury in the patients with rib fracture[J]. Korean J Thorac Cardiovasc Surg, 2012, 45(4): 246-250.
16	Al-Hassani A, Abdulrahman H, Afifi I, et al. Rib fracture patterns predict thoracic chest wall and abdominal solid organ injury[J]. Am Surg, 2010, 76(8): 888-891.
17	Park HS, Ryu SM, Cho SJ, et al. A treatment case of delayed aortic injury: the patient with posterior rib fracture[J]. Korean J Thorac Cardiovasc Surg, 2014, 47(4): 406-408.
18	Pan Y, Shi D, Wang H, et al. Automatic opportunistic osteoporosis screening using low-dose chest computed tomography scans obtained for lung cancer screening[J]. Eur Radiol, 2020, 30(7): 4107-4116.
19	Ringl H, Lazar M, Töpker M, et al. The ribs unfolded—a CT visualization algorithm for fast detection of rib fractures: effect on sensitivity and specificity in trauma patients[J]. Eur Radiol, 2015, 25(7): 1865-1874.
20	高歌, 马帅, 王霄英. 计算机辅助诊断在医学影像诊断中的基本原理和应用进展[J]. 放射学实践, 2016, 31(12): 1127-1129.
21	高歌, 胡娟, 王成彦, 等. 整合临床及多参数MRI信息的前列腺癌CAD系统: 诊断效能研究[J]. 放射学实践, 2016, 31(12): 1143-1145.
22	韩超, 朱丽娜, 刘想, 等. 基于U-Net实现前列腺MR图像上腺体的自动分割和径线测量: 临床植入验证研究[J]. 放射学实践, 2020, 35(4): 519-524.
23	Liu Y, Liu Q, Han C, et al. The implementation of natural language processing to extract index lesions from breast magnetic resonance imaging reports[J]. BMC Med Inform Decis Mak, 2019, 19(1): 288.

[1]	李小敏, 曲扬, 张少霆, 赵亮, 刘畅, 谢帅宁, 戴尅戎, 艾松涛. 人工智能技术在骨肌系统影像学方面的应用[J]. 上海交通大学学报(医学版), 2021, 41(2): 262-266.
[2]	陆言巧，沈兰，何奔. 人工智能在心血管疾病的辅助诊疗中的应用[J]. 上海交通大学学报（医学版）, 2020, 40(2): 259-.
[3]	张璐1，陈强2，蒋蓓蓓1，丁珍红2，张丽3，解学乾1. 基于生成式对抗网络的冠状动脉CT血管成像运动伪影去除的初步研究[J]. 上海交通大学学报(医学版), 2020, 40(09): 1229-1235.
[4]	荣雯雯1，汪刚1，朱其立2. 基于人工智能的病历后结构化专病数据库在临床研究中的价值探讨[J]. 上海交通大学学报(医学版), 2020, 40(07): 995-1000.
[5]	茅怡铭 1，吴长江 2. 老年多发肋骨骨折患者肺部并发症风险预测模型的建立[J]. 上海交通大学学报（医学版）, 2019, 39(1): 69-.
[6]	滕继平, 倪达, 程佑爽, 等. 急诊手术内固定对创伤性连枷胸患者血流动力学的影响[J]. , 2010, 30(12): 1486-.
[7]	滕继平, 程云阁, 倪达, 等. 手术内固定与非手术外固定治疗创伤性连枷胸的效果比较[J]. , 2009, 29(12): 1495-.