目的 分析人工智能(AI)辅助下医师对肝脏MRI影像的诊断效能，探讨AI在肝脏MRI影像诊断中的应用价值。方法 2020年1月-2022年12月行肝脏MRI平扫及增强扫描检查的患者103例，3名低年资(≤5年)和1名高年资(＞5年)影像科医师首先对肝脏MRI影像进行独立阅片。间隔4周后，4名医师均在AI智能辅助诊断软件辅助下对肝脏MRI影像进行第2次独立阅片。记录肝背景类型(正常、肝硬化、脂肪肝、铁过载)、病灶性质(肝实性结节、肝囊肿、肝血管瘤)及阅片时间。比较AI辅助前、后医师阅片时间;以另2名高年资(＞10年)医师一致性结论作为诊断金标准，比较AI辅助前、后医师诊断的灵敏度、特异度;采用Kappa检验评价AI辅助前、后医师诊断结果与金标准的一致性。结果 103例患者中，肝背景正常69例，肝背景异常34例(肝硬化11例，脂肪肝15例，铁过载8例);病灶1 272个，其中肝实性结节508个，肝囊肿721个，肝血管瘤43个。低年资医师、高年资医师AI辅助后阅片时间[(572。46±44。31)、(510。15±52。82)s]均短于 AI 辅助前[(691。67±37。05)、(607。62±40。24)s](t=10。120，P＜0。001;t=5。766，P=0。001)。低年资医师甲、乙、丙AI辅助后诊断肝背景异常的灵敏度(97。06％、100。00％、100。00％)均高于AI辅助前(41。18％、32。35％、50。00％)(P＜0。05)，高年资医师AI辅助后诊断肝背景异常的灵敏度与AI辅助前比较差异无统计学意义(P＞0。05);低年资医师甲、乙、丙及高年资医师AI辅助后诊断肝实性结节(90。75％、92。91％、98。03％、95。87％)、肝囊肿(97。50％、99。86％、99。72％、99。86％)、肝血管瘤(97。67％、100。00％、95。35％、97。67％)的灵敏度均高于 AI 辅助前(79。13％、82。48％、85。04％、83。86％，86。96％、88。63％、90。01％、90。85％，39。53％、60。47％、51。16％、83。72％)(P＜0。05);4名医师AI辅助后诊断肝背景异常、肝实性结节、肝囊肿、肝血管瘤的特异度与AI辅助前比较差异均无统计学意义(P＞0。05)。低年资医师甲、乙、丙及高年资医师AI辅助后诊断肝背景异常、肝实性结节、肝囊肿、肝血管瘤的Kappa值均高于AI辅助前。以金标准为对照，AI独立诊断肝背景异常、肝实性结节、肝囊肿、肝血管瘤的灵敏度分别为 100。00％、96。85％、98。34％、93。02％，特异度分别为 100。00％、99。74％、100。00％、99。67％，假阳性率分别为0、3。53％、2。54％、0。16％。结论 AI智能辅助诊断软件可提高医师对肝脏MRI影像诊断的灵敏度和准确率，缩短阅片时间，提高诊断效率。
Predictive value of artificial intelligence in liver MRI diagnosis
Objective To analyze the efficiency of diagnosis by physicians assisted with artificial intelligence(AI)on liver MRI image,and to discuss the application value of AI in liver MRI diagnosis.Methods From January 2020 to December 2022,103 patients underwent liver MRI plain scan and enhanced scan.Three junior physicians with practicing experience of ≤5 years and 1 senior physician with practicing experience of＞5 year read the films of liver MRI images by themselves.Four weeks later,all 4 physicians reviewed the films for the second time with AI assistance.The background type of liver(normal liver,cirrhosis,fatty liver,iron overload),lesion nature(hepatic solid nodules,hepatic cysts,hepatic hemangioma)and the time of reading films were recorded.The time of reading films was compared with and without AI assistance.Taking the consistent conclusions of two other senior physicians with practicing experience of＞10 years as the gold diagnostic standard,the sensitivity and specificity of diagnosis were compared with and without AI assistance.Kappa test was used to evaluate the consistency of diagnosis by physicians with and without AI assistance with the gold standard.Results Among 103 patients,69 patients had normal liver background,and 34 had abnormal liver background(cirrhosis in 11 patients,fatty liver in 15,iron overload in 8).There were 1 272 lesions,including 508 hepatic solid nodules,721 hepatic cysts and 43 hepatic hemangiomas.The junior and senior physicians took shorter time in reading films with AI assistance[(572.46±44.31),(510.15±52.82)s]than those without AI assistance[(691.67± 37.05),(607.62±40.24)s](t=10.120,P＜0.001;t=5.766,P=0.001).The sensitivities of abnormal liver background diagnosed by three junior physicians with AI assistance(97.06％,100.00％,100.00％)were higher than those without AI assistance(41.18％,32.35％,50.00％(P＜0.05),and showed no significant differences by senior physicians with and without AI assistance(P＞0.05).The sensitivities in hepatic solid nodules(90.75％,92.91％,98.03％,95.87％),hepatic cysts(97.50％,99.86％,99.72％,99.86％),and hepatic hemangiomas(97.67％,100.00％,95.35％,97.67％)diagnosed by three junior physicians and one senior physician with AI assistance were higher than those without AI assistance(79.13％,82.48％,85.04％,83.86％;86.96％,88.63％,90.01％,90.85％;39.53％,60.47％,51.16％,83.72％)(P＜0.05).There were no significant differences in the specificities in abnormal liver background,hepatic solid nodules,hepatic cysts and hepatic hemangiomas diagnosed by four physicians with and without AI assistance(P＞0.05).The Kappa coefficients in abnormal liver background,hepatic solid nodules,hepatic cysts and hepatic hemangiomas diagnosed by four physicians with AI assistance were higher than those without AI assistance.Compared with the gold standard,the sensitivities of AI diagnosis in abnormal liver background,hepatic solid nodules,hepatic cysts and hepatic hemangiomas were 100.00％,96.85％,98.34％and 93.02％,the specificities were 100.00％,99.74％,100.00％and 99.67％,and the false positive rates were 0,3.53％,2.54％and 0.16％,respectively.Conclusion AI assisted diagnosis software can improve the sensitivity and accuracy of liver MRI diagnosis,shorten the time of reading the film,and improve the diagnostic efficiency.
NETL
NSTL
万方数据
