基于深度学习的视网膜毛细血管无灌注区智能识别定量分析系统在CRVO缺血分型诊断中的应用

目的　探讨一种基于深度学习的视网膜毛细血管无灌注区（NPA）智能识别定量分析系统在视网膜中央静脉阻塞（CRVO）缺血分型诊断中的应用价值。方法　回顾性病例系列研究。选取2017年1月至2018年12月经山西省眼科医院门诊确诊为CRVO并行荧光素眼底血管造影（FFA）检查的343例343眼患者纳入本研究。随访并记录患者自发病起1年内是否发生新生血管并发症。应用人工智能诊断系统计算CRVO患者后极部55°视野范围NPA面积，受试者工作特征曲线（ROC）分析该NPA面积对CRVO发生新生血管并发症的诊断价值。结果　343例CRVO患眼中有26眼发生了新生血管并发症，发病率7.58%。依据NPA面积判断CRVO患者是否发生新生血管并发症的曲线下面积为0.889（95%CI:0.799~0.978，P<0.001）。最佳截断值为20.997视盘面积（DA），灵敏度为0.808，26眼新生血管并发症患者中有21眼NPA值大于该值；特异度为0.946，317眼无并发症患者中有300眼NPA值小于该值。结论　基于深度学习的NPA智能识别定量分析系统可为CRVO分型诊断提供决策依据。55°视野范围后极部NPA>20 DA可作为CRVO缺血分型的阈值标准。

Application of intelligent recognition and quantitative analysis system based on deep learning for retinal capillary non-perfusion area in ischemic typing of central retinal vein occlusion

Objective　To explore the application value of an intelligent recognition and quantitative analysis system based on deep learning for the retinal capillary non-perfusion area (NPA) in ischemic typing of central retinal vein occlusion (CRVO). Methods　A total of 343 patients (343 eyes) who were diagnosed with CRVO and underwent fundus fluorescein angiography (FFA) in the outpatient department of Shanxi Eye Hospital from January 2017 to December 2018 were included in this retrospective case series study. All these patients were followed up for neovascularization complications within one year from the onset of the disease. The artificial intelligence-based diagnostic system was used to calculate the NPA area at 55° in the posterior pole of CRVO patients. The receiver operating characteristic (ROC) curve was used to evaluate the diagnostic value of the NPA area on the neovascularization complications in CRVO patients. Results　Neovascularization complications occurred in 26 eyes (7.58%) of 343 CRVO patients. The area under the curve for determining whether CRVO patients developed neovascularization complications based on the NPA area was 0.889 (95%CI: 0.799-0.978, P<0.001). The best cutoff value was 20.997 disk area (DA), the sensitivity was 0.808, and 21 of 26 eyes with neovascularization complications had the NPA area greater than this value. The specificity was 0.946, and 300 of 317 eyes without neovascularization complications had the NPA area less than this value. Conclusion　The NPA intelligent recognition and quantitative analysis system based on deep learning can provide a decision-making basis for CRVO typing diagnosis. NPA>20 DA at 55° in the posterior pole can be used as the threshold for CRVO ischemia typing.