特发性黄斑裂孔的人工智能诊断研究

Artificial intelligence diagnosis of idiopathic macular hole

目的利用计算机图像特征识别和特征参数提取算法实现特发性黄斑裂孔的人工智能诊断。方法收集2018年5月至8月在西安市第四医院眼科诊断为特发性黄斑裂孔患者48例(48眼)和同期健康志愿者48人(48眼)眼底光学相干断层扫描(optical coherence tomography,OCT)图像。通过对收集的OCT图像进行人工智能学习,利用图像处理和特征识别判断技术,提取能够区别正常人眼和特发性黄斑裂孔患眼的特征参数,在此基础上得到诊断的初始阈值。对2018年9-12月于西安市第四医院眼科诊断为特发性黄斑裂孔的患者73例(73眼)和正常51人(51眼)的OCT图像进行1~124随机编号后,使用计算机程序逐一进行处理,对处理后图像进行特征参数提取,然后将特征参数和阈值进行比较。结果经计算得到训练样本中正常人OCT图像特征参数为9,特发性黄斑裂孔患者OCT图像特征参数为23,初始阈值为16。经过计算机智能诊断,124例随机图像中73例OCT特征参数最小值为16.8,最大值为27.5,平均值为23.4,特征参数均大于阈值;51人随机图像OCT特征参数最小值为2.8,最大值为14.7,平均值为8.3,特征参数均小于阈值。经过比对,特征参数大于阈值的73例OCT图像均为特发性黄斑裂孔,特征参数小于阈值的51例OCT图像均为正常人。计算机判断结果与眼科医师判断结果差异无统计学意义(P=0.551)。结论一种基于特征参数提取和智能门限选取的特发性黄斑裂孔自动诊断算法可达到特发性黄斑裂孔智能诊断的目的,能够很好地应用于临床诊断。

Objective To diagnosis the idiopathic macular hole though the artificial intelligence(AI)using computer image feature recognition and feature parameter extraction algorithm.Methods We retrospectively reviewed 48 patients(48 eyes)fundus optical coherence tomography(OCT)images of patients with idiopathic macular hole and 48 images of healthy individuals in Xi’an Fourth Hospital from May 2018 to August 2018.AI learning was performed for the collected OCT images,and image processing and feature recognition technology was applied to extract the feature parameters which could differentiate OCT image of patients with idiopathic macular hole from healthy individuals.On this basis,the initial threshold for diagnosis of was idiopathic macular hole obtained.After randomly numbered fundus OCT images of 73 patients(73 eyes)with idiopathic macular hole and 51 healthy individuals(51 eyes)in Xi’an Fourth Hospital from 1 to 124,and the images was processed by the computer program,and then feature parameters were extracted from the processed images.Finally,we compared the feature parameters with the threshold.Results The feature parameter of OCT image of the normal individuals in the training samples was 9 and 23 for the patients with idiopathic macular hole,and the initial threshold was 16.After computer intelligent diagnosing,it was shown that the minimum value of the OCT feature parameters of 73 patients in 124 random images was 16.8,the maximum value was 27.5,and the average value was 23.4.All the feature parameters were larger than the threshold.The minimum value of OCT feature parameters of the 51 random images was 2.8,with the maximum value of 14.7,the average value of 8.3,and all the feature parameters were less than the threshold.73 OCT images of feature parameter larger than the threshold were all patients with idiopathic macular hole and 51 OCT images feature parameter smaller than the threshold were all healthy individuals.There was no statistical difference between the computer diagnosis and the ophthalmologist diagnosis(P=0.551).Conclusion The intelligent diagnosis algorithm based on feature parameter extraction and intelligent threshold selection can help diagnose idiopathic macular hole and can be well applied to clinical diagnosis.