神经心理发育障碍儿童静态眼睛照片深度学习的病例对照初步研究

Still eye photos in children with neuropsychological developmental disorders by deep learning:A pilot case-control study

背景目前对儿童神经心理发育障碍的诊断主要靠临床表现,尚缺乏客观的生物学指标,一些神经心理发育障碍儿童的眼睛特征可能具有重要的临床价值。目的评价静态眼睛图像是否可以作为儿童神经心理发育障碍的一个潜在的筛查测试。设计病例对照研究。方法双眼图像采集应用Google人脸图片搜索引擎进行图片搜索,以“autism and child/autistic child”和“normal child/healthy child”关键词搜索到的图片分别归为神经心理发育障碍组和对照组。人工删除完全重复的照片、卡通照片、有面部畸形的照片、没有儿童脸的照片和有多个儿童脸部的照片;应用图像工具识别和挖取每张儿童人脸照片中的双眼图像。生成的眼睛图像调整至128×28分辨率(8位灰度)。卷积神经网络(CNN)是基于一台安装了Kera、Scipy和Python成像库(PIL)的Windows 7计算机。采用二进制交叉熵损失函数和RMSprop算法优化器对模型进行训练。将图像数据分为训练集、验证集和测试集。神经心理发育障碍组测试分数>0.5为分类正确,对照组≤0.5为分类正确,训练25次停止。以分组检索结果为“金标准”,以深度学习训练模型分组为待测标准。计算准确率和损失率。主要结局指标根据深度学习训练模型两组特征生成的眼睛图像。结果通过双眼图像采集到826张儿童眼睛图像,神经心理发育障碍组413张,对照组413张。训练集736张(89.1%),验证集44张(5.3%),测试集46张(5.6%)。训练集和验证集准确率随训练次数增加呈升高趋势,训练集和验证集损失率随训练次数增加而逐渐下降。神经心理发育障碍组和对照组测试分数分别为0.66±0.20和0.20±0.15,差异有统计学意义(t=9.03,P<0.001),正确分类例数分别为18/23例(78.3%)和22/23例(95.6%)。准确率为87.0%,敏感度为78.3%(95%CI:58.1%~90.3%),特异度为95.6%(95%CI:79.0%~99.2%)。深度学习训练模型ROC曲线显示,AUC=0.962。由CNN卷积图层的可视化生成神经心理发育障碍组和对照组的眼间距宽,像素数分别为90和70。结论深度学习训练模型可以获得神经心理发育障碍儿童的眼睛特征,这将有利于通过评价眼睛图片提高儿童神经心理发育障碍的早期筛查水平。

Background The diagnosis of neuropsychological developmental disorders in children mainly depends on clinical manifestations at present,and there is still a lack of objective biological indicators.The eye characteristics of some children with neuropsychological developmental disorders may have important clinical value.Objective To evaluate whether still eye photos can be used as a potential screening test for neuropsychological developmental disorders in children.Design Case-control study.Methods The Google image search engine was run with key words autism and child/autistic child/normal child/healthy child separately in order to select the sample images.Then the images were divided into two subsets,corresponding to population labeled with case and control respectively.Duplicated photos,cartoon photos,photos with facial deformity,photos without children's faces and photos with many children's faces hard to determine the research objects were deleted.Face tools identified and extracted the face out of each photograph.The resulting face images were scaled to 128×28 resolution(8-bit grayscale).Convolution neural network(CNN)was used in a machine installed with Keras,Scipy,Python imaging library(PIL).The binary cross entropy loss and rmsprop optimizer were used to train the model.The images were divided into training set,verification set and test set.The test score regarded as correct classification was more than 0.5 in case group and less than 0.5 or equal to 0.5 in the control group,and the training was stopped at 25 times.The grouping search results were regarded as the"gold standard",and the training group was regarded as the test standard to obtain the calculation accuracy and loss rate.Main outcome measures Eye images generated from two sets of features.Results Totally,826 eye images were collected,413 in the neuropsychological developmental disorders group and 413 in the control group.There were 736(89.1%)in the training set,44(5.3%)in the validation set,and 46(5.6%)in the test set.With increasing number of training sessions,the training and validation set accuracy rates showed an increasing trend while the training and validation loss rates gradually decreased.Test scores were 0.66±0.20 and 0.20±0.15,respectively in neuropsychological developmental disorder group and control group(t=9.03,P<0.001),and the number of correctly classified cases was 18/23(78.3%)and 22/23(95.6%).The mean accuracy of the two groups was 87.0%,the sensitivity was 78.3%(95%CI:58.1%~90.3%),and the specificity was 95.6%(95%CI:79.0%~99.2%).The receiver operating characteristic(ROC)curve of the trained model with deep learning showed the area under curve ROC(AUC)was 0.962.Eyes interval in the neuropsychological developmental disorder and control groups was generated by visualization of convolution layer of CNN,with the number of pixels being 90 and 70 in neuropsychological developmental disorder group and control group respectively.Conclusion Supervised deep learning can obtain eye features of children with neuropsychological developmental disorders,which may be helpful to improve the early screening of neuropsychological developmental disorders by evaluating eye photos.