目的：通过对患者临床数据进行数据挖掘分析建立针对小直径角膜的圆锥角膜智能化诊断模型。

方法：诊断性研究。共收集患者830例830眼，其中男338例338眼，女492例492眼，年龄14～36(平均23.19±5.71)岁，其中2020-01/2022-03在重庆南坪爱尔眼科医院已行角膜屈光手术患者731例731眼，2015-01/2022-03确诊圆锥角膜患者99例99眼。所有患者行Pentacam角膜地形图显示角膜直径≤11.1mm。由2位角膜科专家通过Pentacam地形图中Belin/Ambrósio增强扩张显示(BAD)系统将患者数据分类为正常角膜、可疑圆锥角膜、圆锥角膜。采用计算机随机采样方法随机筛选其中665例患者的数据作为训练集，另165例患者的数据作验证集。利用卷积神经网络(CNN)提取7个角膜参数特征，分别采用残差网络(ResNet， Residual Network)、Vision Transformer(ViT)及CNN+Transformer建立模型，通过交叉熵损失函数进行训练并采用样本交叉法验证模型的准确性，并采用受试者工作特征曲线评价模型的敏感度与特异度。

结果：ResNet、ViT和CNN+Transfermer模型诊断正常角膜和可疑圆锥角膜的准确率分别为85.57%、86.11%和86.54%，受试者工作特征曲线下面积(AUC)为0.823、0.830和0.842。诊断可疑圆锥角膜和圆锥角膜的准确率分别为97.22%、95.83%和98.61%，AUC分别为0.951、0.939和0.988。

结论：对于直径≤11.1mm的角膜，借助CNN+Transformer算法建立的数据模型对圆锥角膜有较高的准确率，可为早期筛查提供真实有效的指导作用。

METHODS: Diagnostic study. A total of 830 patients(830 eyes)were collected, including 338 male(338 eyes)and 492 female(492 eyes), with an average age of 14-36(23.19±5.71)years. Among them, 731 patients(731 eyes)had undergone corneal refractive surgery at Chongqing Nanping Aier Eye Hospital from January 2020 to March 2022, and 99 patients had a diagnosed keratoconus from January 2015 to March 2022. Corneal diameter ≤11.1 mm was measured by Pentacam in all patients. Two cornea specialists classified patients' data into normal corneas, suspect keratoconus, and keratoconus groups based on the Belin/Ambrósio enhanced ectasia display(BAD)system in Pentacam. The data of 665 patients were randomly selected as the training set and the other 165 patients as the validation set by computer random sampling method. Seven parametric corneal features were extracted by convolutional neural networks(CNN), and the models were built by Residual Network(ResNet), Vision Transformer(ViT), and CNN+Transformer, respectively. The diagnostic accuracy of models was verified by cross-entropy loss and cross-validation method. In addition, sensitivity and specificity were evaluated using receiver operating characteristic curve.

RESULTS: The accuracy of ResNet, ViT, and CNN+Transfermer for the diagnosis of normal cornea and suspect keratoconus was 85.57%, 86.11%, and 86.54% respectively, and the area under the receiver operating characteristic curve(AUC)was 0.823, 0.830 and 0.842 respectively. The accuracy of models for the diagnosis of suspect keratoconus and keratoconus was 97.22%, 95.83%, and 98.61%, respectively, and the AUC was 0.951, 0.939, and 0.988 respectively.

CONCLUSION: For corneas ≤11.1 mm in diameter, the data model established by CNN+Transformer has a high accuracy rate for classifying keratoconus, which provides real and effective guidance for early screening.