Mirsalehi Maziar, Fassbind Benjamin, Streich Andreas, Langenbucher Achim
Department of Experimental Ophthalmology, Saarland University, Homburg, Germany.
Department of Computer Science, Eidgenössische Technische Hochschule, Zürich, Switzerland.
PLoS One. 2025 Sep 2;20(9):e0311036. doi: 10.1371/journal.pone.0311036. eCollection 2025.
Purpose Prediction of the ectasia screening index, an estimator provided by the Casia2 instrument for identifying keratoconus, from raw optical coherence tomography data using convolutional neural networks. Methods Three convolutional neural networks models (ResNet18, DenseNet121 and EfficientNetB0) were employed to predict the ectasia screening index. Mean absolute error was used as the performance metric for predicting the ectasia screening index by the adapted convolutional neural network models on the test set. Scans with an ectasia screening index value higher than a certain threshold were classified as Keratoconus, while the remaining scans were classified as Not Keratoconus. The architectures' performance was evaluated using metrics such as accuracy, sensitivity, specificity, positive predictive value and F1 score on data collected from patients examined at the eye clinic of the Homburg University Hospital. The raw data from the Casia2 instrument, in 3dv format, was converted into 16 images per examination of one eye. For the training, validation and testing phases, 3689, 1050 and 1078 scans (3dv files) were selected, respectively. Results In the prediction of the ectasia screening index, the mean absolute error values for the adapted ResNet18, the adapted DenseNet121 and the adapted EfficientNetB0, rounded to two decimal places, were 7.15, 6.64 and 5.86, respectively. In the classification task, the three networks yielded an accuracy of 94.80%, 95.27% and 95.83%, respectively; a sensitivity of 92.07%, 94.64% and 94.17%, respectively; a specificity of 96.61%, 95.69% and 96.92%, respectively; a positive predictive value of 94.72%, 93.55% and 95.28%, respectively; and a F1 score of 93.38%, 94.09% and 94.72%, respectively. Conclusions Our results show that the prediction of keratoconus based on the ectasia screening index values estimated from raw data outperforms previous approaches using processed data. adapted EfficientNetB0 outperformed both the other adapted models and those in state-of-the-art studies, with the highest accuracy and F1 score.
利用卷积神经网络从原始光学相干断层扫描数据预测由Casia2仪器提供的用于识别圆锥角膜的扩张筛查指数。方法：采用三种卷积神经网络模型（ResNet18、DenseNet121和EfficientNetB0）预测扩张筛查指数。平均绝对误差用作经适配的卷积神经网络模型在测试集上预测扩张筛查指数的性能指标。扩张筛查指数值高于某一阈值的扫描被分类为圆锥角膜，而其余扫描被分类为非圆锥角膜。使用诸如准确率、灵敏度、特异度、阳性预测值和F1分数等指标，对从洪堡大学医院眼科诊所检查的患者收集的数据评估这些架构的性能。来自Casia2仪器的3dv格式的原始数据，每只眼睛每次检查转换为16张图像。对于训练、验证和测试阶段，分别选择了3689、1050和1078次扫描（3dv文件）。结果：在扩张筛查指数的预测中，经适配的ResNet18、经适配的DenseNet121和经适配的EfficientNetB0的平均绝对误差值，四舍五入到两位小数，分别为7.15、6.64和5.86。在分类任务中，三个网络的准确率分别为94.80%、95.27%和95.83%；灵敏度分别为92.07%、94.64%和94.17%；特异度分别为96.61%、95.69%和96.92%；阳性预测值分别为94.72%、93.55%和95.28%；F1分数分别为93.38%、94.09%和94.72%。结论：我们的结果表明，基于从原始数据估计的扩张筛查指数值对圆锥角膜的预测优于先前使用处理后数据的方法。经适配的EfficientNetB0优于其他经适配的模型以及现有研究中的模型，具有最高的准确率和F1分数。
