Figure 1.

Deep learning framework for the diagnosis of IK based on slit-lamp images. (a) Data set establishment. Slit-lamp images were collected and annotated with etiological diagnosis label. Then labeled images were splitted into training, testing and external validation set, (b) Image pre-processing. After the original input images were resized to same height and width, image normalization, histogram equalization, and augmentation were applied to our data set successively, (c) Model training. Nine representative image classification networks were implemented with training data set. Models with high accuracy were combined using model blending technique to further improve the performance and (d) Model validation. Ten-fold cross-validation, t-SNE, ROC curves, confusion matrix, Grad-CAM visualization as well as comparation to human experts were used to assess the application potential of final model.