Abstract
The objectives of this study are to construct a deep convolutional neural network (DCNN) model to diagnose and classify meibomian gland dysfunction (MGD) based on the in vivo confocal microscope (IVCM) images and to evaluate the performance of the DCNN model and its auxiliary significance for clinical diagnosis and treatment. We extracted 6643 IVCM images from the three hospitals' IVCM database as the training set for the DCNN model and 1661 IVCM images from the other two hospitals' IVCM database as the test set to examine the performance of the model. Construction of the DCNN model was performed using DenseNet-169. The results of MGD classifications by three ophthalmologists were used to calculate the area under the receiver operating characteristic curve (AUROC), accuracy, precision, recall, true negative rate (TNR), true positive rate (TPR), and false positive rate (FPR) of the model. The deep learning (DL) was used to build the model to identify the IVCM images. Model accuracy and loss tests showed that the DCNN model had high accuracy, low loss, and no large fluctuations at an epoch of 175, indicating that DenseNet-169 could enable the dichotomization to proceed stably. The accuracy of each classification of the test set was above 90%, which was highly consistent with the ophthalmologists' diagnosis. The precision of the groups in each classification was more than 90%, or even close to 100%, except for the meibomian gland atrophy with obstruction group in the fifth classification. The recall ranged from 0.8728 to 0.9981, and the FPR was low in the screening and classification diagnoses. The application of DCNN can achieve accurate classification and diagnosis of MGD through IVCM images and has great potential during medical procedures.