Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network

The study utilized in vivo confocal microscopy (IVCM) to observe meibomian glands and employed a ResNet34 deep learning network model to classify 12,630 images into six categories related to meibomian gland dysfunction (MGD). The model was trained on 70% of these images, with the remaining 30% and an additional 12,889 images used for validation. The model demonstrated strong performance, with area under the receiver operating characteristic curves (AUROCs) exceeding 0.95. This indicated the potential for the model to aid in the automatic classification and diagnosis of MGD, supporting clinical diagnosis and disease screening.