Research And Implemention Of Deep-Learning Based Method For Dry-Eye Detection

In this era of advanced and smart technologies,people are having more screen time with their various digital devices,or even get addicted.With their eyes overused,plus the effect of other external environmental factors(e.g.,air pollution,air conditioners,etc.)on the tear film stability,dry eye has become a globally prevalent eye disease.As a comprehensive disease with multiple pathogens,dry eye has complicated and costly diagnosis process.While a convenient and accurate diagnosis method is still urgently needed,there is few research and application on deep-learning based dry-eye detection.Thus,in this thesis,research has been carried out on dry-eye detection with deep learning.The main contents are as follows:(1)To break the number limits of manual annotation of meibomian gland,a Widthaware Multi-supervised learning model for meibomian gland segmentation is proposed in this thesis.In this model,as data pre-processing for meibomian gland segmentation,the tarsus is firstly extracted from the infrared image of meibomian glands and get angle-corrected.Further,a width-aware surrogate task is designed to dig into the potential features of unlabeled meibomian gland images and optimize the segmentation accuracy with multi-supervised learning strategy.(2)To deal with the lack of severity classification of dry eye in previous work,a multimodal feature fusion dry-eye detection classification model based on ocular redness and meibomian gland is proposed.The model uses a meibomian gland segmentation based transfer learning method to alleviate the overfitting problem of deep convolutional networks due to data volume limit.Thereby,a feature fusion component manages to combine hand-craft features and deep-learning features.Finally,a cost-sensitive loss function is used to capture the underlying connections in labels of dry-eye severity,which improves the accuracy of dry-eye detection with further refinement of severity judgement.(3)Based on the models and method mentioned above,a dry-eye detection prototype system was designed and implemented.The system not only lowers the requirement of the user’s professional skills with automatically generated visual reports for diagnosis assistance,but also provides a convenient means of self-examination and outputs accurate diagnostic suggestions based on deep-learning,during which only the image capture requires manual intervention.In summary,the thesis focuses on the research and implementation of deep-learning based methods for dry-eye detection,which has significant implications for further development of related study in this field.