Fine-grained Keratitis Image Classification Based On Convolutional Neural Network

As the rapid development of deep learning represented by convolutional neural network,there are several important breakthroughs in image classification and other computer vison problem.These breakthroughs will bring revolutionary changes for some medical application relied on the medical images.Take keratitis for example,keratitis is one of the most common eye disease and believed to be an important cause of infectious blindness.In therapy,the ophthalmologist uses slit lamp to take eye images that help them get first diagnosis.However,keratitis has various kinds and complex condition,it causes that the ophthalmologist has very low accuracy in the first diagnosis.Thus,this dissertation mainly work as follows:(1)This dissertation reviews the development of hand-crafted feature and deep learning in image classification.It combines the keratitis diagnosis with fine-grained image classification and proposes an image-level convolutional neural network model.It uses whole image feature to classify images and uses focal loss to solve the problem that the dataset is unbalanced and difficult to train.(2)According to the characteristics that keratitis images has high similarity,this dissertation proposes a patch-level model that focus on the local feature.It proposes three unsupervised patch generating and selecting approaches.It also proposes a multitasking model for patch selecting and image classification.The alternative training method is applied for the training.(3)Inspired by the information retrieval and ranking,constructing triplets(anchor,positive sample,and negative sample)by images of one patient at different time,this dissertation proposes a model based on triplets ranking and makes the intra-class distance smaller than inter-class distance.(4)On the preprocessed keratitis dataset,this dissertation designs several experiments comparing with the base model and visualizes intermediate results to analyze the performance.(5)According to the characteristics of keratitis,this dissertation develops a labeling tool and labeling regulation for keratitis imageThe contributions of this dissertation are listed as follows:(1)This dissertation introduces find-grained image classification into keratitis image classification.Combining with the ophthalmologist's experience,it tries to use unsupervised method to extracting local features to improve the performance.The experiments has proven that.(2)According to dataset,it tries to use triplets ranking model to improve the performance of the image classification.The fine-grained visual concept recognition has been applied in 973 national project"ross-media Computing for Public Safety:Theory and Applications”.