Research And Application Of Fundus Retinal Image Caption Based On Deep Networks

This thesis is mainly about an design and implementation of a deep learning model,which can automatically classifies and labels an retinal image without any human work,what's more,this model is designed to be able to generate description of the input image.This work mainly aims on the acquisition,management,and analysis of the data set of fundus retinal images in the field of biological imaging,and strives to exploit and improve several key points of automatic fundus retinal image classifications and description models,which are mainly the following aspects:1.A data augmentation method during the procedure of pre-processing of fundus retinal images is analyzed and optimized.This paper introduces this improved AIFT method,which adopts an outlier detection method based on normal distribution to filtrate these newly generated candidate samples,to deal with the issue of lack of fundus retinal image data sets.Experiments are conducted to compare these two methods' performance on the fundus retinal images,results demonstrate that the improved AIFT method is better at fastening the training process of deep learning model,lessening the time of training as well as boosting model's stability during training.2.Three different activation functions are introduced to compare with the original squashing activation function in Hinton's paper,which was present on 2017.Several contrast tests are executed to compare these 4 functions' performances on the Mnist dataset.Results certify that except sigmoid function,whose performance is the same as the original squashing function,the other two newly introduced functions both get better performances than the one of the primitive paper.Following is the training of the CapsNet with newly proposed squashing activation function,whose dataset is the 15000 fundus retinal images,this model is designed to classify a 6-classification problem.3.Employing the former proposed two methods: an improved AIFT method and the Caps Net based on the newly introduced squashing activation function,this paper designs and implements an automatic model generating description of fundus retinal images.Firstly,a pre-trained darknet model of Yolov2 is fine-tuned on the data set of fundus retinal images,the last two layers of this darknet is replaced with capsule layers,then a RNN is adopted to train the model,finally this new model is used to automatically generate image description on the fundus retinal images.This thesis' s emphasis is on how to employ the now mature and widely-used deep learning neural network to automatically deal with fundus retinal images,so as to promote the real-world application of fundus retinal images of the diagnosis,treatment of diseases,and to assist doctors in disease-diagnosing,decision-making and so on.