Research On Cataracts Image Grading Based On Deep Learning

The correct diagnosis of cataracts is the important premise of the reasonable treatment.However,the increasing cataract patients burdens the doctor heavily and makes it difficult to complete the diagnosis efficiently.Meanwhile,the existing methods of manual diagnosis are highly subjective and easily cause divergence in diagnosis.Therefore,the current cataract diagnosis process urgently requires an efficient,convenient and high-precision computer-aided diagnosis method.Although the artificial intelligence(AI)algorithms represented by depth learning gradually flourishes in the image field,they still need massive samples to improve the accuracy.Due to the low quality and small scale of cataract data caused by the lack of medical resources,AI technology has always been limited for the grading of cataract images.At the same time,the huge computation desire of deep network also makes it difficult to be promoted in practical applications.This thesis presents a method for grading cataract images based on deep learning,which can achieve the task of grading prediction and treatment advising for slit-lamp cataract images.An intelligent diagnosis software platform is further implemented,which is convenient for hospitals to deploy.The results show that the proposed method can effectively solve the impact of poor-quality and small-size dataset and facilitate the application and promotion of AI technology in the field of computer-aided diagnosis.The main research work of this thesis is as follows:(1)According to the characteristics of cataract,a classification model based on image classification was established.Firstly,the qualities of the samples are improved by the methods of generative adversarial network and supervised clipping on the original image data.Then,a step-wise fine-tuning method is proposed to steadily improve the training effect on small-size data by combining data enhancement and transfer learning stage by stage.ResNet based classification model achieves mAP of 84.39% and 97.46% from the perspective of grades and advising,respectively.(2)In order to exclude the effect of non-lesion areas on the classification effect,the target detection framework is introduced to limit the attention area of the classification model.Based on the original classification model,RetinaNet is further trained and optimized by dimension clusters and multi-model ensembles.The mAP indexes of grades and advising are increased to 84.4% and 97.5%,respectively.(3)In order to reduce the complexity of deep network model and make it adapt to small-size sample learning,we propose a method for optimizing model structure based on tensor decomposition.The variational Bayes matrix factorization(VBMF)is used to solve the Tucker factorization rank selection problem of the kernel tensor.After layer-bylayer compression and fine-tuning of the entire network,21% operation acceleration and 76.6% parameter compression can be achieved under the condition of 5.5% accuracy performance loss.(4)In order to facilitate the promotion of computer-aided diagnostic methods,medical image processing and intelligent diagnosis(MIPID)software platforms are further developed to organize the proposed method and migrate them to the low-profile computers.Eventually,it occupies only 2GB memory of the E5-2620 machine to achieve the calculation speed of 1.7RPS.