Research On Intelligent Detection Algorithm And Application Of Fundus Diseases Based On Mobile Device

Currently,ophthalmic diagnosis observation and analysis based on fundus image disease detection is often labor-intensive and subjective in judgment,and there is a lack of relevant expert diagnosis for relatively underdeveloped medical regions.It is important to develop and design a tool that can help medical practitioners to screen fundus diseases and assist professional doctors in diagnosis by using artificial intelligence technology combined with fundus image datasets.Based on this background,combined with the rapid development of deep learning technology and the concept of intelligent medicine in recent years,this thesis focuses on the automatic classification of fundus image diseases and the development of APP applications for mobile devices.Firstly,with the goal of achieving automatic classification of five categories of fundus images,namely diabetic retinopathy(DR),glaucoma(GLC),cataract(CAT),age-related macular degeneration(ARMD),and normal fundus(NL),on a mobile device,the dataset for this study was formed by combining five categories of fundus images from several large scale public datasets in this field at present.The final obtained dataset is also subjected to image pre-processing,and the training image dataset is expanded using changing color parameters such as training image brightness,random flip,etc.Secondly,based on the application deployment requirements at the mobile device side,the LM＿MBANet convolutional network model which is partially improved based on the backbone network PPLCNet by incorporating multi-scale feature transform and bilinear channel attention module and introducing h-swish activation function suitable for mobile side computing is proposed.For the initial part of the network setup,a448×448 input image size is used and a feature extraction convolutional layer is incorporated.At the same time,the training hyperparameter experiments are designed,and the experimental results show that the learning rate and the drop rate of 0.2 are dynamically adjusted using Adam optimizer,and the randomly deactivated neurons achieves the best training effect.The comparison experiments show that the final obtained model has only a small increase in the number of parameters,with the number still about 7.7M,and only increases the amount of floating point computation by about300 M relative to the backbone network,but each module proposed is proved to be effective in increasing the classification accuracy.By ablation experiments,all modules are effective in increasing classification accuracy and fundus disease recognition,and the highest comparison accuracy of 98.61% and 87.5% is obtained on our-Fiveclass Dataset and Fourclass Dataset respectively.The accuracy of the overall model for fundus disease classification is better than that of various lightweight and nonlightweight networks in the same period.Finally,the optimized LM＿MBANet model and its weight file are deployed as the core of the classification system.The results of the model running time and frame rate show that the model can run stably and smoothly on the Android-based mobile devices.Overall,this work optimizes the PPLCNet model as the backbone network after preprocessing the extracted and combined dataset images to obtain a lightweight network model LM＿MBANet that can be run on mobile devices,and finally deploys the model to Android mobile devices to form a runnable application.The visualization results show that the APP achieves the expected purpose of running the intelligent detection algorithm for fundus diseases based on deep learning and mobile devices.