Research On Retinal Multi-disease Classification Based On Deep Learning

With the increasing prevalence of eye diseases caused by the widespread use of electronic products and unhealthy eye habits,diabetic retinopathy,glaucoma,cataracts,age-related macular degeneration,hypertension,and high myopia have become the leading causes of retinal diseases.Currently,screening for retinal diseases mainly relies on manual diagnosis by professional ophthalmologists,which is time-consuming,laborintensive,and has a low success rate.Most intelligent auxiliary diagnostic systems focus on detecting a single retinal disease.Therefore,this paper proposes a joint diagnosis algorithm for retinal multi-disease classification,considering the complexity of retinal images and the importance of retinal vascular structures.The proposed algorithm can identify and classify the six common retinal diseases simultaneously,achieving good results and optimizing the screening process for retinal diseases.It effectively improves the diagnostic efficiency and reduces the diagnosis cost.This paper mainly includes the following contents:(1)Completed the dataset analysis and image processing of retinal fundus images.Summarize and collect multiple datasets,and normalize and unify them.Clear abnormal images and remove interference labels.Handle data imbalance,expand the number of minority samples in the dataset using methods such as data augmentation,resampling,and class weight adjustment.Preprocess retinal images to remove redundant images,enlarge disease details,and enhance visual effects and recognition performance.(2)Completed the retinal fundus image vascular structure segmentation.Two methods for retinal fundus image blood vessel segmentation were designed,including a morphology-based method using filtering,morphology processing,and threshold segmentation,and a deep learning-based method(BCL-Net model)incorporating bidirectional convolutional LSTM modules,densely connected convolutional modules,BN modules,and attention mechanisms into the standard U-Net model.Experimental results showed that the BCL-Net model improved the accuracy by 11.87% and 1.56%compared to traditional algorithms and the standard U-Net model,respectively.(3)Completed the joint diagnosis of six retinal diseases based on retinal fundus images.A multi-disease joint diagnosis algorithm(FCML algorithm)based on fourchannel multi-task learning was proposed,which incorporated four-channel inputs,multitask learning,and multi-model fusion modules,and optimized the algorithm performance with attention mechanisms,binocular supervised learning,and model randomization.Several evaluation metrics were integrated,and VGG-16,Inception-V3,Res Net-50,and Efficient Net-B4 were selected as the basic models for subsequent algorithms.Algorithm ablation experiments were designed to prove the effectiveness of the model optimization strategies and the important auxiliary role of the retinal blood vessel structure.After adding the SE module,the accuracy of the binary classification model increased from83.45% to 90.31%.The FCML algorithm accuracy was improved by 1.11% and 1.71%,respectively,after adding different model optimization strategies.However,the accuracy of the algorithm decreased by 5.21% after removing the blood vessel image input.Algorithm comparison experiments were designed to prove that this algorithm has reached the level of first-class models in the ODIR competition and has exceeded similar models in accuracy,sensitivity,and other indicators.