Construct Grading Model For Diabetic Retinopathy Based On Deep Learning

Objective To construct a diabetic retinopathy assisted diagnosis model based on the principle of deep neural network by grading the fundus images of diabetic retinopathy,providing a convenient and fast method for the screening of this type of disease in the population.Methods The image enhancement technique was used to pre-process the fundus image data,and the deep convolutional neural network Inceptionv4 and SENet154 were optimised using Focal Loss,cosine annealing,weighted random sampling and image Gaussian filtering hybrid weighting methods to construct a grading model for diabetic retinopathy fundus images,and the model was validated using published Messidor data.The model was validated using fundus images of diabetic retinopathy and macular edema from the published Messidor dataset,and finally,the activation-like map was used to mark the lesion areas for visualization of the grading to assist doctors in rapid diagnosis.Results 1.In the diabetic retinopathy grading model,the AUCs of the model on the test set for classes 0,1,2,and 3 were 0.925(95% CI: 0.898-0.952),0.738(95% CI: 0.711-0.764),0.873(95% CI: 0.848-0.898),0.977(95% CI: 0.957-0.997);sensitivities of 0.750,0.435,0.662,0.842;specificities of 0.893,0.869,0.895,0.958;and Jorden The adjustment of the cosine annealing dynamic learning rate improved the model classification performance most significantly,the weighted random sampling improved the AUC of hard-to-learn samples in the unbalanced dataset more compared to other methods,and the model output class activation maps accurately depicted the suspected lesion regions.2.In the diabetic macular edema risk class model,the optimisation methods constructed based on the diabetic retinopathy model,multiple optimisation methods were overlaid with class 0,1 and 2 AUCs of 0.965(95% CI: 0.948-0.981),0.881(95% CI:0.852-0.909),0.963(95% CI: 0.941-0.985);sensitivities of 0.976,0.545,0.822;specificities of 0.821,0.979,0.981;and Jordans of 0.643,0.304,0.557,respectively.Conclusion We have constructed a deep learning-based risk classification model for diabetic retinopathy and diabetic macular edema with automatic classification of fundus images,and achieved high AUC values.In the case of small sample size and class imbalance in the dataset,the model can effectively improve the classification performance of difficult-to-learn classes by superimposing cosine annealing learning rate,Focal Loss function,Ben’s image enhancement and weighted random sampling method.At the same time,the model draws class activation maps to accurately locate suspicious lesion areas,which are visualized in the form of heat maps,and can visually assist physicians to facilitate quick diagnosis.