Class-oriented Method of Fundus Images Augmentation

Abstract

Innovative class-oriented method of fundus images augmentation is proposed. The advantages of this method compared to the standard method are given. The augmentation strategy for images with the existing signs of glaucoma, cataracts, diabetic retinopathy and a healthy eye is substantiated and selected according to the specifics of each class. Augmentation approximates real clinical variation. Resistance to precisely those variations that are characteristic of a specific disease has been improved. Sensitivity and specificity specifically to pathologies in medical images is increased. A neural network model was developed using the class-oriented method. GlobalAveragePooling layers, Dropout with rate 0.5, Dense with 1024 neurons, l2-regularization with a value of 0.001 and categorical crossentropy loss function, Dense classification layer with 4 neurons and softmax activation function were added to the EfficientNetB3 core network by learning methods transfer. Half of the layers of the base model were frozen. The model was compiled using an Adam compiler with an initial learning rate of 0.0001 and a cost feature of categorical crossentropy. In pre-processing, image sizes resized to 224 by 224. The normalization of images automatically occurred during the generation of data for training. The following callback functions were used to adjust the model learning: ModelCheckpoint — to save the best model, EarlyStopping — with to stop in the absence of improvement of the val accuracy metric during 15 eras, ReduceLROnPlateau — to reduce the learning rate by 3 times when stagnating. As a result of the training, high indicators of metrics were obtained. The trained model was compressed by quantification for later use on mobile and disabled devices. The proposed approach makes it possible to increase the overall accuracy and robustness of the neural network, to overcome the limitations of the traditional method.