A New Technique For Retinal Disease Treatment

Structured DR Image Database Construction and accurate annotation: retrospective (from January 1, 2016 to the day of ethical review) and prospective (from the day of ethical review to December 31, 2023) collected FFA and other multimodal image data. Several ophthalmologists and senior experts of fundus diseases made diagnostic evaluation on each image of each patient and completed the accurate grading diagnosis of the data Finally, a structured Dr database was established step by step. This paper uses the theory of computer vision to quantify the quality distortion of FFA image, guides the computer to configure the existing image enhancement and noise reduction algorithms adaptively, and completes the preprocessing of fundus image data.

Construction of Dr intelligent grading diagnosis system based on fundus image: firstly, the fundus image is used as the fundus data training database, and according to the international clinical Dr grading diagnosis standard, many doctors mark the fundus image accurately. International clinical Dr grading criteria: grade 0, no obvious retinal abnormalities; grade 1, only microangioma; grade 2, more severe than microangioma, but less severe than severe; grade 3, four quadrants, each quadrant has more than 20 retinal hemorrhage, more than two quadrants have definite venous beads, more than one quadrant has obvious Irma, no signs of proliferative retinopathy; grade 4, neovascularization, vitreous hemorrhage Volume blood, pre retinal hemorrhage. On the basis of Dr grading intelligent diagnosis standard, convolution neural network is constructed to train and grade fundus images. After repeating this process many times for each image in the training set of fundus images, the deep learning system learns how to classify all the data in the training set to accurately diagnose the fundus images.

Convolution neural network construction for FFA image focus area: the convolution neural network of deep learning is composed of millions of parameters, which is used to train and perform given tasks. The output generated by each linear convolution operation is regularized by nonlinear activation function, combined with the dimensionality reduction of pooling layer and full connection layer, so that the optimization process of deep neural network not only overcomes the gradient dispersion, but also helps to generate features similar to the hierarchical perception mechanism of human neural cells to visual signals. The FFA image is used as the fundus data training database. Based on the accurate labeling of the lesion area (no perfusion area, microangioma area and leakage area), the FFA image needs to be treated for the intelligent recognition of the lesion area. In the training process, the parameters of the neural network are initially set to random values. Then, for each image, the results given by the function are compared with the known results of the training set to optimize the parameters of the function. After repeating this process many times for each image in the training data set, the deep learning system learned how to classify all the data in the training set to accurately predict the Dr lesions on FFA images.

Construction of intelligent fundus laser navigation model based on FFA image and fundus image registration: the Dr lesion intelligent recognition system on the above FFA image accurately identifies the areas that need fundus laser treatment, helps doctors determine the lesions that need to be treated, and based on the image matching of machine learning, provides the registration image of fundus image and FFA combination, which is set according to the location and size information of the lesion area According to the matching retinal diameter and the arrangement of different laser spots, the personalized laser treatment scheme is formulated, and the intelligent fundus laser treatment guidance model is constructed.