| Summary: | Background. One of the serious consequences of diabetes is the negative impact
on the visual system. The aim of the work is to review the sources devoted to the task
of diagnosing diabetic retinopathy from eye images using neural networks. Materials
and methods. The use of modern methods, approaches and algorithms at such stages as data
collection and preparation, data preprocessing, image recognition task, transfer training,
comparison of methods, ensembles of models, system development is considered. Possible
promising further steps in the future research are outlined. Results. In the process of analyzing
publications on methods of diagnosing diabetic retinopathy from eye images based on
neural networks, the following areas were identified to improve the existing results: increasing
image data sets, image preprocessing methods, interpretation of the neural network
model, computational power of algorithms for implementation on mobile devices, classification
and segmentation tasks of eye lesions, false-negative and false-positive diagnoses,
ensembles of models, the use of recurrent and capsule neural networks. Conclusions. Based
on the results of the study, the directions for improving the achievements in the task of eye
image recognition for the diagnosis of diabetic retinopathy were identified.
|
|---|