Author: Juan Donate-Lopez
Co-authors: ROdrigo Abreu-Gonzalez, Natan Rodriguez.Martin, Joseph Blair, Sandro de Zanet, Jose Julio Rodrigo, Carlos Bermudez-Perez
Abstract
PurposeTo obtain a screening algorithm whose performance is equivalent to that of medical specialists in the diagnosis of DR and that can be used in real clinical practice.
Methods
Two deep convolution networks with different architectures (ResNet50 and EfficientNet B2) have been trained and combined by means of a "perceptron" type neural network. This training has been carried out on a data set previously divided into three subsets (training, validation and test), each one of them subjected to an automatic quality control carried out by a neural network trained for the detection of low quality images. and subsequently reviewed by specialists. In turn, the subset of training data has been subjected to "data augmentation" techniques that introduce variability as well as different degrees of noise so that the algorithm is capable of working in conditions equivalent to those that doctors operate in clinical practice.
Results
A set of 146,414 images was used for the training phase, 21,517 images for the validation phase, and 10,000 for the test phase. An area under the ROC curve of 0.969, a sensitivity of 0.816 and a specificity of 0.942 were obtainde.
Conclusions
The algorithm presents the necessary performance to be subjected to a field test through a multicenter clinical study, already in progress. This study is the step prior to obtaining certification as a medical grade product suitable for use in real clinical practice. These algorithms are far from a perfect result to the same extent that the data used for their training is far from a perfect label, but our tests prove that they can perform in an equivalent way to specialists and can become a key tool for them in the performance of their work, raising the standard of ophthalmological management of these patients.
Disclosure:
The author reports no conflicts of interest in this work.