Detection of glaucoma using three-stage training with EfficientNet

Abstract

This paper sets forth a methodology that is based on three-stage-training of a state-of-the-art network architecture previously trained on Imagenet, and iteratively finetuned in three steps; freezing first all layers, then re-training a specific number of them and finally training all the architecture from scratch, to achieve a system with high accuracy and reliability. To determine the performance of our technique a dataset consisting of 17.070 color cropped samples of fundus images, and that includes two classes, normal and abnormal, is used. Extensive evaluations using baselines models (VGG16, InceptionV3 and Resnet50) are carried out, in addition to thorough experimentation with the proposed pipeline using variants of EfficientNet and EfficientNetV2. The training procedure is described accurately, putting emphasis on the number of parameters trained, the confusion matrices (with analysis of false positives and false negatives), accuracy, and F1-score obtained at each stage of the proposed methodology. The results achieved show that the intelligent system presented for the task at hand is reliable, presents high precision, its predictions are consistent and the number of parameters needed to train are low compared to other alternatives.