Improving Regularization of Deep Learning Models in Fundus Analysis

With the advancement of deep learning in many computer vision applications, the development of computer-aided diagnosis systems has been focused on recently. The deep learning approaches for glaucoma assessment through fundus analysis have shown extremely high performance, especially for early detection, in many studies. In addition, researchers also found that the deep learning models detect not only the subtle changes that reflect RNFL thinning but also the morphological features that are present outside the optic disc in fundus photography. It suggests that the deep learning models have great potential for extracting detailed morphologies which may not be aware by human’s vision. However, the deep learning models also suffered from the problem of poor regularization when dealing with cross-institutional datasets. That is, the model’s performance dropped dramatically on the input images from different institutions. The feature extraction by deep learning models can detect detailed characteristics for distinguishing different categories; nonetheless these detailed features also cause the overfitting problem for cross-institutional data. Therefore, in this study, we aim to verify our proposed hypothesis by the application of fundus analysis for glaucoma assessment using deep learning approaches. The features with better regularization, called robust features, were selected to reduce the influences from the non-robust features that overfit the training dataset from a specific source. With feature selection, the performance of the glaucoma assessment cross-institutional dataset can be improved from 62.7% to 81% in accuracy.