Residual Attention Network for distinction between visible optic disc drusen and healthy optic discs

Abstract

In this study the authors propose a new solution for distinguish healthy cases and those with optic disc drusen (ODD) utilizing Residual Attention Network (RAN). This network architecture, which employs convolutional layers, integrates a diverse attention mechanism within its deep structure. In this study, an unique approach is adopted, involving the iterative division and subsequent recombination of a single image B-scan obtained using OCT-A. Overall, 116 images of ODD obtained using optical coherence tomography - angiography (OCT-A) have been analysed and compared to images of healthy optic discs. A sequence of trials was conducted, considering the random partitioning of data into training, validation, and test elements, with proportions of 60 %, 20 %, and 20 %, respectively. The minimum accuracy exceeded 86 %, while the maximum values were higher than 98 %. The accuracy for analyzing healthy cases and ones with ODD has gained very satisfactory results. Our study shows that RAN is a suitable tool for distinguishing between optic disc drusen and normal optic discs on the basis of OCT-A B-scans of the optic nerve head. Deep learning may be used as a possible solution to screen for patients with ODD.