Towards automated eye cancer classification via VGG and ResNet networks using transfer learning

Abstract

Complex tasks such as disease diagnosis or semantic segmentation are now becoming easier to tackle in part due to increasing advances in computing and storage. This study provides an exploratory approach utilising convolutional neural networks (CNNs) to detect ocular abnormalities with an illustrative case of uveal melanoma (UM), a type of ocular cancer. In previous studies UM has been researched employing different computational techniques focusing on discriminative features using fuzzy systems, neural networks, and adaptive neuro-fuzzy systems. However, given the inheritable nature of the problem, it was decided to use CNNs with transfer learning as a promising alternative to improve the accuracy of the results. As for the main contributions, the results outperforms different state-of-the-art computational algorithms studied to detect UM, in particular improvements in sensitivity, precision and accuracy, achieving 99%, 98% and 99%, respectively. Besides, two algorithms were implemented to reduce the bias of the dataset: a data augmentation algorithm using the Gabor filter, and an algorithm to remove light spots using Navier–Stokes approach.