Eye diseases diagnosis using deep learning and multimodal medical eye imaging

The present study carries out an empirical evaluation and comparison of the seven most recent deep Convolutional Neural Network (CNN) techniques (VGG19, DenseNet121, InceptionV3, InceptionResNetV2, Xception, ResNet50V2, MobileNetV2) for eye disease classification into normal and diseased using mono-modality and late fusion multimodality over three datasets. All empirical evaluations were carried out using: (1) six classification metrics, (2) the Scott-Knott Effect Size Difference (SK-ESD) statistical test, (3) Borda count method, and (4) publicly available datasets: DHS (combining three datasets: DRIVE, STARE, and HRF), FFA, and Macula. The results showed that DenseNet121 and ResNet50V2 were the top performing and less sensitive techniques on the three datasets using mono-modality with accuracy values of 99.57% and 99.51% respectively. As for late fusion techniques, they outperformed mono-modality across the three datasets, regardless of the mono-modality used. Moreover, ResNet50V2 late fusion was the best late fusion technique and scored 100% in accuracy across all three datasets. Additionally, our proposed ResNet50V2 late fusion model trained on the Macula dataset outperforms current state-of-the-art models trained on more than one eye disease (accuracy of: proposed ResNet50V2 late fusion = 100%, New CNN = 81.33%), and is similar to best ranking feature-level fusion one eye disease model with accuracy equal to 100%.