| Abstrakt: |
With the progressive growth of artificial intelligence technologies, applying deep learning methods for medical imaging have successfully solved various medical imaging problems with high accuracy, efficiency and stability, particularly in the field of ophthalmology. In this vein, there is practically an urgent need to develop new tools by relying on artificial intelligence to early detect keratoconus and to lately prevent its progression as a disease. Two data augmentation methods, known as SyntEyes models and GANs networks, are applied to generate synthetic corneal topographic maps and increase data availability. Accordingly, in this paper, seven different deep learning models, based on CNN architecture, and which allow an efficient classification of corneal topographic and maps, are introduced. Thus, accuracy ranges from 95.31% to 99.74%, recall between 98.71% and 95.74% and lastly precision between 99.10% and 93.42%. The findings show that the notably customised classic CNN model outperforms the two hybrid models named CNN-SVM and CNN-LSTM and the four transfer learning models, known as VGG19, Xception, ResNet 50 and MobileNetV2, not only with respect to accuracy, recall, precision and F1 score, but also computation time and model complexity. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink