Classification of Eye Diseases in Humans Using Vision Transformer Architecture Model

The eye is a very important sensory tool of living things, especially in humans. Where if the eye retina is damaged or affected by disease, especially cataract eye disease, the quality of the eye retina cannot be used properly. So, a digital approach is needed in order to recognize the eye is positive or negative cataract quickly, precisely, efficiently and accurately. So, this research makes an investigation of human eye disease, especially cataracts through eye retina images in the form of classification quickly, efficiently, and accurately. The Classification is performed using the Vision Transformer algorithm model of Conventional Neural Network(CNN) by utilizing six scenarios tested, namely using training datasets with Hyperparameter Epoch 25 and 50, Input Shape $224 \times 224 \times 3$ (RGB channel), batch size 32 and Optimizer (Adam, RMSprop, SGD) and Learning Rate 0.009 and 0.001 with a dataset of 1120 retinal images. The thing that makes this research use the vision transformer architecture model is in the architectural design and the way they process information visually more effectively and efficiently. While CNN rely on the use of convolutional layers to extract features hierarchically, Vision Transformers utilize self-attention mechanisms to capture global dependencies and relationships between image patches directly so that remote interactions in images are more effective than CNN. The results obtained from Hyperparameter Epoch 25 and 50, Input Shape $224 \times 224 \times 3$ (RGB channel), batch size 32 and Optimizer (Adam, RMSprop, SGD) and Learning Rate 0.009 and 0.001 with a dataset of 1120 retinal images. The results obtained from the Vision Transformer model that utilizes Hyperparameter get consecutive results with Accuracy 84.82% in scenario 1, Accuracy 89.29% scenario 2, Accuracy 89.29% scenario 3, Accuracy 90.18% scenario 4, Accuracy 92.86% scenario 5, Accuracy 92.86% scenario 6 and then evaluated to obtain the best accuracy, namely in Accuracy 92.86%, Precision 100%, Recall 85.72%, F1-score 92.31% in scenario 6 with validation los 0.2504.