A Combined ResNet50 - Restricted Boltzmann Machine for Multilabel Eye Diseases Classification

This research aims to perform multilabel classification of eye diseases using fundus retina images. The condition of the dataset used in this study includes both labeled and unlabeled datasets. To utilize this unlabeled dataset, the unlabeled dataset is transformed into a dataset with pseudo labels using a semi-supervised learning approach. ResNet50 is the backbone model to extract features from the original labeled dataset. Then, the features from this modeling will be used in the label spreading algorithm to label the unlabeled dataset, resulting in a dataset with pseudo labels, which is subsequently combined with the original labeled dataset, referred to as the combined dataset. The combined dataset will be used to train the next model, ResNet50, with the same experimental setup as the previous backbone model and the combination model of ResNet50 and Restricted Boltzmann Machine (RBM), hereafter referred to as the ResNet50-RBM model. The model evaluation results show that using the combined dataset with label spreading can improve the performance of ResNet50. ResNet50 with the original labeled dataset achieved an F1-Score of 81%, while ResNet50 with the combined dataset achieved an F1-Score of 93%. The combined dataset was also used to train the proposed ResNet50-RBM model, which achieved the best performance with an F1-Score of 96%, precision of 98%, and Recall of 95%. The results of this study indicate that the labelspreading method effectively enriches the dataset, and the ResNet50-RBM model can improve the performance of ResNet50 in performing multilabel classification of eye diseases.