Hyperspectral Image Augmentation Technique for the Classification of Diabetic Retinopathy

This paper leverages the advantages of Hyperspectral Imaging, which can show different images with distinct features highlighted in each spectrum. Amongst the domains of Artificial Intelligence (AI), the Medical Domain has significant challenges in constructing high-quality datasets due to various social, ethical, and legal issues. Furthermore, available constructed datasets struggle to escape from Long-Tailed (Data Imbalance) distributions. Various techniques such as Data Augmentation and Data Sampling have been proposed and employed to address these issues, yet a complete solution has not been presented. Furthermore, the primary drawback of existing Data Augmentation techniques is their difficulty in the proper incorporation of professional medical knowledge. Therefore, this paper proposes a new Augmentation method that utilizes AI to convert RGB Imaging into Hyperspectral Imaging for employment to an AI model. Through this experiment, the information on lesions that were previously not found in the original RGB Imaging were additively used in the training process, resulting in increased performance in different performance evaluation metrics, such as the F'1-Score, Accuracy, Sensitivity, ROC-AVC Score, etc. This confirmed that Data Augmentation through Hyperspectral Imaging can have a positive impact on resolving Data Imbalance in the Medical Domain. Furthermore, the extension of this research is expected to provide applications to a wider domain, other than the Medical Domain, that requires a solution to the Data Imbalance problem.