Dealing with Data Imbalance in Exudate Detection of Retinal Fundus Image

Many medical image diagnostic datasets do not provide adequate training data for conventional learners because some classes are overrepresented while others are underrepresented. Imbalanced dataset complicates machine learning categorization, and predicting an outcome becomes harder when insufficient data for learning are there. The goal of this study is to use various data imbalance solutions to get balanced classes and enhance multi-class classification performance. Exudates are early signs of diabetic retinopathy, a condition that can lead to blindness in diabetics. A crucial and difficult diagnostic challenge is recognising exudates with good accuracy and sensitivity. We employed the exudate classification dataset, which is made up of publicly accessible fundus image dataset DIARETDB and is composed with colour, texture, sharpness, and intensity. This dataset is divided into three classes, each of which is significantly imbalanced. Data, algorithm and hybrid level approaches are applied to address this class disparity. The pre-processing stage employs a data-level method known as resampling, with resampled data categorised using a random forest classifier. Random oversampling (ROS), synthetic minority oversampling technique (SMOTE) and adaptive synthetic (ADASYN) methods are the data-level approaches employed in this study. Cost sensitive method with random forest and ensemble method as multi-layer stack (MLS) ensemble are evaluated as algorithm level. The hybrid-level strategy incorporates the benefits of both data-level and algorithm-level approaches. Exudate dataset is subjected to a comparative analysis using several class imbalance techniques. The proposed method improves class wise and overall classification performance.