Logarithmic Progressive-SMOTE: Oversampling Minorities in Retinal Fundus Multi-disease Image Dataset

Abstract

Multiple retinal diseases co-occur more frequently. It is extremely challenging to effectively diagnose several diseases in an image and provide appropriate treatment for those diseases with high accuracy. The performance of such an imbalanced dataset, which has minority classes, can be enhanced. The Retinal Fundus Multi-Disease Image Dataset (RFMiD) is a collection of 3200 Multi-labeled Imbalanced Dataset (MLID) marked with 46 different disease labels. The suggested Logarithmic Progressive Synthetic Minority Oversampling Technique (LP-SMOTE) is intended to oversample the minority classes in imbalanced dataset. The Imbalance Ratio Per Label (IRPL) and Mean Imbalance Ratio (MeanIR) are used and assessed to distinguish the minority and majority classes. The logarithmic progressive sampling per label is applied to minority classes. The Oversampling is implemented to level up the samples in minority classes. The proposed technique improved the overall minority class samples by 5.4 times. After employing LP-SMOTE technique, the overall validation accuracy and test accuracy is improved by 2.81% and 4.53% respectively on VGG19 pre-trained model.