A SELECTIVE GAN-BASED DATA AUGMENTATION METHOD FOR MEDICAL IMAGE CLASSIFICATION

Abstract

Although deep learning has significantly improved classification tasks, the available medical image datasets are often inadequate and imbalanced, hindering the development of accurate models for disease diagnosis and treatment.

In this research, a selective GAN-based data augmentation method for imbalanced classification tasks is proposed. The method involves two selection steps based on label certainty and triplet loss to control the quality of synthetic samples generated, and make the dataset more balanced by generating more synthetic samples for minority classes.

The research evaluates the method on the HyperKvasir dataset. Eventually, we achieved better classification performance than the baseline method, with an accuracy of 86.80 %, a precision of 76.64%, a recall of 77.39%, a F1-score of 76.92% and an AUC-ROC of 0.8966.

Overall, the results demonstrated significant improvements in the classification performance, particularly for the minority classes, which shows potential in enhancing the classification performance of imbalanced medical image datasets.