Analisis Pengaruh Strategi Augmentasi Data Terhadap Performa Model Hybrid Cnn–Swin Transformer dalam Klasifikasi Citra Mikroskopis Malaria

Abstract

Malaria diagnosis through microscopic images still faces challenges due to variations in the shape, size, and quality of blood cell images that can reduce the generalization ability of deep learning models. One approach to address this problem is data augmentation, but the effectiveness of various augmentation strategies on a hybrid Convolutional Neural Network (CNN) and Swin Transformer model has not been systematically compared. This study aims to compare the effect of five data augmentation strategies on the performance of the Hybrid CNN–Swin Transformer model in classifying malaria microscopic images. The NIH Malaria dataset consisting of 27,558 blood cell images was divided using the stratified split method with a ratio of 80:10:10. The model was trained using five augmentation strategies, namely Baseline (A0), Basic (A1), Geometric (A2), Advanced (A3), and RandAugment (A4), then evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and Train-Val Gap. All strategies produced accuracy above 96% and ROC-AUC above 99%. Strategy A3 provided the highest classification performance with an accuracy of 97.79%, an F1-score of 97.78%, and an ROC-AUC of 99.52%, while strategy A2 showed the most stable generalization ability based on the Train-Val Gap value. The results of this study provide empirical evidence that the choice of augmentation strategy affects the performance and generalization of the model, and serve as a reference in the development of a Hybrid CNN–Swin Transformer-based malaria classification system.