Improving Malaria Parasite Life Multi-Stage Classification Accuracy via Noise-Augmented Feature Extraction and Data Augmentation

The eradication of malaria by 2030 is one of the top priorities of the United Nations' Sustainable Development Goals and a fundamental goal for the world. In Indonesia, especially in resource-constrained eastern regions, accurate and comprehensive diagnostics are crucial to reducing the high mortality rates due to malaria. This research aims to improve the accuracy of malaria parasite life multi-stage classification to contribute to better malaria diagnosis by integrating deep learning and digital image processing techniques. Using a publicly available MP-IDB dataset of malaria parasite images, this paper proposes a large multi-stage classification framework that distinguishes 13 classes comprising different species of malaria parasites and their life stages. The dataset was split into 85% training-validation (Train-Val) and 15% testing. The Train-Val set was divided into 70% training and 30% validation after augmentation. The framework utilized EfficientNet-B7 for classification tasks. This study has significant potential for integration into real-world diagnostic workflows, particularly in resource-limited healthcare systems. The integration of advanced preprocessing techniques, such as contrast stretching, HSV conversion, and Otsu’s thresholding, along with noise generation in augmentation methods, ensured enhanced data diversity. The framework achieved a classification precision of up to 70%, demonstrating its potential for addressing complex, large, multi-stage classification tasks in constrained malaria microscopic image datasets. Furthermore, it attained an accuracy of 64.71%, outperforming other models, including EfficientNet-V2M, DenseNet-201, ResNet-152, MobileNet- V3Large, and EfficientNet-B0, under the same evaluation scenario.