Optimization of Malaria Cell Image Classification Using Pretrained Resnet50 Architecture with Data Augmentation and Fine-Tuning

Abstract

Malaria remains a significant health concern, particularly in tropical regions such as Indonesia, where timely and accurate diagnosis is crucial for reducing transmission and mortality. Conventional diagnosis through microscopic examination is labor-intensive, time-consuming, and highly dependent on expert availability. This study proposes an automated malaria cell image classification model using a deep learning approach based on the pretrained ResNet50 architecture. The research framework adopts the SEMMA (Sample, Explore, Modify, Model, Assess) methodology to structure the development workflow. A total of 27,558 labeled blood cell images comprising two balanced classes, Parasitized and Uninfected, were used for training and evaluation. Two model configurations were tested: a baseline model without data augmentation or fine-tuning, and an optimized model that integrates both. Augmentation techniques such as rotation, flipping, shearing, zoom, and brightness adjustment were applied to increase data diversity, while fine-tuning involved unfreezing the last 20 layers of ResNet50 to adapt pretrained features to the malaria domain. Performance was evaluated using accuracy, precision, recall, F1-score, loss, and AUC-ROC. The optimized model achieved 97.63% accuracy, 0.996 AUC-ROC, and 0.2472 loss, outperforming the baseline accuracy of 92.84%. An ablation study analyzed the individual contributions of augmentation and fine-tuning, showing that both techniques play complementary roles, with fine-tuning having the greater impact. A McNemar test confirmed that the improvements were statistically significant (p < 0.05). These findings demonstrate that the optimized ResNet50 model is effective for malaria detection and holds promise for integration into real-time diagnostic systems in resource-constrained environments.