Effective preprocessed thin blood smear images to improve malaria parasite detection using deep learning

Malaria can be difficult to detect from thin blood smears. Image recognition methods such as convolutional neural network can be used to detect malaria, but the training process takes a long time. Previous research created a new architecture and compares it to several other architectures such as VGG-16 and ResNet. The effect of preprocessing is analyzed in this research. VGG-16, ResNet, and the custom architecture created by the previous research are being used in this study. The preprocessing methods being analyzed in this research include gray-world normalization and comprehensive normalization. The highest accuracy improvement per epoch (0.5256% using ResNet-50 and 0.0352% using custom architecture) is achieved through gray-world normalization, that also improves final accuracy (90.1% using ResNet-50 and 93.1% using custom architecture) when compared to other methods with the same epochs for ResNet and custom architecture.