Malaria diagnosis by use of automated systems in place of interpreted microscopic blood smear images can be used to increase the screening throughput and reduce diagnostic result variation. Four deep learning models are compared and analysed on a balanced dataset of 27,560 labelled cell images (13,780 Parasitized, 13,780 Uninfected), which were divided into training (80) and test (20) datasets, in this study. Two tailor-made convolutional neural networks (CNNs), Mo-bileNetV2 with transfer learning, and a Vision Transformer (ViT) are the models. The initially trained CNN, which was trained on 64x64 and most common types of augmentation, had a test accuracy of 96.06%. Mo-bileNetV2 was tested at 93.56% test accuracy on 128x128 images and has been demonstrated to be a lightweight alternative whilst being pretrained on ImageNet and a custom head. A deeper CNN, the second, was trained with 20 epochs with learning rate scheduling and regularization methods, yielding an accuracy of 94.32% on the test set, as well as on classification, and the accuracy during the validation varied over the training. ViT model was trained on 224x 224 images and optimized with Adam production which provided the best results 97.59 test accuracy. Finally, the great-est accuracy was achieved by attention-based models, yet by use of a CNN architecture, models were still competitive, and with less computationally costly, could permit repartitioning of computational resources and make a viable diagnostic choice in low-resource contexts. It is a complete reference to the classification of malaria cells and allusions to the choice of the model in situations of biomedical image classification.

CNN ViT Mobilenet Transfer learning Transformer Models