Comparative Analysis of Deep learning architectures for automated diagnosis of malaria parasite.
This paper proposes the study of implementation and comparative analysis of pre-trained Convolutional Neural Network (CNN) models architectures on the basis of accuracy for malaria parasite detection. Malaria caused by the Plasmodium parasites, is a blood disorder, which is transmitted through the bite of a female Anopheles mosquito. It is detected by trained microscopists who analyse microscopic blood smear images. In recent times, Machine learning technologies have been used for automated diagnosis of malaria. Manual evaluation for diagnosis requires various steps to be performed. Moreover, this process leads to overdue and misguided analysis, even when it comes to the hands of expertise. The performance of pre-trained Convolutional Neural Network (CNN) dependent DL models as feature extractors for classifying parasitized and uninfected cells is evaluated in this study to help in improved disease screening. We take advantage of transfer learning methodology by examining pre-trained VGG-16, Xception, convolutional neural network (CNN) models with adjusted, densely connected classifiers. We compare the obtained results and visualized them. The proposed models have been evaluated on a dataset containing 13789 healthy and 12188 infected images. The dataset used was taken from National Institute of Health named NIH Malaria Dataset. The use of pre-trained Convolutional Neural Network (CNN) as a promising method for feature extraction for this reason is demonstrated by statistical confirmation of the results.