Malarial Parasite Classification Using Machine Learning on Microscopic Images of Peripheral Blood Smears

Abstract

Malaria is one of the leading causes of mortality in tropical and sub-tropical countries. The gold standard in malaria detection has been conventional microscopy which is fraught with challenges ranging from timeliness to difficulty in reproducing results which may be attributable to inter-person variation in reading of microscopy slides. The purpose of this project was to develop a Machine Learning model for malaria parasites diagnosis in Giemsa-stained microscopic cell images and build a reference application for the same. This would combat some of the prevalent contemporaneous issues in malaria diagnosis like the lack of expert technicians in resource constrained areas and the high cost associated with more sophisticated methods. In distinction to most of the preceding work, our work used microscopic images instead of decision tree algorithm and employed regularization and data augmentation techniques amongst other fine-tuning techniques. We also used thin blood smears as opposed to thick smears. Whilst thin blood samples are preferred for diagnostic tasks, they make the diagnostic task harder. We used transfer learning coupled with various data augmentation and regularization techniques. The data was obtained from the National Institute of Health’s data repository for malarial microscopic cell images. The characteristic parameters such as learning rate and weight decay were optimized through cyclical learning rate algorithm. The optimal batch size and the various augmentation and regularization parameters were determined through experimentation with various input parameters and adjusting based on error rates. The architecture that provided the best generalization and performance was that which used ResNet-152 for transfer learning with dropout and batch normalization. This recorded an accuracy of 0.9982, sensitivity of 0.9989, specificity of 0.9975and scored 0.9964 on the Mathew’s Correlation Coefficient (MCC). This model outperformed state-of-the-art Plasmodium diagnostic models on all metrics. Therefore, convolutional neural networks perform better in classifying malaria parasites when used in conjunction with regularization techniques such as dropout and batch normalization, and data augmentation combined with fine-tuned models and are such, better suited for such tasks. We have provided a pre-trained model that can be used for transfer learning in microscopy tasks. We have provided reference implementation and API documentation that was used.